A dual-spectral ship small target registration method for wide-area video surveillance

Abstract

The present invention discloses a wide-area video surveillance dual-spectrum ship small target registration method, which belongs to the field of image processing technology. The method includes obtaining image pairs and screening image pairs of qualified quality, coarse registration, extracting candidate areas of ship targets, pairing candidate areas based on spatial distance constraints to obtain spatially close candidate area pairs, performing image interpolation enhancement on the candidate area pairs, constructing a multi-scale image pyramid, multi-scale step-by-step registration optimization, and outputting high-precision registration results. The method solves the technical problems of improving the resolution of dual-spectrum ship small targets and improving the cross-modal registration accuracy in wide-area monitoring scenarios. The present invention improves the focus of small target registration, alleviates the mismatch problem caused by cross-spectral feature heterogeneity, strengthens the local feature expression of low-resolution targets, improves registration robustness, balances computational efficiency and accuracy, and adapts to actual engineering deployment requirements.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a wide-area video surveillance dual-spectrum ship small target registration method. Background Art

[0002] Current bispectral image registration techniques primarily rely on global feature matching (such as SIFT and SURF) or regional similarity optimization (such as mutual information and phase correlation). Some research has attempted to improve cross-modal alignment capabilities by incorporating deep learning (such as CNN-based feature extraction). For maritime surveillance scenarios, existing techniques often directly register the entire image and then filter the target area through filtering or template matching.

[0003] Deficiencies of existing technology:

[0004] Traditional global feature matching in low-resolution small target scenes is prone to uneven distribution of feature points or mismatching due to sparse texture and low signal-to-noise ratio in the target area.

[0005] Due to the differences in imaging mechanisms between visible light and infrared images (such as texture dependence vs. thermal radiation dependence), traditional similarity metrics (such as mutual information) are insufficient in modeling cross-spectral target correlations and have a high mismatch rate.

[0006] Registration methods based on deep learning rely on large amounts of labeled data and have high computational overhead, making them difficult to adapt to the real-time requirements of wide-area monitoring. Lightweight methods (such as regional registration) are susceptible to background interference and have a high missed detection rate. Summary of the Invention

[0007] The purpose of the present invention is to provide a wide-area video surveillance dual-spectrum ship small target registration method, which solves the technical problems of improving the resolution of dual-spectrum ship small targets and improving the cross-modal registration accuracy in wide-area monitoring scenarios.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A wide-area video surveillance dual-spectrum ship small target registration method includes the following steps:

[0010] Step 1: Establish a bispectral ship small target registration system, obtain a bispectral image pair sequence, perform preprocessing and quality assessment calculations on each pair of images in the sequence, select image pairs with quality differences less than a preset threshold as the image pairs to be processed, and construct a set of image pairs to be processed;

[0011] Step 2: performing initial registration processing on any one of the image pairs to be processed in the set of image pairs to be processed, specifically including extracting background feature points and calculating coarse geometric transformation to generate a coarsely registered image pair; the coarsely registered images include a visible light image and an infrared image;

[0012] Step 3: Identify the sea area in the coarsely registered image pair, extract significant candidate regions in the sea area, and generate a set of candidate target regions;

[0013] Step 4: Preset spatial distance constraints, pair the candidate regions in the visible light image and the infrared image, generate candidate region pairs, and construct a set of candidate region pairs with similar spatial positions;

[0014] Step 5: Perform local image enhancement processing on the candidate region pairs, improve the image resolution through multiple interpolation, and construct a multi-scale image pyramid on the enhanced image to enhance the regional features again and generate enhanced multi-scale candidate region pairs;

[0015] Step 6: On the enhanced multi-scale candidate region pairs, according to the multi-scale registration strategy, features are extracted step by step and the registration transformation is optimized to obtain the target region registration result;

[0016] Step 7: Process all the images to be processed in the set of image pairs to be processed according to the methods of steps 2 to 6.

[0017] Preferably, the dual-spectral ship small target registration system includes an image analysis module, a target region extraction module, a constraint pairing module, a feature enhancement module and a fine registration module;

[0018] The image analysis module, target region extraction module, constraint pairing module, feature enhancement module and fine registration module communicate with each other through the Internet.

[0019] Preferably, when executing step 1, the following steps are specifically included:

[0020] Step 1-1: The image analysis module obtains a bispectral image pair sequence and performs image preprocessing on each pair of images in the bispectral image pair sequence;

[0021] Step 1-2: Perform quality evaluation calculations on the preprocessed images, select image pairs with small quality differences as the image pairs to be processed, and construct a set of image pairs to be processed.

[0022] Preferably, when executing step 2, the target region extraction module retrieves a set of image pairs to be processed and processes the image pairs to be processed one by one, specifically including the following steps:

[0023] Step 2-1: The image analysis module extracts background feature points of the image pair to be processed;

[0024] Step 2-2: Match background feature points;

[0025] Step 2-3: Calculate the coarse geometric transformation to generate a coarse registration image pair;

[0026] Step 2-4: Create a mapping table between the coarsely registered image pairs and the image pairs to be processed.

[0027] Preferably, when executing step 3, the following steps are specifically included:

[0028] Step 3-1: The target region extraction module retrieves the coarse registration image pair;

[0029] Step 3-2: performing sea area semantic segmentation on the visible light image in the coarsely registered image pair to obtain a sea area mask;

[0030] Step 3-3: Mapping the sea area mask onto the infrared image in the coarsely registered image pair to perform consistent positioning of the sea area in the bispectral image;

[0031] Step 3-4: In the sea area, use the spectral residual method or local contrast method to extract significant areas as candidate areas and generate a set of candidate areas for ship targets.

[0032] Preferably, when executing step 4, the following steps are specifically included:

[0033] Step 4-1: The constraint matching module retrieves the candidate vessel target area set;

[0034] Step 4-2: Calculate the geometric center of each candidate region in the visible light image and the infrared image;

[0035] Step 4-3: Enumerate all paired combinations of cross-modal candidate regions;

[0036] Step 4-4: Calculate the spatial distance of all the pairing combinations one by one;

[0037] Step 4-5: Compare the result calculated in step 4-4 with the preset distance threshold, filter out the pairing combinations that do not meet the spatial distance constraint, and generate a set of spatially close candidate region pairs.

[0038] Preferably, when executing step 5, the following steps are specifically included:

[0039] Step 5-1: The feature enhancement module retrieves a set of spatially close candidate region pairs;

[0040] Step 5-2: Using the multiple Lánczos interpolation method, perform local image resolution enhancement on each candidate region in the set of spatially close candidate regions to generate an enhanced image region;

[0041] Step 5-3: Construct a multi-scale Gaussian pyramid on the enhanced image area to generate enhanced multi-scale candidate region pairs.

[0042] Preferably, when executing step 6, the following steps are specifically included:

[0043] Step 6-1: The fine registration module retrieves the enhanced multi-scale candidate region pairs;

[0044] Step 6-2: At the lowest resolution scale, extract feature points using a cross-modal robust feature descriptor;

[0045] Step 6-3: Optimize the matching through the RANSAC algorithm to obtain the initial registration transformation;

[0046] Step 6-4: The initial transformation is transferred to a higher resolution scale step by step, and feature matching and transformation optimization are repeated at each level to obtain a high-precision registration result and generate a precisely registered target area image.

[0047] The present invention discloses a method for wide-area video surveillance dual-spectrum small-target ship registration, which solves the technical problems of improving the resolution of dual-spectrum small-target ship registration and cross-modal registration accuracy in wide-area monitoring scenarios. The method collaboratively narrows the search range through sea area semantic segmentation and saliency detection, reduces background interference, and improves the focus of small-target registration. It also introduces a candidate region pairing strategy with geometric center distance constraints to alleviate the mismatching problem caused by cross-spectral feature heterogeneity. It combines interpolation super-resolution with multi-scale pyramid optimization to enhance the local feature expression of low-resolution targets and improve registration robustness. Through staged processing (coarse registration → fine registration) and task decoupling, the method balances computational efficiency and accuracy to meet the needs of actual engineering deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the main flow chart of the present invention;

[0049] Figure 2 is a flow chart of step 4 of the present invention;

[0050] Figure 3 It is a schematic diagram of the data flow between the various modules of the present invention. DETAILED DESCRIPTION

[0051] Depend on Figure 1-Figure 3 The method for dual-spectral ship small target registration in wide-area video surveillance includes the following steps:

[0052] Step 1: Establish a bispectral ship small target registration system, obtain a bispectral image pair sequence, perform preprocessing and quality assessment calculations on each pair of images in the sequence, select image pairs with quality differences less than a preset threshold as the image pairs to be processed, and construct a set of image pairs to be processed;

[0053] When executing step 1, the specific steps include:

[0054] Step 1-1: The image analysis module obtains a bispectral image pair sequence and performs image preprocessing on each pair of images in the bispectral image pair sequence, such as image denoising (mean filtering / median filtering), histogram equalization (contrast enhancement), etc.

[0055] Step 1-2: Perform quality evaluation calculations on the preprocessed images, select image pairs with small quality differences as the image pairs to be processed, and construct a set of image pairs to be processed.

[0056] In this embodiment, the grayscale histograms of the visible light image and the infrared image are calculated respectively, and the grayscale histogram correlation coefficient is further calculated to evaluate the similarity of the image content between the two;

[0057] The grayscale histogram correlation coefficient threshold is pre-set to 0.6 (typical value range is 0.5-0.7). If the threshold is lower than this, it is judged that the quality difference of the bispectral images is large, the possibility of successful registration is low, and the image pair is excluded;

[0058] Only image pairs with grayscale histogram correlation coefficient greater than 0.6 are retained for subsequent processing to ensure the registration success rate.

[0059] Grayscale histogram correlation reflects the consistency of the grayscale distribution of two images and is an effective indicator for measuring the consistency of the content of bispectral images. The higher the correlation, the greater the possibility of successful image matching.

[0060] Step 2: performing initial registration processing on any one of the image pairs to be processed in the set of image pairs to be processed, specifically including extracting background feature points and calculating coarse geometric transformation to generate a coarsely registered image pair; the coarsely registered images include a visible light image and an infrared image;

[0061] When executing step 2, the target region extraction module retrieves the set of image pairs to be processed and processes each of the image pairs to be processed one by one, specifically including the following steps:

[0062] Step 2-1: The image analysis module extracts background feature points of the image pair to be processed;

[0063] Step 2-2: Match background feature points;

[0064] Step 2-3: Calculate the coarse geometric transformation to generate a coarse registration image pair;

[0065] Step 2-4: Create a mapping table between the coarsely registered image pairs and the image pairs to be processed.

[0066] In this embodiment, deep learning image semantic segmentation technology (such as U-Net or DeepLabV3 network) is used to perform sea area semantic segmentation on the registered visible light image, thereby obtaining a high-precision sea area mask.

[0067] The obtained sea area mask is applied to the infrared image to achieve consistent positioning of the sea area in the dual-spectrum image.

[0068] Within a defined sea area, a method based on image spectrum residuals or local contrast saliency detection is used to extract candidate small target regions, resulting in a set of candidate ship target regions. This method effectively highlights the salient features of the ship region while suppressing interference from non-ship regions.

[0069] Background feature points typically originate from areas with stable structures and high cross-modal texture consistency (e.g., horizons and coastlines). We use the ORB feature extraction algorithm with a KNN matcher, and the RANSAC method to eliminate mismatches and estimate the affine transformation matrix for initial coarse registration.

[0070] Step 3: Identify the sea area in the coarsely registered image pair, extract significant candidate regions in the sea area, and generate a set of candidate target regions;

[0071] When executing step 3, the specific steps include:

[0072] Step 3-1: The target region extraction module retrieves the coarse registration image pair;

[0073] Step 3-2: performing sea area semantic segmentation on the visible light image in the coarsely registered image pair to obtain a sea area mask;

[0074] Step 3-3: Mapping the sea area mask onto the infrared image in the coarsely registered image pair to perform consistent positioning of the sea area in the bispectral image;

[0075] Step 3-4: In the sea area, use the spectral residual method or local contrast method to extract significant areas as candidate areas and generate a set of candidate areas for ship targets.

[0076] The spectral residual method is based on retaining high-frequency abnormal areas in Fourier transform to highlight small targets.

[0077] The local contrast method emphasizes the grayscale mutation of a region within its neighborhood.

[0078] In this embodiment, the spectral residual method performs Fourier transform on the image, calculates its spectral residual map, and inversely transforms it to form a saliency map, thereby highlighting areas with unique frequency characteristics; the local contrast method analyzes the local grayscale contrast intensity of each pixel point through a sliding window, which is suitable for capturing small targets with sharp edges.

[0079] Step 4: Preset spatial distance constraints, pair the candidate regions in the visible light image and the infrared image, generate candidate region pairs, and construct a set of candidate region pairs with similar spatial positions;

[0080] When executing step 4, the specific steps include:

[0081] Step 4-1: The constraint matching module retrieves the candidate vessel target area set;

[0082] Step 4-2: Calculate the geometric center of each candidate region in the visible light image and the infrared image;

[0083] Step 4-3: Enumerate all paired combinations of cross-modal candidate regions;

[0084] Step 4-4: Calculate the spatial distance of all the pairing combinations one by one;

[0085] Step 4-5: Compare the result calculated in step 4-4 with the preset distance threshold, filter out the pairing combinations that do not meet the spatial distance constraint, and generate a set of spatially close candidate region pairs.

[0086] In this embodiment, the spatial positions (eg, coordinates of the region centroid or geometric center) of the visible light image candidate region and the infrared image candidate region are calculated respectively.

[0087] Calculate the spatial distance between each candidate region and set the distance threshold to the maximum expected displacement range (typically 20 to 50 pixels).

[0088] In this embodiment, in dual-spectral image registration, due to the different optical properties of infrared and visible light imaging, the position of the same target (such as a ship) in the two band images may be offset to a certain extent. This offset is generally caused by factors such as camera calibration error, imaging perspective deviation, time asynchrony (position change caused by motion), feature extraction deviation caused by spectral differences, etc.

[0089] However, this offset is usually limited, especially in wide-area fixed monitoring, where large-scale misalignment is unlikely. Therefore, this embodiment presets a maximum tolerable spatial offset, namely, the "maximum expected displacement range."

[0090] The typical value used in this embodiment is 20 to 50 pixels, which is set based on the actual system characteristics, image resolution, and target scale experience, such as:

[0091] In common surveillance images, the size of ship targets is within a few dozen pixels, and a single area with a displacement of more than 50 pixels is often not the same target.

[0092] If the camera fields of view are aligned or calibrated, the offset is typically <30 pixels, with 50 pixels being a conservative upper limit.

[0093] In continuous frames or short-term images, the displacement caused by ship movement is also small, especially in wide-area monitoring where the change is small relative to the camera.

[0094] Allowing a certain degree of registration error tolerance makes the matching more robust and can filter out "false positive" region pairs that are obviously in the wrong position.

[0095] Therefore, in this embodiment, to control the error range of cross-modal candidate region pairing, a spatial distance threshold is preset to the maximum expected displacement range, typically ranging from 20 to 50 pixels. This threshold is set empirically based on image resolution, target scale, and bispectral imaging bias. It is used to exclude regions with significant spatial differences (such as ocean waves and strong reflective areas) while retaining candidate region pairs with similar geometric positions, thereby improving pairing accuracy and reducing mismatching rates.

[0096] Eliminate candidate region pairs whose spatial distance obviously exceeds the threshold (such as water surface waves, ripples and other regions with large spectral differences), and only retain candidate region pairs with similar spatial positions to improve the accuracy of candidate region pairs and reduce the mismatch rate.

[0097] In this embodiment, all visible light and infrared candidate region pairs are enumerated, and the Euclidean distance between the geometric centers is calculated. ,by The pixel is used as the threshold, and only region pairs with distances within this range are retained, excluding mismatched combinations with significant spatial deviations, thereby improving the effectiveness of candidate region pairing. The specific calculation method is as follows:

[0098] Let the center of the candidate region of the i-th visible light image be , x and y are coordinates, then the center of the jth infrared image candidate area is , calculate the Euclidean distance between the two:

[0099] ;

[0100] set up ,like , then the candidate region pair is retained, otherwise it is deleted.

[0101] represent the visible light candidate region and infrared candidate region respectively, Indicates the geometric center point.

[0102] Step 5: Perform local image enhancement processing on the candidate region pairs, improve the image resolution through multiple interpolation, and construct a multi-scale image pyramid on the enhanced image to enhance the regional features again and generate enhanced multi-scale candidate region pairs;

[0103] The principle of Gaussian pyramid is to construct a multi-scale sequence of images through Gaussian filtering + downsampling, which can be used to optimize matching layer by layer in registration.

[0104] When executing step 5, the specific steps include:

[0105] Step 5-1: The feature enhancement module retrieves a set of spatially close candidate region pairs;

[0106] Step 5-2: Using the multiple Lánczos interpolation method, perform local image resolution enhancement on each candidate region in the set of spatially close candidate regions to generate an enhanced image region;

[0107] Step 5-3: Construct a multi-scale Gaussian pyramid on the enhanced image area to generate enhanced multi-scale candidate region pairs.

[0108] In this embodiment, the Lánczos interpolation method is used to perform 2-fold, 4-fold, and 8-fold interpolation magnification on each candidate region to improve the image resolution of the local region.

[0109] A multi-scale Gaussian pyramid is constructed on the enhanced local area, including the original scale, 2x scale, 4x scale and 8x scale, to improve the number and quality of feature points in the candidate area.

[0110] After feature enhancement processing, the number of feature points in the local area increases to 2 to 3 times that before enhancement (typically from 10 to 20 to 20 to 50), improving the reliability of subsequent fine alignment.

[0111] In this embodiment, preliminary region registration is first performed at the lowest resolution scale (such as the original scale), and feature points are extracted using a cross-modal robust feature descriptor (such as Census transform).

[0112] First, preliminary region registration is performed at the lowest resolution scale (such as the original scale), and feature points are extracted using a cross-modal robust feature descriptor (such as Census transform).

[0113] The low-scale registration transformation matrix is passed to the next scale (2 times the scale) as the initial transformation parameter, and the registration transformation matrix is refined step by step.

[0114] On the original scale (highest resolution) image, the feature matching results are further refined to complete the final high-precision region registration. The final target region registration error is less than 1 pixel, achieving a high-precision registration effect.

[0115] In this embodiment, after interpolating and magnifying the candidate region, a pyramid is constructed consisting of four layers: the original scale, 2×, 4×, and 8×. Each layer is filtered using a Gaussian blur kernel and downsampled. This pyramid structure helps to stably capture coarse features at low resolution and achieve fine optimization at high resolution.

[0116] Step 6: On the enhanced multi-scale candidate region pairs, according to the multi-scale registration strategy, features are extracted step by step and the registration transformation is optimized to obtain the target region registration result;

[0117] When executing step 6, the specific steps include:

[0118] Step 6-1: The fine registration module retrieves the enhanced multi-scale candidate region pairs;

[0119] Step 6-2: At the lowest resolution scale, extract feature points using a cross-modal robust feature descriptor;

[0120] Step 6-3: Optimize the matching through the RANSAC algorithm to obtain the initial registration transformation;

[0121] Step 6-4: The initial transformation is transferred to a higher resolution scale step by step, and feature matching and transformation optimization are repeated at each level to obtain a high-precision registration result and generate a precisely registered target area image.

[0122] In this embodiment, the Census descriptor is used to maintain structural consistency across different modalities, making it suitable for matching infrared and visible light images. RANSAC iteratively selects feature point pairs and calculates a transformation matrix to eliminate mismatched point pairs, improving robustness.

[0123] Step 7: Process all the images to be processed in the set of image pairs to be processed according to the methods of steps 2 to 6.

[0124] The dual-spectral ship small target registration system includes an image analysis module, a target area extraction module, a constraint pairing module, a feature enhancement module and a fine registration module;

[0125] The image analysis module, target region extraction module, constraint pairing module, feature enhancement module and fine registration module communicate with each other through the Internet.

[0126] The image analysis module is used to obtain a sequence of bispectral image pairs and perform image preprocessing, quality assessment and initial coarse registration on the image pairs;

[0127] The target region extraction module is used to perform semantic segmentation on the registered image pairs, identify the sea area, and extract the salient candidate regions in the sea area;

[0128] The constraint pairing module is used to perform cross-modal pairing of the candidate regions based on spatial constraint rules to generate a set of spatially close candidate region pairs;

[0129] The feature enhancement module is used to perform image enhancement and multi-scale image construction on candidate region pairs to improve the matching of regional images;

[0130] The fine registration module is used to perform feature matching and registration optimization on a multi-scale image pyramid and output high-precision target registration results.

[0131] The present invention discloses a method for wide-area video surveillance dual-spectrum small-target ship registration, which solves the technical problems of improving the resolution of dual-spectrum small-target ship registration and cross-modal registration accuracy in wide-area monitoring scenarios. The method collaboratively narrows the search range through sea area semantic segmentation and saliency detection, reduces background interference, and improves the focus of small-target registration. It also introduces a candidate region pairing strategy with geometric center distance constraints to alleviate the mismatching problem caused by cross-spectral feature heterogeneity. It combines interpolation super-resolution with multi-scale pyramid optimization to enhance the local feature expression of low-resolution targets and improve registration robustness. Through staged processing (coarse registration → fine registration) and task decoupling, the method balances computational efficiency and accuracy to meet the needs of actual engineering deployment.

Claims

1. A dual-spectral ship small target registration method for wide-area video surveillance, characterized by: The steps include: Step 1: Establish a bispectral ship small target registration system, obtain a bispectral image pair sequence, preprocess and evaluate the quality of each image pair in the sequence, select image pairs with quality differences less than a preset threshold as the image pairs to be processed, and construct a set of image pairs to be processed; Step 2: performing initial registration processing on any one of the image pairs to be processed in the set of image pairs to be processed, specifically including extracting background feature points and calculating coarse geometric transformation to generate a coarsely registered image pair; the coarsely registered images include a visible light image and an infrared image; Step 3: Identify the sea area in the coarsely registered image pair, extract significant candidate regions in the sea area, and generate a set of candidate target regions; Step 4: Preset spatial distance constraints, pair the candidate regions in the visible light image and the infrared image, generate candidate region pairs, and construct a set of candidate region pairs with similar spatial positions; Step 5: Perform local image enhancement processing on the candidate region pairs, improve the image resolution through multiple interpolation, and construct a multi-scale image pyramid on the enhanced image to enhance the regional features again and generate enhanced multi-scale candidate region pairs; Step 6: On the enhanced multi-scale candidate region pairs, according to the multi-scale registration strategy, features are extracted step by step and the registration transformation is optimized to obtain the target region registration result; Step 7: Process all the images to be processed in the set of image pairs to be processed according to the methods of steps 2 to 6.

2. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 1, characterized in that: The dual-spectral ship small target registration system includes an image analysis module, a target area extraction module, a constraint pairing module, a feature enhancement module and a fine registration module; The image analysis module, target region extraction module, constraint pairing module, feature enhancement module and fine registration module communicate with each other through the Internet.

3. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 1, the specific steps include: Step 1-1: The image analysis module obtains a bispectral image pair sequence and performs image preprocessing on each pair of images in the bispectral image pair sequence; Step 1-2: Perform quality assessment on the preprocessed images, select image pairs with small quality differences as the image pairs to be processed, and construct a set of image pairs to be processed.

4. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 2, the target region extraction module retrieves the set of image pairs to be processed and processes each of the image pairs to be processed one by one, specifically including the following steps: Step 2-1: The image analysis module extracts background feature points of the image pair to be processed; Step 2-2: Match background feature points; Step 2-3: Calculate the coarse geometric transformation to generate a coarse registration image pair; Step 2-4: Create a mapping table between the coarsely registered image pairs and the image pairs to be processed.

5. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 3, the specific steps include: Step 3-1: The target region extraction module retrieves the coarse registration image pair; Step 3-2: performing sea area semantic segmentation on the visible light image in the coarsely registered image pair to obtain a sea area mask; Step 3-3: Mapping the sea area mask onto the infrared image in the coarsely registered image pair to perform consistent positioning of the sea area in the bispectral image; Step 3-4: In the sea area, use the spectral residual method or local contrast method to extract significant areas as candidate areas and generate a set of candidate areas for ship targets.

6. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 4, the specific steps include: Step 4-1: The constraint matching module retrieves the candidate vessel target area set; Step 4-2: Calculate the geometric center of each candidate region in the visible light image and the infrared image; Step 4-3: Enumerate all paired combinations of cross-modal candidate regions; Step 4-4: Calculate the spatial distance of all the pairing combinations one by one; Step 4-5: Compare the result calculated in step 4-4 with the preset distance threshold, filter out the pairing combinations that do not meet the spatial distance constraint, and generate a set of spatially close candidate region pairs.

7. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 5, the specific steps include: Step 5-1: The feature enhancement module retrieves a set of spatially close candidate region pairs; Step 5-2: Using the multiple Lánczos interpolation method, perform local image resolution enhancement on each candidate region in the set of spatially close candidate regions to generate an enhanced image region; Step 5-3: Construct a multi-scale Gaussian pyramid on the enhanced image area to generate enhanced multi-scale candidate region pairs.

8. The wide-area video surveillance dual-spectrum ship small target registration method according to claim 2, characterized in that: When executing step 6, the specific steps include: Step 6-1: The fine registration module retrieves the enhanced multi-scale candidate region pairs; Step 6-2: At the lowest resolution scale, extract feature points using a cross-modal robust feature descriptor; Step 6-3: Optimize the matching through the RANSAC algorithm to obtain the initial registration transformation; Step 6-4: The initial transformation is transferred to a higher resolution scale step by step, and feature matching and transformation optimization are repeated at each level to obtain a high-precision registration result and generate a precisely registered target area image.

Images