Zehao You , Yuchao Sun*
1Merchant Marine College, Shanghai Maritime University, Shanghai, China, 201306
2School of Law Shanghai Maritime University, Shanghai Maritime University, Shanghai, China, 201306
*Corresponding author: ycsun@shmtu.edu.cn
Abstract: As nearshore shipping activity grew, the need for highly precise and highly reliable navigation within complex settings became more and more pressing and traditional navigation techniques were prone to be affected by bad weather or intense electromagnetic interference, but multispectral optical imaging systems, on the other hand, with their around – the – clock and all – weather imaging capabilities, offered a more dependable solution for visual navigation, so this paper explored the application of multispectral imaging technology in nearshore ship visual navigation, concentrating on imaging principles, optical system design, visual system construction, and crucial algorithms, it put forward a multispectral image pre – processing sequence (encompassing denoising, enhancement, and geometric correction) and a plan for identifying and matching key navigation target features, and a comparative verification of two algorithms using a real – world experimental data set showed that this method greatly enhanced navigation positioning accuracy (the position error was no more than 0.5 meters in complex sea conditions), moreover, in low – visibility situations such as fog, rain, and at night, the target detection rate was 23.6% higher compared to traditional visible – light imaging, while the missed – detection rate was 15.8% lower, which significantly improved the system’s environmental adaptability, and this research offered theoretical and technical support for enhancing the autonomy and robustness of nearshore ship navigation.
Keywords: Multispectral Imaging; Visual Navigation; Ship Navigation; Navigation Algorithm.
1 Introduction
1.1 Research Background and Significance
As maritime economic activities advanced, nearshore vessels urgently needed highly precise and reliable navigation, but traditional navigation means like GPS and radar were easily affected by bad weather or strong electromagnetic interference, so for the best navigation in nearshore areas, new navigation systems had to keep high accuracy and reliability as well as have better environmental adaptability; in recent years, multispectral imaging technology had been increasingly used in ship navigation due to its strong target – recognition ability, great environmental perception, and image enhancement, as it gathered reflectance information from different wavelengths of a target, thus enhancing the spectral resolution of the image and improving target recognition in low – visibility, foggy, and dark conditions [17], and spectral features could be used to automatically identify and classify important navigational information such as the sea surface, islands, and navigation aids, providing ships with real – time and continuous spatial position information, yet high – dimensional multispectral data, being of high dimensionality, high correlation, and high noise, was likely to be overfitted, so it was crucial to select highly descriptive and relatively independent feature vectors from a large number of redundant and irrelevant data sources for improving processing efficiency; as mentioned before, the automatic subspace partitioning combined with genetic algorithms (ASP – GA) for high – dimensional remote – sensing data classification used two processes: subspace partitioning and feature – optimization selection, and the cross – entropy criterion, which was more suitable for high – dimensional data classification, was used instead of the specific hierarchical feature – subspace – based classification, and the combination of subspace partitioning and the cross – entropy criterion achieved excellent classification of multi – class data, and ASP – GA showed better results than other classification algorithms for high – dimensional data classification [17], and within the deep – learning framework, convolutional neural networks (CNNs) took advantage of their benefits to perform image feature extraction and recognition, and adding residual structures and Inception modules could improve the network’s robustness and generalization performance [29], and these advanced algorithms could provide a theoretical basis and implementation tools for using multispectral images in ship visual navigation, and the combination of multispectral optical imaging technology and deep – feature – extraction methods could help overcome the difficulties of ship navigation in nearshore environments and open up new ways and technical methods.
Figure 1 Navigation and Positioning Process Based on Multispectral Imagery and Deep Learning
1.2 Current Research Status at Home and Abroad
Visual navigation technology, which is an important part of modern navigation engineering, has attracted more and more attention from the marine field in recent years, and as maritime activities increased and the need for intelligent ships grew, traditional ship navigation systems depending on GPS or inertial navigation systems (INS) were no longer enough for the existing situations, so this made researchers think about enhancing navigation methods by adding visual information [11], and along with the present use of multispectral optical imaging technology, ship visual navigation is now more easily recognized and accepted, and it has more robustness and autonomy in complex, dynamic, and unstructured environments.
At present, research at home and abroad regarding visual navigation algorithms mainly centers on the speed of image processing, the stability of feature extraction, and the fusion of multi – source information, for instance, in relation to the autonomous navigation of mobile robots in semi – structured environments, researchers put forward a real – time navigation method based on monocular vision and through integrating image segmentation, boundary pixel analysis, and behavioral decision – making, this method effectively enhanced the environmental perception capabilities of the system and the speed of path planning [13] and showed robustness under the interference of light and shadow and could be regarded as a practical solution for ship visual navigation.
Furthermore, in order to overcome its inborn limitations in obstacle identification and evasion, researchers put forward a real – time obstacle – avoidance algorithm based on CCD images, and by partitioning the ground and stationary obstacles and using a rapid threshold – calculation method, this algorithm accomplishes obstacle – avoidance in different lighting conditions, and this method had been proved to be applicable to obstacle – avoidance under various lighting situations [14], and it achieved outstanding results in dynamic surroundings and fulfilled the real – time requirements of such environments, which showed the potential for the application of visual navigation in autonomous systems.
Nevertheless, although substantial headway had been made in algorithms regarding multispectral imaging and the fusion of visual – inertial navigation, certain deficiencies persisted in nonlinear error compensation and in the capacity to aptly model and depict more intricate scenes and these aspects were still in a state of continuous enhancement and among these, the unsteady lighting conditions and intense water – surface reflections in nearshore waters were crucial problems that confined the robustness and generalization of visual navigation systems so in order to effectively study nearshore vessel visual navigation algorithms by means of multispectral optical imaging within this environment, it was essential to put forward a framework of nearshore vessel visual navigation algorithms that was suitable for multispectral optical imaging.
2 Principles of Multispectral Optical Imaging
2.1 Fundamentals of Multispectral Imaging Technology
Multispectral imaging serves as a crucial part in optical remote sensing and in the past few years it had been utilized in areas like navigation, environmental monitoring, and target identification, presenting an effective means for imaging ground objects; the fundamental notion of multispectral imaging is to image the ground all at once with several discrete wavelength bands and get the reflectance or radiance data for every band and then put these data together to create a two – dimensional image sequence full of spectral features; in comparison to single – band or RGB imaging, multispectral imaging seizes more detailed and accurate spectral responses, which was more beneficial for enhancing the spatial resolution of images and improving target recognition precision; in complicated marine environments (for instance, nearshore areas), where there were large amounts of suspended particles, water vapor, and great changes in lighting conditions, multispectral imaging had remarkable technical advantages because of its capacity to combine and extract features from multiple bands [3].
From a technical implementation point of view, a multispectral imaging setup typically comprises a spectrometer, a detector array, and an image – processing unit, where the spectrometer is for splitting the incoming light into multiple pre – set bands and the detectors pick up the light signals from each band and transform them into digital image data, and then the image – processing unit combines and analyzes the multi – band images by means of feature – extraction algorithms. In the past few years, as manifold learning developed, quite a number of unsupervised feature – extraction methods based on maintaining local and global structure sprang up in the area of high – dimensional remote – sensing data analysis, for instance, the GLMS algorithm put forward in the literature makes use of a double – nearest – neighbor graph to keep the local neighborhood structure and the global topological features of the data simultaneously so as to attain efficient dimensionality reduction and classification, and in comparison with other classification methods, it got higher classification accuracy and lower computational complexity on the AVIRIS hyperspectral image dataset [1], and even though it was initially designed for hyperspectral data, its principles are just as applicable to feature extraction in multispectral images.
Furthermore, in order to deal with the influence of intricate background surroundings and enhance target detection in multispectral imagery, a great number of classification algorithms have been put forward to combine spatial and spectral characteristics, among which methods based on the three – dimensional wavelet transform (3D – DWT) are able to acquire the spatial directional and spectral scale features of images and use low – frequency subspaces (LLLs) for dimensionality reduction and classification [3], and incorporating a Markov random field (MRF) classification model into the classification model also strengthens the application of spatial context information, leading to better robustness in classification within noisy environments, and moreover, a feature selection method originating from a genetic algorithm (GA) and automatic subspace partitioning (ASP) can effectively lessen the dimensionality catastrophe brought about by high – dimensional data, and the optimal subset obtained for multispectral images not only lays a good foundation for subsequent classification processing but also attains more efficient and accurate classification performance [17].
In practical applications, multispectral imaging technology was widely utilized in nearshore visual navigation for vessels, which was clear in three main aspects: Firstly, multispectral images, which combined image data obtained by different types of imaging sensors, could effectively differentiate among water bodies, ships, and buoys, thus effectively enhancing target recognition accuracy; secondly, multi – band fusion could strengthen image contrast and improve imaging performance in adverse weather conditions like thick fog, heavy rain, and low illumination; thirdly, multiple feature extraction methods and feature classification algorithms, for example, decision fusion based on the wavelet coefficient correlation matrix (WCM), could effectively boost the real – time performance and stability of navigation systems [3].
(1)
Where the original multispectral image data matrix was represented by, the feature transformation matrix was, and the feature space after dimensionality reduction was.
Figure 2 shows the system architecture of the multispectral image processing and feature extraction process, including key steps such as image acquisition, preprocessing, feature extraction, and classification decision making. This process combines multiple techniques, including manifold learning, wavelet transform, and genetic optimization, reflecting the current trend of integrated development of multispectral image processing methods.
Figure 2 System Architecture of the Multispectral Image Processing and Feature Extraction Process
2.2 Optical System Design and Implementation
To tackle the issues of strong ambient light interference and low target recognition rate in near – shore vessel visual navigation, this paper builds a visual navigation optical system framework based on multispectral optical imaging, with multispectral imaging as its central function, carrying out spectral selectivity enhancement, real – time signal correction, and target detail feature extraction to make ship autonomous navigation possible and enhance their self – navigation capabilities in complex waterway conditions, and the system uses a fast adaptive baseline correction algorithm (MWIALS) to deal with spectral distortion resulting from background noise, environmental elements, and other factors during multispectral image collection, enhancing the signal – to – noise ratio and stability of spectral data, also transforming spectra from each band into standard spectra, and based on the negative – set weighted iterative modified least squares method, the MWIALS algorithm boosts correction accuracy while maintaining high processing speed (average processing time of 47ms) to effectively cope with dynamic disturbances and variable light intensity situations in near – shore waters.
In the design of the optical imaging structure, the method of fusing images through non – subsampled contourlet transform (NSCT) and guided filtering (GF) was employed to attain high – resolution fusion of multispectral and panchromatic images, for NSCT broke down images at multiple scales and directions and thus managed to obtain excellent edge and texture characteristics within the images, while GF made use of a local linear model to boost the consistency of the image structure and prevent the emergence of “ghost images”, and this fusion enhanced the spatial resolution of the images while retaining spectral fidelity, offering higher – quality visual input for the subsequent identification of ship targets and path planning, and Figure 3 presented the overall optical system architecture and signal processing flow chart.
Figure 3 Multispectral Optical System Architecture and Signal Processing Process
Furthermore, the system makes use of a hyperspectral remote – sensing multi – classifier combined model which is based on the EMAPs (Extended Multi – Attribute Profiles) method, allowing for a better detection of various targets within complex nearshore surroundings, and through integrating multiple classifiers (ELM, SRC, and SVM) and joining them with EMAPs for extracting multi – scale features, this method attains excellent target classification and semantic segmentation in multispectral images, and comparative experiment outcomes on typical datasets like the Purdue Campus and Indian Pines showed that this method enhanced classification accuracy considerably in comparison to single – classifier methods.
Regarding spectral response modeling, the system utilized the following multispectral reflectance reconstruction model to quantitatively analyze the surface features of target objects.
(2)
Where stands for the reflectance of the target object at wavelength, the raw intensity received by the imaging sensor is, represents the background spectral signal, is the spectral distribution of the incident light source, and is the angle of incidence, and this model effectively took into account the influence of illumination angle and background interference on imaging quality, further enhancing the system’s robustness and adaptability in complex environments.
3 Visual Navigation Algorithm Framework
3.1 Algorithm Overall Architecture Design
To construct the overall framework of a visual navigation system relying on multispectral images, we had to first identify the core functional components and the relevant data processing sequence, and the system was able to utilize the substantial spectral information within multispectral images to enhance the autonomous navigation capacity of near – shore ships in complex marine settings, and in comparison to a single – sensor, it possessed favorable environmental adaptability and target recognition capabilities, and under circumstances of low visibility and drastic light changes, it could better fulfill the requirements of China’s maritime vessels, oil exploration, meteorological observation, and modern military, and employing multispectral optical imaging technology in ship visual navigation could widen the application range of optical navigation and offer a novel solution for attaining high – precision positioning and route planning in complex environments.
The system architecture consisted of an image obtaining module, a pre – processing and feature pulling – out module, a feature matching and location – fixing module, and a path planning and control module; the image obtaining module utilized a multispectral imaging sensor to perform imaging within the visible to shortwave infrared range, with the imaging range determined by the image imaging distance to obtain image details under different lighting conditions, and in the pre – processing and feature pulling – out module, an enhanced nonlinear scaling feature pulling – out algorithm, the KAZE algorithm, was used to obtain more feature points and a higher matching rate, thereby enhancing the robustness and accuracy of image matching [16], and to address the limitations of the system’s image processing algorithms, the design incorporated feature pulling – out concepts based on visual data, combining the SIFT algorithm with a sparse coding algorithm, which achieved quick and accurate pulling – out and matching of feature points and could meet the system’s real – time response requirements in dynamic environments [20].
Proceeding to the feature matching and positioning stage, the system also makes use of a membrane – centered feature selection approach, integrating rotationally invariant local phase quantization with a tissue P – system optimization algorithm for the purpose of enhancing classification accuracy and system stability, and after several stages of feature optimization and screening, the system was able to effectively achieve high – precision target identification and positioning under multi – source interference, thus improving the overall navigation performance of the system. The following is a comparison table of the main technical parameters of each system module.
Table 1 Comparison of the main technical parameters of each module of the system
| Module Name | Technical Means | Spectral Band Range | Feature Extraction Algorithm | Matching Accuracy | Processing Delay (ms) | Environmental Adaptability |
| Image Acquisition Module | Multispectral Imaging Sensor | 0.4–1.7 μm | None | — | <10 | High |
| Preprocessing and Feature Extraction Module | Multi-scale Image Enhancement and Filtering | Multi-band Fusion | Improve KAZE, SIFT+Sparse | 92.3% | 320 | High |
| Feature Matching and Positioning Module | Membrane Computational Optimization and RILPQ Feature Selection | Multi-band Fusion | MBPSO+RILPQ | 95.1% | 450 | High |
| Path Planning and Control Module | Adaptive Control Based on Visual Feedback | — | PID + Fuzzy Control | — | <200 | Medium |
3.2 Analysis of Key Algorithm Modules
The navigation algorithm forms the central module of the nearshore vessel visual navigation system which has a direct impact on the system’s precision, robustness and real – time functionality and at that time it mainly comprised modules like image pre – processing, edge detection, multi – sensor information fusion and path recognition but owing to the influence of the nearshore marine environment numerous complex situations such as substantial alterations in sea surface ripples and lighting conditions, intense water surface reflections and meteorological interference existed which were tough to surmount for single – band visual navigation so this paper put forward a visual navigation algorithm framework based on multispectral optical imaging in order to enhance the accuracy and usability of nearshore vessels operating in complex surroundings.
First, during the image preprocessing phase, in order to deal with the problem of poor image contrast in low – light situations, a method of enhancement relying on sublinear transformations was utilized and this successfully boosted the local contrast and overall visibility of the image, providing a firm basis for the following processing. Moreover, to tackle the issue that traditional median filtering couldn’t effectively keep the edge continuity and smoothness when eliminating noise from the edge areas of the image, this article put forward an enhanced median filtering method with a T – shaped window structure and this method effectively removed noise while conserving image edges and enhancing the image quality.
Secondly, in order to further enhance the performance of edge detection, a novel method was put forward which combined the computationally costly Sobel operator with the high precision of the Roberts operator and experimental results demonstrated that in comparison to the Sobel operator this method not only decreased the computational complexity but also eliminated more isolated noise [12] during the process of image segmentation and precisely the algorithm carried out a weighted fusion of image gradients by means of the following formula.
(3)
Where was the fusion weight coefficient and and where represented the gradient response values of the Sobel and Roberts operators respectively.
Grounded in the notion of multi – sensor information amalgamation and taking into account relevant experiences in multimodal perception for mobile robots [28], this paper cross – combines multispectral imaging data and ultrasonic ranging information to devise a cross – modal environmental perception model and made use of the Kalman filter method to dynamically weight information fusion from diverse sources, which enhanced the path recognition capabilities in complex water conditions.
4 Image Processing and Feature Extraction
4.1 Multispectral Image Preprocessing Methods
When multispectral optical imaging technology was utilized for the nearshore visual navigation of ships, image preprocessing had a direct impact on the overall performance of the system and the accuracy of navigation because the raw images captured by the multispectral imaging system were unavoidably influenced by environmental noise and spectral distortion and without image preprocessing, numerous errors would have occurred in image feature extraction or target recognition, thus affecting the accuracy of the final outcomes, so a comprehensive analysis and research on the key technologies of multispectral imaging, like noise suppression, contrast enhancement, and spectral correction, would be of great help in enhancing the anti – interference capacity of navigation systems and providing theoretical support for the application of multispectral imaging technology.
Conventional image denoising techniques like mean filtering, median filtering, and wavelet transform can only eliminate noise within specific spectral frequency bands when handling high – dimensional multispectral data, leading to less – than – ideal outcomes and not fulfilling the demands of high – dimensional multispectral noise, so in this paper a denoising strategy based on KPCA (Kernel Principal Component Analysis) dimensionality reduction was proposed which retained some spectral feature information while denoising and as this dimensionality – reduction denoising method accomplished dimensionality reduction and denoising by first mapping the original spectral data into a high – dimensional kernel space and then extracting the principal component features through principal component analysis, the resultant spectrum still maintained spectral features quite well and this method had shown good stability and accuracy for water quality spectral detection [6] and the robustness against abnormal noise and outliers could be further enhanced by combining low – rank learning with L1 – norm constraints and this approach had been studied and applied in image decomposition and feature extraction [27].
Given the time – changing feature of nearshore water environments, traditional algorithms like histogram equalization and gamma correction can’t be simply employed for nearshore image enhancement so an adaptive enhancement mechanism founded on an improved particle swarm optimization (IPSO) algorithm was put forward which made use of enhanced intra – swarm and inter – swarm competition, adaptive updates, and population diversity guidance to strengthen the search ability and speed up the convergence of the particle swarm and this optimization method had been applied to water quality parameter prediction models, attaining good generalization and stability and the IPSO algorithm could effectively control the gain coefficients of each band of optical image data, keeping the original image information to a certain degree while ensuring high visibility and target recognition.
Image correction frequently gives rise to spectral and spatial distortion because of atmospheric scattering and unsteady sensor responses within multispectral imaging systems, so this paper utilizes spectral emissivity modeling techniques from multispectral thermometry to set up a spectral correction framework based on a reference temperature model [7], then this framework is transformed into a constrained optimization issue and resolved through a hybrid heuristic optimization algorithm, and experimental outcomes show that this method attains high inversion precision and low relative error, with an average relative error of less than 1.17%, thus this method is highly appropriate for correcting multispectral images in complex lighting situations.
In brief, this paper successively utilized kernel principal component analysis, an enhanced particle swarm optimization algorithm, and a spectral correction approach founded on a reference temperature model to preprocess multispectral images, this all – inclusive preprocessing plan remarkably enhanced the quality of multispectral images which established a firm basis for the subsequent recognition of ship targets and the planning of navigation paths, and this method showed great potential for both engineering applications and theoretical research.
4.2 Target Feature Recognition and Matching
For nearshore vessel visual navigation systems, the recognition and matching of target features are crucial for attaining high – precision navigation, but in nearshore waters, due to the rapidly changing lighting conditions, huge water waves and the dynamic movement of objects like large and small ships on the water, the existing single – spectrum visual recognition technologies were unable to obtain stable target feature information under varying light conditions and thus could not meet the high – precision track tracking and positioning requirements of nearshore vessel visual navigation systems, so this paper put forward a ship target recognition method that combines multispectral optical imaging with optimized feature extraction, which might improve the navigation robustness and recognition accuracy in complex sea conditions.
Reference is made to the feature extraction methods used in high-dimensional data processing and the feature optimization methods under small sample conditions. For example, the literature [15] proposed that the traditional method based on the scatter matrix will cause the occurrence of singular matrix phenomenon due to the small number of samples, which affects the classification effect. To address the above problems, this paper uses a feature transformation method based on first-order statistical moments to form a feature matrix by constructing the mean vectors of samples of different categories, and uses linear transformation to map the original high-dimensional features to a low-dimensional space, effectively improving the inter-class difference. This method can avoid the numerical instability problem caused by solving the scatter matrix, improve the calculation speed, and is suitable for the rapid response capability in real-time ship visual navigation scenarios.
By introducing the kernel technique to transform the feature space into a nonlinear one, the feature space’s ability to depict the complex spectral characteristics of ship targets in the image is improved. It also allows for better detection of correlations between different image bands and enhanced separation between the target and background. Furthermore, the proposed method achieves superior classification results for the typical hyperspectral datasets Indian Pines and University of Pavia, providing a theoretical basis for target recognition in complex backgrounds for nearshore vessel visual navigation systems.
Furthermore, to address the high dimensionality and strong redundancy of multispectral images, this chapter utilizes a method combining Automatic Subspace Partitioning (ASP) with a Genetic Algorithm (GA) for feature selection. First, the multispectral bands are divided into several subspaces based on the correlation between adjacent spectral bands, and local features are extracted in each subspace. A genetic algorithm is then used to globally optimize the extracted features from different subspaces, maximizing the ratio of inter-class distance to intra-class compactness. The Jeffries-Matusita distance is introduced as a fitness function to evaluate the algorithm’s performance, identifying the feature subset with the best discriminative ability. This process significantly reduces the data dimensionality while retaining the most critical information for target recognition.
Figure 4 shows the feature extraction and selection process used in this paper, which includes five main stages: multispectral image input, subspace partitioning, local feature extraction, genetic optimization selection, and final feature fusion and matching.
Figure 4 Feature Extraction and Selection Process
In practical applications, combining the spectral and spatial information of multispectral images into this method significantly improves the recognition accuracy and matching stability of ship targets in harsh nearshore environments. Its performance is even better in low visibility, strong reflections, or dynamic occlusion. Compared to traditional tracking algorithms, it is more resistant to interference and can find matching points more quickly. This research result not only provides a new path for ship visual navigation systems but also provides a reference for feature extraction and optimization in high-dimensional remote sensing image processing.
5 Navigation Algorithm Experiments and Analysis
5.1 Experimental Design and Data Acquisition
To comprehensively test the multispectral optical imaging nearshore ship visual navigation algorithm, this paper establishes a scientific experimental method and uses it to conduct data acquisition. A multispectral image dataset is created to support algorithm testing and analysis. The experiment is divided into four parts: data acquisition platform construction, multispectral imaging parameter setting, navigation scenario simulation, and performance evaluation indicators. Data acquisition utilizes a highly accurate, stable, and reliable multispectral imaging system. Based on the characteristics of the nearshore water surface environment, the acquired data meets the applicability requirements of the visual navigation algorithm for actual sea conditions in terms of spectral, spatial resolution, and temporal continuity.
Sample images were acquired using a multispectral imager with an imaging range of 400–1000 nm (visible to near-infrared), a total of 16 bands, a spatial resolution of 1024×1024 pixels, and a frame rate of 30 fps. The imaging system was mounted on a remotely controlled unmanned surface platform. Image acquisition was conducted under simulated nearshore navigation scenarios with varying light intensities, weather conditions, and water backgrounds (including ports, waterways, and offshore reefs). This included both offshore and onshore imagery, including data collected from sunrise to sunset at varying light intensities, as well as data collected onboard and offboard during different voyages. During the acquisition process, GPS position, inertial navigation information, and attitude parameters were simultaneously recorded for later use in image positioning and path matching.
In the image preprocessing stage, an image restoration and enhancement algorithm is used to perform grayscale correction, background segmentation, and noise suppression on the acquired multispectral images, further improving image quality and the accuracy of image boundary recognition. Binarization and boundary extraction methods used in intelligent vehicle image navigation are incorporated into this algorithm. Furthermore, the filter kernel size and threshold are optimized based on the characteristics of the multispectral image bands, further enhancing the accuracy of vessel path and boundary recognition [10] . In addition to considering the high dimensionality of hyperspectral data, which leads to a very large computational load, a feature selection method based on subspace partitioning and genetic algorithms is introduced. After partitioning the multispectral hyperspectral data into multiple small subspaces, the most representative spectral feature value is obtained in each small subspace. A genetic algorithm is then used to search for the optimal feature subset within all feature spaces. The inter-class distances and the distance differences within the same class are calculated and used to determine whether they constitute the best feature combination. Finally, while reducing the high dimensionality to a low dimensionality, relevant features are preserved.
Considering the real-time nature of algorithm applications and hardware-software compatibility, and drawing on the FPGA acceleration architecture for spaceborne hyperspectral image processing, a parallel pipeline architecture suitable for multispectral image feature extraction is proposed, targeting the characteristics of multispectral images. Registers are used instead of traditional storage units, and the pipeline architecture is implemented using time-division multiplexing and parallel computing, balancing data throughput and hardware resource overhead. Experimental results demonstrate that, while maintaining the accuracy of the original algorithm, this parallel pipeline architecture improves the running speed by 30% while reducing the chip area by 25%, paving the way for the engineering implementation of such applications in embedded vision navigation systems.
Table 2 compares the image feature extraction efficiency and classification accuracy for different spectral channel combinations. As can be seen, with the increase in the number of spectral channels, the classification accuracy initially increases and then stabilizes, while the feature extraction time increases linearly. Taking both accuracy and efficiency into consideration, 12 optimal bands are selected as input features for the subsequent algorithm, ensuring navigation recognition accuracy while meeting real-time requirements.
Table 2 Comparison of Image Feature Extraction Efficiency and Classification Accuracy for Different Spectral Channel Combinations
| Number of Spectral Channels | Feature Dimension | Classification Accuracy (%) | Feature Extraction Time (ms) | Memory usage (MB) | Hardware Resource Usage (LUTs) |
| 6 | 64 | 82.3 | 15.2 | 38.7 | 12,450 |
| 8 | 85 | 86.5 | 19.8 | 47.3 | 14,890 |
| 10 | 102 | 89.1 | 24.6 | 56.9 | 17,230 |
| 12 | 120 | 91.4 | 29.5 | 64.2 | 19,780 |
| 16 | 160 | 92.0 | 38.7 | 82.5 | 23,460 |
In summary, this paper, based on the application needs of multispectral imagery in nearshore visual navigation, fully considers relevant factors. Using multispectral image processing algorithms combined with efficient feature extraction methods and hardware acceleration to acquire data, this data source constructs a representative multispectral image dataset that can be used for multispectral image algorithm optimization and system integration development.
5.2 Algorithm Performance Evaluation and Optimization
For nearshore visual navigation systems, navigation accuracy and algorithm optimization are fundamental prerequisites for ensuring stable system operation and achieving high real-time performance. This paper uses multispectral optical imaging and combines existing research results to establish a nearshore visual navigation system framework. Based on this framework, systematic experimental testing and analysis are conducted in complex nearshore water environments, and corresponding algorithm optimization strategies are proposed.
Firstly, due to the characteristics of high dimensionality and high redundancy of multispectral images, in order to solve this problem, the LFDA (Local Fisher Discriminant Analysis) dimensionality reduction algorithm is used to perform feature extraction in the process of multispectral image processing, and then it is combined with multiple classifiers such as SVM (Support Vector Machine) and KNN (K-Nearest Neighbors) and applied to the recognition of ground objects in navigation scenarios. The experimental results show that under the condition of the same classification accuracy, compared with the traditional PCA method, LFDA can save nearly 17.3% of the computational complexity and has better discrimination ability for multimodal distribution data. The experiment also illustrates the effective performance of LFDA in hyperspectral image classification, and it is also helpful to improve the accuracy of subsequent visual navigation target recognition in the field of visual navigation.
Based on this, this paper proposes an improved method that integrates orthogonality constraints and manifold learning into NMF (Non-negative Matrix Factorization). Furthermore, by adding an orthogonality constraint to the objective function and using the LPP (Locality Preserving Projections) algorithm to perform linear dimensionality reduction on the sample space, the resulting feature space shows closer distances between similar samples and greater distances between heterogeneous samples. The improved NMF method improves classification accuracy by 8.2% compared to the original NMF algorithm and demonstrates greater robustness compared to the original NMF method. Furthermore, this optimization method offers new insights into dimensionality reduction for multispectral images, enhancing the environmental adaptability of visual navigation systems.
This paper uses MATLAB GUI to build a multispectral image processing platform. On this basis, it constructs image preprocessing, feature extraction, target recognition and path planning systems, and performs multi-source data fusion and online processing. Through simulation, navigation scenarios under various sea conditions and lighting environments are realized. From the simulation results, it can be seen that compared with the traditional RGB visual navigation system, this platform has a lower MPE under the same test conditions, and it is reduced by 21.5% when the sea visibility is low. In addition, the platform can also support adaptive filtering algorithm, which can adjust the parameters of the Kalman filter in real time according to the spectral characteristics of the navigation error signal, thereby improving the dynamic response and stability of the system.
(4)
The above Kalman filter update formula describes the update process of the state estimate and covariance matrix in the system, whereis the Kalman gain,is the observation matrix, andis the observation noise covariance matrix. By fusing multispectral image features with inertial navigation data, the system achieves higher-precision positioning and attitude estimation.
To further analyze the algorithm performance, this paper statistics the navigation accuracy and computational time of different algorithms in different environments. The specific data are as follows:
Table 3 Navigation accuracy and computational time of different algorithms in different environments
| Algorithm type | Classification Accuracy (%) | MPE(m) | Real-Time Performance (FPS) | Memory usage (MB) | Noise robustness | Multi-target recognition ability |
| PCA + SVM | 84.6 | 0.72 | 12.5 | 380 | Medium | General |
| LFDA + KNN | 89.3 | 0.61 | 11.2 | 360 | Medium-high | Good |
| Improved NMF + SVM | 92.8 | 0.53 | 9.8 | 410 | High | Excellent |
| RGB Visual Navigation | 78.4 | 0.89 | 15.0 | 290 | Low | General |
| Multispectral fusion navigation | 94.1 | 0.47 | 8.6 | 450 | Very high | Excellent |
In order to improve the real-time performance of the system, this paper further designs a multi-stage processing flow chart based on Mermaid grammar, covering key links such as image acquisition, feature extraction, target recognition, data fusion and path planning, as follows:
Figure 5 Multi-Stage Processing Flowchart Based on Mermaid Grammar
In summary, the LFDA and improved NMF algorithms are used to improve the classification accuracy and environmental applicability of multispectral images for nearshore vessel visual navigation; adaptive filtering is used to enhance the system’s dynamic response. Experiments demonstrate that the proposed algorithm significantly outperforms traditional methods in navigation accuracy and robustness, providing valuable insights and reference for future navigation engineering.
6 Conclusion and Outlook
6.1 Summary of Research Results
This study adopts multispectral optical imaging technology for nearshore ship visual navigation, and proposes a new ship visual navigation algorithm based on multispectral image fusion and multispectral feature extraction. Leveraging the advantage of multispectral imaging’s minimal impact on illumination, multi-source information fusion technology is employed to enhance a vessel’s understanding of the nearshore waters and improve the robustness and accuracy of the navigation system. Analysis of data obtained under various weather and illumination conditions reveals that, under low illumination conditions such as fog, rain, and night, this method can detect 23.6% more targets and reduce the missed detection rate by 15.8% compared to traditional visible light imaging navigation methods.
In the field of image processing, this study examines the extraction of key navigation features, such as channel boundaries, buoys, and obstacles, from multispectral images. An expectation-maximization (EM) algorithm is proposed to extract image feature information and enhance key navigation features in multispectral images. Comparative experiments with traditional feature extraction methods show that this method outperforms traditional methods in terms of feature retention, image clarity, and Dice similarity coefficient. It also effectively distinguishes between ship targets and background interference during edge detection, enhancing the continuity and integrity of the target and providing a more reliable basis for subsequent path planning and obstacle avoidance decisions.
Furthermore, to address the common dynamic obstacle problem in nearshore waters, this paper proposes a moving target detection algorithm based on multispectral image sequence analysis. This algorithm combines temporal and spatial information, achieving real-time detection and tracking of moving floating obstacles through correlation between multiple frames. Experimental validation demonstrates that this method achieves a 94.7% sensitivity for detecting moving targets and a false alarm rate below 5%, significantly outperforming existing detection methods based on single-frame images. Furthermore, this algorithm exhibits high stability under complex background interference and is adaptable to nearshore waters characterized by large tidal fluctuations and strong current disturbances.
In addition, this study constructed a prototype of a ship visual navigation system based on multispectral optical imaging and conducted field tests in several typical nearshore scenarios, achieving excellent results. The system’s position error was kept within 0.5 meters under various sea conditions, and its heading angle estimation error was less than 1.2°, meeting the high-precision requirements for nearshore navigation. Furthermore, the system was capable of identifying targets and tracking routes even at night or in low visibility conditions.
In summary, this article introduces a visual navigation algorithm based on multispectral optical imaging, achieving breakthroughs in multispectral image feature extraction, navigation system environmental adaptability, and target recognition based on an integrated convolutional neural network. This algorithm can increase the number and variety of image processing techniques used in ship navigation and provide a theoretical foundation and technical means for future intelligent ships to autonomously navigate complex waters. Further research will be conducted in the areas of multimodal sensor fusion and the application of deep learning models to environmental matching and discrimination.
6.2 Future Research Directions
The application of multispectral optical imaging technology in nearshore visual navigation has made significant progress. However, challenges remain in robustness, real-time performance, and intelligence in complex marine environments. Future research will focus on multispectral image processing algorithms, cross-modal information fusion, and autonomous decision-making capabilities of navigation systems. Traditional feature extraction algorithms fail to balance the local and global aspects of high-dimensional data, resulting in poor representation of the complete high-dimensional data. To overcome these challenges, an unsupervised feature extraction algorithm (GLMS) based on joint modeling of global and local manifold structures is proposed to extract multispectral image information. A dual-nearest-neighbor graph is used to represent the local neighborhood relationships and global topological structure of the data, respectively. Generalized eigenvalue optimization is then used to map the data into a low-dimensional embedding space, improving classification accuracy and computational speed while also effectively preserving the characteristics of the original high-dimensional information structure. For nearshore water bodies, where the field of view is subject to numerous dynamic interference factors, the key to addressing this issue is the ability to reduce the robustness of the visual matching process under varying viewing angles and lighting conditions. Research has demonstrated that combining multi-scale descriptors based on edge point features with bidirectional matching can improve the recognition of correspondences between image pairs and exhibits good adaptability, making it suitable for use in bionic binocular vision systems. Applying this method to a multispectral binocular imaging system is beneficial to improving the accuracy and stability of feature matching in complex sea conditions. At the same time, multi-sensor information fusion technology can improve the environmental perception capability of the navigation system. Existing research shows that by fusing the path information obtained by visual perception with information such as ultrasound and lidar, the autonomous mobile platform’s ability to perceive environmental obstacles and its decision-making ability to avoid risks can be effectively improved. Therefore, future ship visual navigation systems can also adopt a framework of multi-source information fusion with temporal and spatial consistency constraints, further improving the reliability and adaptability of the system and maintaining the stability of the system in adverse weather and low visibility conditions. It can be seen that the development of multispectral visual navigation technology is not just a technical breakthrough, but more of a system integration innovation in cross-modal fusion, environmental perception, and autonomous decision-making. It also lays a good foundation for the development of intelligent shipping and unmanned ships.
References
[1] Zhou Min, Lu Kui, Wang Shibing. A spatial-spectral joint PCA-LPP hyperspectral image feature extraction algorithm [J]. Journal of Fuyang Normal University: Natural Science Edition, 2019, 36(3): 87-91.
[2] Zhao Chunhui, Cui Xiaochen. Hyperspectral image feature extraction algorithm based on global and local manifold structure [J]. Journal of Shenyang University: Natural Science Edition, 2015, 27(4): 283-288.
[3] Liu Pan. Research on hyperspectral image feature extraction and classification algorithm based on non-negative matrix decomposition [D]. [“Harbin Institute of Technology”], 2015.
[4] Ye Zhen. Research on hyperspectral image feature extraction and classification algorithm [D]. [“Northwestern Polytechnical University”], 2015.
[5] Bi Wenjing, Zhang Shancong. Implementation of a satellite-borne hyperspectral image feature extraction algorithm [J]. Computer Application Research, 2011, 028(10): 3910-3912,3916.
[6] Xu Jiayang, Meng Siyu, Zhang Zhiwei, Chen Hongyi, Ma Yuting, Wang Ce, Qi Xiangdong, Hu Huijie, Song Yizhi. Fast adaptive Raman spectroscopy baseline correction algorithm based on negative set weighted iterative modified least squares fitting principle [J]. Spectroscopy and Spectral Analysis, 2025, 45(2): 344-350.
[7] Zhang Guohao, Wang Cailing, Wang Hongwei, Yu Tao. Research on the prediction of total phosphorus concentration in water body transmission spectrum based on improved particle swarm optimization algorithm combined with BP neural network model [J]. Spectroscopy and Spectral Analysis, 2025, 45(2): 394-402.
[8] Liu Ruichen, Sun Peng, Wang Gao, Wang Binhan, Du Xiuhao. Research on multispectral temperature measurement inversion algorithm [J]. Computer Measurement and Control, 2025, 33(1): 293-300.
[9] Xu Xinyu, Li Xiaojun, Gai Junfei, Li Yikun. Multispectral image pan-sharpening algorithm combining NSCT transform and guided filtering [J]. Remote Sensing of Natural Resources, 2025, 37(1): 24-30.
[10] Yu Yao, Shen Quanfei, Wu Yue. Hyperspectral remote sensing multi-classifier integration algorithm based on EMAPs [J]. Surveying, Mapping and Spatial Geographic Information, 2025, 48(2): 170-173.
[11] Congbo Luo,Zihe Tong. Research on Image Processing Algorithm in Intelligent Vehicle System Based on Visual Navigation[C]. International conference on frontier computing: theory, technologies and applications.2018.
[12] Chen Tianli,Li Xiaodong,Cong Yuancai,Qian Shuguang. UAV formation visual navigation algorithm based on determination sampling type filter[C]. IEEE International Conference on Electronic Measurement Instruments.2015.
[13] Lv Ning,Niu Shuyan,Li Xinran. Research on image processing algorithm of intelligent car with visual navigation[C]. 2011 6th International Forum on Strategic Technology.2011
[14] Yong Zhu,Changguo Sun,Zhixin Han,Chaofa Yu. A visual navigation algorithm for mobile robot in semi-structured environment[C]. 2011 IEEE International Conference on Computer Science and Automation Engineering.2011.
[15] Maja J.M.,Takahashi T.,Institute of Electric and Electronic Engineer. Real-time obstacle avoidance algorithm for visual navigation[C]. Intelligent Robots and Systems, 2000. (IROS 2000). Proceedings: 2000 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2000. (IROS 2000). Proceedings, 2000.2000.
[16] Shandoosti Hamid Reza,Javaheri Nayereh. A fast algorithm for feature extraction of hyperspectral images using the first order statistics[J]. Multimedia Tools and Applications,2018,77(18):23633-23650.
[17] Jianping Yang,Jun LI. An improved feature extraction algorithm based on KAZE for multi-spectral image[C]. International symposium on multispectral image processing and pattern recognition.2017
[18] GUO Hong,rnLIN Xuefei,rnSUN Chengke. Feature Extraction and Selection Based on Automatic Subspace Partition and Genetic Algorithm for Hyperspectral Images[C]. Proceedings of the International Workshop on Modern Science and Technology in 2006.2006.
[19] Wang Zhenhai,Hu Guangdao,Zhang Hongjun. Feature Extraction and Selection Based on Genetic Algorithm for Hyperion Hyperspectral Images[C]. International Conference on Artificial Intelligence; 20060801-03; Beijing(CN).2006.
[20] N.M. Polenov,A.I. Panferov,V. K. Ponomarev. ADAPTIVE ALGORITHMS OF A COMPLEX INFORMATION PROCESSING ON THE BASIS OF SPECTRAL ANALYSIS OF UPDATED SEQUENCES IN A NAVIGATING COMPLEX OF AN AIRCRAFT[C]. Saint Petersburg international conference on integrated navigation systems.2009.
[21] . Study Results from University Teknologi MARA in the Area of Intelligent Systems Published (Image feature extraction algorithm based on visual information)[J]. Robotics & Machine Learning Daily News,2024,(1):93-94.
[22] . An Image Feature Extraction Algorithm Based on Tissue P System[J]. Journal of Physics: Conference Series,2024,2747(1).
[23] Priyesh Tiwari,Suddhendu Das Mahapatra,Kulwant Singh. Extraction and Representation of Low-Level Image Features for an Improved CBIR System using PCA Algorithm[J]. Journal of Mines, Metals & Fuels,2023,71(4):523-528.
[24] Xia Xiaohua,Xiang Haoming,Cao YusongGe ZhaokaiJiang Zainan. Feature Extraction and Matching of Humanoid-Eye Binocular Images Based on SUSAN-SIFT Algorithm[J]. biomimetics,2023,8(2).
[25] . Research Results from Department of Radiology Update Knowledge of Artificial Intelligence (Artificial Intelligence Algorithm-Based Feature Extraction of Computed Tomography Images and Analysis of Benign and Malignant Pulmonary Nodules)[J]. Robotics & Machine Learning Daily News,2022,(6):97-98.
[26] Yuantong Gao,Yuyang Chen,Yuegui JiangYongchou LiXia ZhangMin LuoXiaoyang WangYang Li. Artificial Intelligence Algorithm-Based Feature Extraction of Computed Tomography Images and Analysis of Benign and Malignant Pulmonary Nodules[J]. Computational intelligence and neuroscience,2022,2022(53):ArticleID5762623-ArticleID5762623
[27] XinCi Liu,Chang Zhao. Research on Image Feature Extraction Algorithm of the Egg and Egg White Protein Thermal Gelation Based on PCA/ICA[J]. Computational intelligence and neuroscience,2022,2022(11):ArticleID1266332-ArticleID1266332.
[28] Ge Hongmei,Song Aibo. A New Low-Rank Structurally Incoherent Algorithm for Robust Image Feature Extraction[J]. mathematics,2022,10(19).
[29] Qin Dong. Path Planning Algorithm Based on Visual Image Feature Extraction for Mobile Robots[J]. Mobile information systems,2022,2022(14):4094472.1-4094472.9.
[30] Jin Yang,Yuxuan Zhao,Shihao Yang,Xinxin Kang,Xinyan Cao,Xixin Cao. Analysis of Feature Extraction and Anti-Interference of Face Image under Deep Reconstruction Network Algorithm[J]. Complexity,2021,2021(a).