Underwater intercept net image enhancement method and system

Abstract

The application provides an underwater net image enhancement method and system, the method comprising: obtaining a multi-angle polarization image under active polarization illumination to obtain a linear polarization degree image DoLP; taking one of the angle images as a main channel to obtain a background scattering field, removing the background scattering field from the main channel image by using the DoLP as a weighting coefficient to obtain an extinction image; performing multi-scale decomposition on one of the angle images to extract at least two high-frequency detail layers of different scales, generating an adaptive weight negatively correlated with the DoLP value according to the image DoLP, and injecting the weighted multi-scale high-frequency detail layer into the extinction image to obtain a detail enhancement image; taking one of the angle images as a guide image to perform guided filtering on the detail enhancement image to obtain a composite correction image, and performing local contrast adaptive stretching on the composite correction image to obtain a final enhancement image. The application improves the recognition degree of the underwater net image through despeckling, detail enhancement and filtering stretching.

Description

Technical Field

[0001] This invention relates to the fields of underwater photoelectric imaging and computational imaging technology, and more specifically, to an image enhancement method and system for underwater interception nets. Background Technology

[0002] In deep-sea or high-turbidity underwater environments, natural light is insufficient, necessitating active illumination. However, existing technologies suffer from the following shortcomings: Regarding backscattering interference, active light, after being scattered by suspended particles, forms a strong background halo, obscuring the fine textures of targets such as interceptor nets and reducing image clarity and recognizability; Regarding uneven illumination and limited dynamic range, high-power illumination at close range easily results in overly bright central areas and darker edge areas. If global equalization is used, it often leads to truncation of details in bright areas and amplification of noise in dark areas; Regarding sensor artifacts, focal plane polarization cameras are prone to producing checkerboard array artifacts at high-contrast edges, interfering with the recognition of net ropes and mesh structures; Regarding feature loss, traditional defogging or extinction methods, while suppressing the scattering background, often simultaneously smooth the high-frequency edges of the target, resulting in blurred lines and incomplete structural information of the interceptor net. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an underwater interception net image enhancement method and system, addressing the problems existing in the prior art.

[0004] The technical solution adopted by this invention to solve its technical problem is: a method for enhancing underwater interception net images, comprising the following steps: Step S1: Acquire multi-angle polarization images under active polarization illumination, and calculate the linear polarization degree image DoLP; take one angle image as the main channel, acquire the background scattering field, and use DoLP as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image; Step S2: Perform multi-scale decomposition on one of the angle images, extract at least two high-frequency detail layers at different scales, generate adaptive weights that are negatively correlated with the DoLP value based on the image DoLP, and inject the weighted multi-scale high-frequency detail layers into the extinction image to obtain a detail-enhanced image. Step S3: Using one of the angle images as a guide image, guide filtering is applied to the detail enhancement image to obtain a synthetic correction image, which is then subjected to local contrast adaptive stretching to obtain the final enhanced image.

[0005] Further, the multi-angle polarization image includes polarization grayscale images at four angles: 0°, 45°, 90°, and 135°; in step S1, the calculation of the linear polarization degree image DoLP includes: Calculate the total intensity image: ; Calculate the linear polarization components: ; ; Calculate polarization intensity: ; Calculate the degree of linear polarization: DoLP = clip( (0.01, 0.99); in, , , , These are the pixel values ​​of the images at polarization directions of 0°, 45°, 90°, and 135°, respectively. It is a very small positive number, used to prevent division by zero.

[0006] Further, in step S1, the step of using one angle image as the main channel to obtain the background scattering field, and using DoLP as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image includes: 90° channel image I 90 As the main channel, a Gaussian low-pass filter is used to obtain the background scattering field. B field : B field =G σlarge ( I 90 ); Among them, G σlarge It is a two-dimensional Gaussian low-pass filter used to extract the slowly changing background halo intensity distribution; Using the linear polarization degree image DoLP as a weighting coefficient, a soft extinction method is employed to remove the background scattering field from the main channel image: Where ⊙ represents pixel-by-pixel multiplication, k is a preset extinction intensity coefficient, and DoLP is the linear polarization degree image; right I descat The image is cropped pixel by pixel to the range of [0, 255] to obtain an extinction image.

[0007] Further, in step S2, the multi-scale decomposition of one angle image to extract at least two high-frequency detail layers at different scales includes: 90° channel image I 90 As a decomposition object: Extracting micro-detail layers D micro = I 90 -G σsmall (I 90 ), 5×5 cores, ( ≈1.1); Extracting the meso-detail layer D macro = I 90 -G σmed ( I 90 ), 15×15, ( ≈2.6); Among them, G σsmall and G σmed These are the first Gaussian low-pass filter and the second Gaussian low-pass filter, respectively, with the size of the second Gaussian kernel being larger than the size of the first Gaussian kernel. denoted as the standard deviation of the Gaussian kernel.

[0008] Further, in step S2, generating adaptive weights that are negatively correlated with the DoLP value based on the image DoLP includes: Generate adaptive weights based on image DoLP. W : W =1.0+(1.0-DoLP); The adaptive weights are negatively correlated with the DoLP value.

[0009] Further, in step S2, injecting the weighted multi-scale high-frequency detail layer into the extinction image to obtain a detail-enhanced image includes: The weighted multi-scale high-frequency detail layer is injected into the extinction image. I descat : I enh = I descat +(0.5· D micro +0.8· D macro )⊙W; Where ⊙ represents pixel-by-pixel multiplication, and 0.5 and 0.8 are the injection intensity coefficients for the micro detail layer and the meso detail layer, respectively; Will I enh Output as an image with enhanced detail.

[0010] Further, in step S3, the step of using one of the angle images as a guide image to perform guided filtering on the detail enhancement image to obtain the synthetic correction image includes: 90° channel image I 90Normalize to the range [0, 255] to obtain the guide graph. G guide ; Enhance the details of the image I enh Normalize the image to the range [0, 255] to obtain the image to be filtered. I src ; The image to be filtered is filtered using a guided filtering algorithm, with the guided image as the filter. G guide As a guide, a synthetic correction image is obtained. I guided : I guided =GuidedFilter( G guide , I src r=10 η =80); Where r is the radius of the filtering window. η This is a regularization parameter used to control the smoothing intensity.

[0011] Further, in step S3, the locally contrast-adaptive stretching is performed: For the synthetic correction map I guided Apply a contrast-restricted adaptive histogram equalization.

[0012] Furthermore, the parameters for the contrast-limited adaptive histogram equalization are set as follows: contrast limit threshold ClipLimit=2.0, block size gridSize=8×8 pixels; the processed image is output as the final enhanced image.

[0013] The present invention also provides an underwater interception net image enhancement system, including a processor and a memory storing a computer program, wherein the processor, when executing the computer program, implements the steps of any of the above-described underwater interception net image enhancement methods.

[0014] The beneficial effects of this invention are that it provides an underwater interception net image enhancement method, comprising: Step S1: acquiring a multi-angle polarization image under active polarization illumination, and calculating the degree of linear polarization image DoLP; using one angle image as the main channel, acquiring the background scattering field, and using DoLP as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image; Step S2: performing multi-scale decomposition on one angle image, extracting at least two high-frequency detail layers at different scales, generating adaptive weights negatively correlated with the DoLP value based on the image DoLP, and injecting the weighted multi-scale high-frequency detail layers into the extinction image to obtain a detail-enhanced image; Step S3: using one angle image as a guide image, performing guided filtering on the detail-enhanced image to obtain a synthetic correction image, and then performing local contrast adaptive stretching to obtain the final enhanced image. This invention improves the image recognition of underwater interception nets through DoLP weighted descattering, negative correlation injection of detail to enhance edges, guided filtering, and contrast stretching. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a logic flowchart of an embodiment of an underwater interception net image enhancement method; Figure 2 This is a logic flowchart of another embodiment of the underwater interception net image enhancement method. Detailed Implementation

[0016] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the invention are now described in detail with reference to the accompanying drawings. In the following description, specific details such as particular structures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known devices, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0017] like Figure 1 As shown, Figure 1 This is a flowchart of the underwater interception net image enhancement method.

[0018] This invention provides a method for enhancing images of underwater interception nets, comprising: Step S1: Acquire multi-angle polarization images under active polarization illumination, and calculate the linear polarization degree image DoLP; take one angle image as the main channel, acquire the background scattering field, and use DoLP as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image; In this step, it's important to note that the DoLP value in the linear polarization degree image (DoLP) reflects the polarization characteristics of each pixel: a larger DoLP value indicates stronger polarization, corresponding to areas of scattered light from water bodies; a smaller DoLP value indicates weaker polarization, corresponding to areas of polymer materials such as intercepting nets. Utilizing this physical characteristic, this step uses DoLP as a weighting coefficient to distinguish between water mist areas and net / rope areas, achieving differentiated suppression of background scattered light. Specifically, areas with larger DoLP values ​​(water mist areas) have more background scattered light subtracted from the main channel image when used as a weighting coefficient, resulting in stronger extinction intensity; areas with smaller DoLP values ​​(net / rope areas) have less background scattered light subtracted when used as a weighting coefficient, resulting in weaker extinction intensity, thus preserving the net / rope target while removing water mist.

[0019] Step S2: Perform multi-scale decomposition on one of the angle images, extract at least two high-frequency detail layers at different scales, generate adaptive weights that are negatively correlated with the DoLP value based on the image DoLP, and inject the weighted multi-scale high-frequency detail layers into the extinction image to obtain a detail-enhanced image. In this step, it's important to note that the high-frequency detail layer is extracted using a "subtract the blurred image from the original image" method, with different scales corresponding to different granularities of texture information. The microscale (small kernel) corresponds to fiber-level details, while the mesoscale (medium kernel) corresponds to the mesh structure. The adaptive weights are negatively correlated with the DoLP value; that is, regions with smaller DoLP values ​​(like rope areas) receive greater enhancement weights, thus primarily injecting details into suspected rope areas. Conversely, regions with larger DoLP values ​​(like water mist areas) receive smaller enhancement weights, avoiding unnecessary noise amplification in water mist areas. Through this differentiated enhancement mechanism, targeted enhancement of the rope target texture is achieved.

[0020] Step S3: Using one of the angle images as a guide image, guide filtering is applied to the detail enhancement image to obtain a synthetic correction image, which is then subjected to local contrast adaptive stretching to obtain the final enhanced image.

[0021] In this step, it's important to note that guided filtering uses a guide image (one angle of the image) as a reference structure to smooth the detail-enhancing image while preserving edges. Locations in the guide image with drastic changes (net edges) automatically have their filter weights reduced, thus protecting the true edges of the net from being smoothed out; locations in the guide image with gradual changes (artifact areas, flat areas) automatically have their filter weights increased, thus smoothing out the checkerboard array artifacts inherent in the focal plane sensor. Local contrast adaptive stretching divides the image into several local blocks and performs contrast stretching on each block individually, achieving uniform brightness distribution across the entire image and making the mesh clearly discernible.

[0022] Furthermore, the multi-angle polarization image includes polarization grayscale images at four angles: 0°, 45°, 90°, and 135°; in step S1, the linear polarization degree image DoLP is calculated, which includes: calculating the total intensity image. ; represents the average intensity of the four polarization states at this pixel, used as the denominator for subsequent normalization; calculate the linear polarization components: Intensity difference in the direction of (approximately horizontal / vertical polarization axis); Intensity difference along (approximately diagonal polarization axis) directions; Calculate polarization intensity: ; Calculate the degree of linear polarization: DoLP = clip( (0.01, 0.99) where, , , , These are the pixel values ​​of the images at polarization directions of 0°, 45°, 90°, and 135°, respectively. The value is a very small positive number to prevent division by zero; clip limits DoLP to [0.01, 0.99] to avoid extreme values ​​causing distortion in subsequent weights. The larger the DoLP, the stronger the polarization of the pixel (mostly scattering / background); the smaller the DoLP, the weaker the polarization (high molecular targets such as interception nets often have low DoLP).

[0023] Further, in step S1, using one angle image as the main channel, the background scattering field is obtained, and DoLP is used as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image, including: using the 90° channel image I 90 As the main channel, a Gaussian low-pass filter is used to obtain the background scattering field. B field : B field =G σlarge ( I 90 ); where G σlarge A two-dimensional Gaussian low-pass filter is used to extract the slowly varying background halo intensity distribution; the DoLP linear polarization image is used as a weighting coefficient, and a soft extinction method is employed to remove the background scattering field from the main channel image. Where ⊙ represents pixel-by-pixel multiplication, k is a preset extinction intensity coefficient, and DoLP is the linear polarization degree image; for I descat The image is cropped pixel by pixel to the range of [0, 255] to obtain an extinction image.

[0024] B fieldThe physical meaning of is the slowly varying background halo intensity; that is, it is not high-frequency texture information in the image, but rather a spatially varying background component formed by backscattering of active light sources by suspended particles in the water. The choice of kernel size for Gaussian low-pass filtering directly affects B field The estimation accuracy depends on the kernel size. A larger kernel size results in a smoother estimation of the background field, but may blur true background changes. A smaller kernel size preserves more detail in the background field, but may misclassify low-frequency components of the netting as background. In this embodiment, the 101×101 pixel kernel size is an optimized choice for typical underwater interception net imaging scenarios (netting diameter approximately 2-5 pixels, grid period approximately 20-50 pixels). If the background halo changes relatively smoothly (e.g., in scenes with clear water and uniform lighting), a larger kernel size (e.g., 151×151 pixels) can be used to further suppress residual textures in the background. If the background halo changes rapidly (e.g., in scenes with turbid water and uneven lighting), a smaller kernel size (e.g., 51×51 pixels) can be used to better follow background changes. The value of k can be adjusted within the range of 0.3 to 1.0, preferably 0.5 to 0.8. A larger k value results in greater dehazing intensity, but an excessively large k value may lead to over-weakening of the net area; a smaller k value results in less dehazing intensity, but an excessively small k value may lead to water mist residue. In the embodiment, k=0.7 is a compromise value optimized through experiments in various underwater scenarios, achieving a good balance between dehazing and target preservation. The purpose of preventing full-black truncation is to prevent the display from being completely black after subtraction overflow becomes negative; this invention does not perform additional global normalization to avoid changing the local contrast relationship. This is because global normalization changes the relative brightness ratio between different areas of the image, which may cause the originally darker net area to be over-brightened and lose its realism, or cause the originally brighter water mist area to be darkened and obscure the target.

[0025] It is understood that the main channel is not limited to the 90° channel. In other embodiments, other angle channels that are closest to orthogonal to the backscattering polarization direction can be selected as the main channel based on the actual polarization angle configuration. For example, in some polarization camera configurations, the definitions of the polarization directions of 0° and 90° may differ from those in this invention. In this case, the angle closest to orthogonal to the backscattering direction can be selected as the main channel. The key principle is that the main channel should be selected based on the polarization direction least affected by backscattering, so that the subsequent extinction operation can achieve the optimal signal-to-noise ratio.

[0026] Further, in step S2, multi-scale decomposition is performed on one of the angle images to extract at least two high-frequency detail layers at different scales, including: using the 90° channel image I 90 As a decomposition object: Extracting micro-detail layers D micro = I90 -G σsmall ( I 90 ), 5×5 ( ≈1.1); Extracting the meso-detail layer D macro = I 90 -G σmed ( I 90 ), 15×15 ( ≈2.6); Among them, G σsmall and G σmed These are the first Gaussian low-pass filter and the second Gaussian low-pass filter, respectively, with the size of the second Gaussian kernel being larger than the size of the first Gaussian kernel. denoted as the standard deviation of the Gaussian kernel.

[0027] Specifically, each scale layer is a high-pass residual of the "original image minus blurred image," extracting high-frequency information that was smoothed away by Gaussian filtering. The kernel size determines the scale of detail extraction: the smaller the kernel, the finer the details preserved; the larger the kernel, the coarser the details preserved. Microscopic detail layer D micro It captures the texture information of individual fibers in the rope, which appear as fine lines 2-3 pixels wide in the spatial domain, with a high frequency; the meso-detail layer... D macro The captured information is the mesh structure, specifically the polygonal openings formed by the intersecting netting ropes. These features are approximately 5–15 pixels wide and have a relatively low frequency (though still high frequency relative to low frequency). This embodiment employs both microscopic and mesoscopic scales based on the multi-scale characteristics of the interceptor net target: the interceptor net contains both fiber-level fine textures (determining the material quality of the netting ropes) and mesh-level macroscopic structures (determining the shape and distribution of the mesh openings). Both scales together comprehensively describe the visual characteristics of the interceptor net, and neither can be omitted.

[0028] Further, in step S2, generating adaptive weights that are negatively correlated with the DoLP value based on the image DoLP includes: generating adaptive weights based on the image DoLP. W : W =1.0+(1.0-DoLP); Among them, the adaptive weights are negatively correlated with the DoLP value.

[0029] Specifically, when DoLP approaches 0 (pure mesh region), W approaches 2.0, resulting in the maximum enhancement; when DoLP approaches 1 (pure water mist region), W approaches 1.0, resulting in the minimum enhancement (only retaining the basic injection intensity). The lower limit of the weight is 1.0, ensuring a basic level of detail injection even in the pure water mist region (DoLP close to 1), guaranteeing that the most basic texture of the entire image is not completely lost. The upper limit of the weight is 2.0, preventing the detail injection from being infinitely amplified even in the pure mesh region (DoLP close to 0), thus avoiding texture distortion and noise amplification caused by over-enhancement. The 1.0 bias term design ensures both the physical meaning of the weights (the minimum enhancement weight is 1.0, i.e., neither amplification nor attenuation) and simplifies the weight normalization process (no additional normalization operation is required). It is understandable that the specific mapping relationship of the adaptive weights is not limited to a linear mapping. W =1.0 + (1.0 - DoLP). Other forms of negative correlation mapping functions can also be used in other embodiments. As long as the negative correlation relationship of lower DoLP and higher weight is satisfied, it is within the protection scope of this invention. Different mapping functions are suitable for different types of images and enhancement needs. Linear mapping is simple and intuitive and suitable for most general scenarios; nonlinear mapping can provide more aggressive enhancement in the rope region (low DoLP) or provide smoother enhancement in the transition region, which is suitable for scenarios with special requirements for specific textures.

[0030] Further, in step S2, injecting the weighted multi-scale high-frequency detail layer into the extinction image to obtain the detail-enhanced image includes: injecting the weighted multi-scale high-frequency detail layer into the extinction image. I descat : I enh = I descat +(0.5· D micro +0.8· D macro )⊙W; where ⊙ represents pixel-wise multiplication, and 0.5 and 0.8 are the injection intensity coefficients for the micro-detail layer and meso-detail layer, respectively; I enh Output as an image with enhanced detail.

[0031] Specifically, different injection intensity coefficients are used for the micro-detail layer and the meso-detail layer, based on the different contributions of the two types of details to visual perception: micro-details (fiber texture) contribute relatively little to visual sharpness, and excessive micro-detail injection may result in an unnatural "grainy" appearance in the image, so a smaller coefficient of 0.5 is used; meso-details (mesh structure) contribute more to mesh recognition and require stronger enhancement, so a larger coefficient of 0.8 is used. This differentiated coefficient setting reflects respect for the hierarchical nature of target features. Mesh structure is the most important visual cue for identifying the interceptor mesh, while fiber texture is a secondary material feature.

[0032] The injection intensity coefficient can be adjusted according to the actual enhancement needs: if it is necessary to enhance the microfiber texture (such as for material recognition or classification tasks), the micro-layer coefficient can be increased (e.g., 0.6–0.8); if it is necessary to enhance the mesh structure (such as for target detection or localization tasks), the meso-layer coefficient can be increased (e.g., 0.9–1.2). The ratio between the two coefficients can also be adjusted according to the thickness of the target mesh. The thinner the mesh, the more important the micro-details, and the micro-coefficient can be appropriately increased; the thicker the mesh, the more obvious the mesh structure, and the meso-coefficient can be appropriately increased. It is recommended that the sum of the injection coefficients be controlled between 1.0 and 2.0. Too large a sum may lead to over-sharpening of the image, while too small a sum will result in an insignificant enhancement effect.

[0033] It is important to emphasize that the injection in this step is performed pixel-by-pixel, meaning that the enhancement amount at each pixel location depends on the DoLP weight W of that pixel and the detail layer value at that location. This pixel-by-pixel injection mechanism ensures the spatial adaptability of the enhancement, with the rope area receiving more detail injection and the water mist area receiving less detail injection, achieving a refined "on-demand enhancement" process.

[0034] Further, in step S3, using one of the angle images as a guide image, the detail enhancement image is subjected to guided filtering to obtain the synthetic correction image, including: using the 90° channel image I 90 Normalize to the range [0, 255] to obtain the guide graph. G guide Enhance image details I enh Normalize the image to the range [0, 255] to obtain the image to be filtered. I src A guided filtering algorithm is used to filter the image to be filtered, in order to guide the image. G guide As a guide, a synthetic correction image is obtained. I guided : I guided =GuidedFilter( Gguide , I src r=10 η =80) where r is the radius of the filtering window, η This is a regularization parameter used to control the smoothing intensity.

[0035] Specifically, the guided filtering correction stage uses an angle image from the original polarization image as a structural reference. Specifically, the 90° channel image is normalized to the range of 0 to 255 using minimum and maximum values ​​to obtain a guide image, which provides a reliable prior for edge structure. Simultaneously, the detail enhancement image is also normalized to the range of 0 to 255 to obtain the image to be filtered. Then, a guided filtering algorithm is used to filter the image to be filtered, using the guide image as a guide to obtain a synthesized correction image. The filtering window radius is 10 pixels, and the regularization parameter is 80 to control the smoothing intensity. During this process, the weight of the guided filter automatically decreases at the edges of the guide image, achieving anisotropic edge-preserving smoothing, i.e., less smoothing along the edge direction and more smoothing across the edge direction. Its edge-preserving mechanism is that the guided filter outputs an approximation of the linear transformation of the image to be filtered under the guidance of the guide image within each local window. When the guide image changes significantly (such as at the edge of a net), the linear coefficient changes significantly, and the filtering hardly changes the image to be filtered at that location, thus forming a "forced protection" effect. In flat areas and periodic artifact areas, the guide image changes less and is effectively smoothed. Therefore, this step can eliminate the inherent 2×2 periodic checkerboard pattern artifacts of the focal plane sensor while ensuring that the gradient of the intercept net lines remains consistent with the guide map, and the geometric skeleton is not flattened. The protection effect can be quantitatively evaluated by the edge gradient retention rate (Sobel gradient correlation coefficient before and after filtering) or the contrast retention ratio of the net area.

[0036] Further, in step S3, local contrast adaptive stretching is performed on the synthesized correction image. I guided Apply a contrast-adaptive histogram equalization constraint.

[0037] This processing avoids overexposure of bright spots and amplification of noise in dark areas caused by global equalization. Global histogram equalization maps the grayscale distribution of the entire image to the entire grayscale range. In the presence of large bright or dark areas, this can cause the histogram peaks to be overstretched, resulting in an unnatural visual appearance. CLAHE overcomes this defect fundamentally through local processing and amplitude limiting, making the brightness distribution of the entire image more uniform and the mesh easier to identify.

[0038] Furthermore, the parameters for limiting the contrast adaptive histogram equalization are set as follows: contrast limit threshold ClipLimit=2.0, block size gridSize=8×8 pixels; the processed image is output as the final enhanced image.

[0039] Specifically, ClipLimit controls the maximum histogram clipping range; a larger value results in greater contrast stretching, but excessively large values ​​may amplify noise. GridSize controls the size of local blocks; a smaller value provides stronger local adaptability, but excessively small values ​​may lead to unnatural transitions between blocks. In this example, ClipLimit=2.0 means that the maximum amplification factor for each gray level in the histogram is twice the average number of pixels per level. This value effectively enhances contrast while keeping noise amplification within a reasonable range. The 8×8 pixel block size is a trade-off between the mesh size (typically 20–50 pixels) and the rope width (2–5 pixels) in underwater interception net images. The block size is much larger than the rope width but smaller than the mesh period, ensuring that each block contains sufficient gray-level distribution information for histogram statistics.

[0040] It is understood that the above parameters can be adjusted according to the actual image resolution and contrast distribution: for high-resolution images (such as 4K resolution), larger blocks (such as 16×16 or 32×32 pixels) can be used to reduce computational load and ensure statistical reliability; for low-resolution images (such as VGA resolution), smaller blocks (such as 4×4 pixels) can be used to improve local adaptability. For low-contrast scenes (such as extremely turbid water or severely insufficient lighting), a larger ClipLimit value (such as 2.5 to 3.0) can be used to obtain a stronger contrast stretching effect; for high-contrast scenes (such as relatively clear water or sufficient lighting), a smaller ClipLimit value (such as 1.0 to 1.5) can be used to avoid over-enhancement. The specific values ​​in the examples are not limited to those in the embodiments.

[0041] In other embodiments, additional denoising steps (such as median filtering, BM3D, etc.) can be added after CLAHE processing to further optimize image quality. CLAHE may slightly amplify existing noise when processing low-contrast areas; adding additional denoising steps can eliminate this side effect, resulting in a cleaner final output image. The added noise suppression step preferably employs an edge-preserving denoising algorithm (such as guided filtering, bilateral filtering, BM3D, etc.) to maintain the sharpness of the net edges while denoising.

[0042] This processing avoids overexposure of bright spots and amplification of noise in dark areas caused by global equalization. Global histogram equalization maps the grayscale distribution of the entire image to the entire grayscale range. In the presence of large bright or dark areas, this can cause the histogram peaks to be overstretched, resulting in an unnatural visual appearance. CLAHE overcomes this defect fundamentally through local processing and amplitude limiting, making the brightness distribution of the entire image more uniform and the mesh easier to identify.

[0043] like Figure 2 As shown, Figure 2This is a logic flowchart of another embodiment, which is related to... Figure 1 The core processing logic is the same, the difference lies in... Figure 1 The three steps are broken down into five stages.

[0044] Input the four-channel original polarization image group (including grayscale images of four polarization directions: 0°, 45°, 90°, and 135°), and proceed to stage (1) full component polarization information calculation and energy reconstruction. This stage is implemented through three sub-steps: Step 1.1 Full component energy field reconstruction, calculate the total intensity image. This provides the defogging light source as the brightness reference for subsequent defogging operations; Step 1.2 Stokes vector solution calculates the linear polarization components. and The original physical structure parameters are provided; Step 1.3 Calculates the degree of linear polarization. Based on the calculated image DoLP, a source of detail is provided to guide subsequent dehazing and enhancement operations.

[0045] Entering stage (2) Local physical extinction background correction. Using the 90° channel image I90 as the main channel, the 90° polarization background field is extracted. B field A large-scale smoothing process is performed on the 90° channel image using a two-dimensional Gaussian low-pass filter to obtain a slowly varying background halo intensity distribution; background soft extinction correction is then performed based on DoLP, using the formula... Background scattering is removed from the main channel image, where a higher DoLP reduces the background more and a lower DoLP reduces the background less to preserve the target, resulting in an extinction image.

[0046] Entering stage (3): DoLP-guided multi-scale high-frequency detail extraction and feature synthesis. A parallel scale decomposition model is used to filter the 90° channel image with Gaussian kernels of 5×5 and 15×15 respectively, and the micro-detail layer is extracted by subtracting the blur from the original image. D micro (5×5 scale) and meso-detail layer D macro (15×15 scale); simultaneously construct adaptive weights based on image DoLP. W =1.0+(1.0-DoLP) to achieve differentiated control of high enhancement in low DoLP (net rope region) and low enhancement in high DoLP (water mist region); finally, the weighted multi-scale detail layer is backfilled into the extinction image through feature synthesis to obtain the detail-enhanced image.

[0047] Entering stage (4) Guided filtering correction based on reference structure. Using the original 90° channel image as the guide image, guided filtering is performed on the detail enhancement image. While removing the 2×2 periodic checkerboard array artifacts inherent in the focal plane sensor, the true edge skeleton of the net rope is protected from being smoothed, and a synthetic correction image is output.

[0048] Entering stage (5) Local contrast adaptive stretching. Apply CLAHE (Constrained Contrast Adaptive Histogram Equalization) to the synthesized correction image to optimize the energy distribution of the entire image, making the brightness uniform and the mesh clear, and finally outputting a reconstructed image with enhanced transparency.

[0049] The beneficial effects of this invention are as follows: (1) Extremely high transparency: Compared with the traditional physical extinction method, this invention effectively preserves the high-frequency information of the target during the descattering process by injecting multi-scale details, solving the common problem of "visual fog" residue in traditional methods, greatly improving the clarity of the output image and significantly enhancing visual transparency. (2) Precise target extraction: By using the DoLP weighting mechanism, targeted enhancement of the polymer material grid lines is achieved, which significantly improves the edge gradient of the net area and makes the contrast difference with the background area more obvious, which is beneficial to subsequent target detection and automatic recognition tasks. (3) Artifact suppression and stability: Through four-channel reconstruction and structure-guided filtering, the checkerboard array artifacts and spatial inconsistency noise generated by the focal plane polarization camera under high-intensity light source illumination are significantly reduced, ensuring the quality stability of the enhanced image. (4) Strong dynamic range adaptability: This invention can handle active lighting environments with high lumen levels well. While effectively suppressing the overexposure of the central spot, it clearly presents the interception details of the dark edge area, making the brightness distribution of the whole image more balanced and the mesh structure clearly distinguishable.

[0050] The present invention also provides an underwater interception net image enhancement system, including a processor and a memory storing a computer program, wherein the processor, when executing the computer program, implements the steps of any of the above-described underwater interception net image enhancement methods.

[0051] It is understood that the above embodiments only illustrate preferred embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can freely combine the above technical features without departing from the concept of the present invention, and can also make several modifications and improvements, all of which fall within the protection scope of the present invention. Therefore, all equivalent transformations and modifications made with respect to the scope of the claims of the present invention should fall within the scope of the claims of the present invention.

Claims

1. A method for enhancing images of underwater interception nets, characterized in that, include: Step S1: Acquire multi-angle polarization images under active polarization illumination and calculate the linear polarization degree image DoLP; Using one angle image as the main channel, the background scattering field is obtained. DoLP is used as a weighting coefficient to remove the background scattering field from the main channel image to obtain an extinction image. Step S2: Perform multi-scale decomposition on one of the angle images, extract at least two high-frequency detail layers at different scales, generate adaptive weights that are negatively correlated with the DoLP value based on the image DoLP, and inject the weighted multi-scale high-frequency detail layers into the extinction image to obtain a detail-enhanced image. Step S3: Using one of the angle images as a guide image, guide filtering is applied to the detail enhancement image to obtain a synthetic correction image, which is then subjected to local contrast adaptive stretching to obtain the final enhanced image.

2. The underwater interception net image enhancement method according to claim 1, characterized in that, The multi-angle polarization image includes polarization grayscale images at four angles: 0°, 45°, 90°, and 135°; in step S1, the calculated linear polarization degree image DoLP includes: Calculate the total intensity image: ; Calculate the linear polarization components: ; ; Calculate polarization intensity: ; Calculate the degree of linear polarization: DoLP = clip( (0.01, 0.99); in, , , , These are the pixel values ​​of the images at polarization directions of 0°, 45°, 90°, and 135°, respectively. It is a very small positive number, used to prevent division by zero.

3. The underwater interception net image enhancement method according to claim 2, characterized in that, In step S1, the step of obtaining the background scattering field by using one angle image as the main channel, and removing the background scattering field from the main channel image using DoLP as a weighting coefficient to obtain an extinction image includes: 90° channel image I 90 As the main channel, a Gaussian low-pass filter is used to obtain the background scattering field. B field : B field =G σlarge ( I 90 )； Among them, G σlarge It is a two-dimensional Gaussian low-pass filter used to extract slowly changing background halo intensity distribution; Using the linear polarization degree image DoLP as a weighting coefficient, a soft extinction method is employed to remove the background scattering field from the main channel image: Where ⊙ represents pixel-by-pixel multiplication, k is a preset extinction intensity coefficient, and DoLP is the linear polarization degree image; right I descat The image is cropped pixel by pixel to the range of [0, 255] to obtain an extinction image.

4. The underwater interception net image enhancement method according to claim 1, characterized in that, In step S2, the multi-scale decomposition of one angle image to extract at least two high-frequency detail layers at different scales includes: 90° channel image I 90 As a decomposition object: Extracting micro-detail layers D micro = I 90 -G σsmall ( I 90 ), 5×5 cores, ( ≈1.1); Extracting the meso-detail layer D macro = I 90 -G σmed ( I 90 ), 15×15, ( ≈2.6); Among them, G σsmall and G σmed These are the first Gaussian low-pass filter and the second Gaussian low-pass filter, respectively, with the size of the second Gaussian kernel being larger than the size of the first Gaussian kernel. denoted as the standard deviation of the Gaussian kernel.

5. The underwater interception net image enhancement method according to claim 4, characterized in that, In step S2, generating adaptive weights that are negatively correlated with the DoLP value based on the image DoLP includes: Generate adaptive weights based on image DoLP. W : W =1.0+(1.0-DoLP); The adaptive weights are negatively correlated with the DoLP value.

6. The underwater interception net image enhancement method according to claim 5, characterized in that, In step S2, injecting the weighted multi-scale high-frequency detail layer into the extinction image to obtain a detail-enhanced image includes: The weighted multi-scale high-frequency detail layer is injected into the extinction image. I descat : I enh = I descat +(0.5· D micro +0.8· D macro )⊙W； Where ⊙ represents pixel-by-pixel multiplication, and 0.5 and 0.8 are the injection intensity coefficients for the micro detail layer and the meso detail layer, respectively; Will I enh Output as an image with enhanced detail.

7. The underwater interception net image enhancement method according to claim 1, characterized in that, In step S3, the step of using one of the angle images as a guide image to perform guided filtering on the detail enhancement image to obtain the synthetic correction image includes: 90° channel image I 90 Normalize to the range [0, 255] to obtain the guide graph. G guide ; Enhance the details of the image I enh Normalize the image to the range [0, 255] to obtain the image to be filtered. I src ; The image to be filtered is filtered using a guided filtering algorithm, with the guided image as the filter. G guide As a guide, a synthetic correction image is obtained. I guided : I guided =GuidedFilter( G guide , I src ,r=10, η =80); Where r is the radius of the filtering window. η This is a regularization parameter used to control the smoothing intensity.

8. The underwater interception net image enhancement method according to claim 7, characterized in that, In step S3, the local contrast adaptive stretching is performed as follows: For the synthetic correction map I guided Apply a contrast-restricted adaptive histogram equalization.

9. The underwater interception net image enhancement method according to claim 8, characterized in that, The parameters for the contrast-limited adaptive histogram equalization are set as follows: contrast limit threshold ClipLimit=2.0, block size gridSize=8×8 pixels; the processed image is output as the final enhanced image.

10. An underwater interception net image enhancement system, comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the underwater interception net image enhancement method according to any one of claims 1 to 9.

Images