Underwater image recognition method based on deep learning

Abstract

The invention provides an underwater image recognition method based on deep learning, and relates to the technical field of information, and the method comprises the steps: preprocessing an original image, recognizing image texture interference caused by the density of suspended particles in an interfered region, separating a noise signal, and obtaining a first processed image; performing deep analysis on the multi-scale feature response values and the key feature points in combination with an underwater target recognition task, and judging the category and position information of the target object to obtain a preliminary recognition result; comparing the preliminary identification result with an actual marine environment condition to obtain a matching degree between the identification result and an expected target feature, and if the matching degree is lower than a preset matching threshold value, adjusting a scale weight parameter and a feature point screening threshold value to obtain corrected identification data; and analyzing changes caused by environmental condition fluctuation according to the corrected identification data, generating target identification output, and determining the accurate position and category information of the underwater target through confidence weighted fusion and coordinate precision correction processing of the identification data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to an underwater image recognition method based on deep learning. BACKGROUND

[0002] As a core supporting technology in fields such as marine engineering, underwater robot navigation and marine biology monitoring, the accuracy and stability of underwater image recognition technology directly determine the safety and efficiency of underwater operations. With the growing demand for deep sea exploration, underwater vision systems have become an irreplaceable key technology. The current mainstream underwater image recognition method mainly relies on traditional image preprocessing technology and fixed parameter deep learning model. These schemes show obvious lack of adaptability when facing complex and changeable underwater environment. The existing technology cannot dynamically adjust according to the real-time changes of water conditions, resulting in a sharp decline in recognition performance in water areas with varying turbidity. The presence of suspended particles in the underwater environment constitutes the primary technical obstacle faced by image recognition. The distribution of suspended particles of different densities will significantly change the texture feature performance of the image, so that the originally clear texture information is covered or distorted by particle noise. This degradation of texture features directly affects the accurate perception ability of deep learning model to the surface details of target objects. When the water turbidity changes, the interference of suspended particles to the light propagation path further aggravates, causing the blur degree of the edge contour of the target object to increase significantly. The loss of edge information causes traditional feature point detection algorithms to frequently misjudge, incorrectly identifying suspended particles as target feature points, or missing real target boundary information. The increase in feature point detection error rate directly leads to the inability of the convolution layer in the deep learning network to obtain reliable input features, and the existing fixed weight distribution strategy cannot cope with this dynamic change in feature quality, thereby affecting the stability and accuracy of the entire recognition system. How to establish a quantitative relationship model between the density distribution of suspended particles and the degradation degree of image texture features, and realize the adaptive dynamic adjustment mechanism of convolution layer weights based on water turbidity changes, so as to maintain the stability of the feature extraction process while ensuring the accuracy of target recognition, has become the key problem for the breakthrough of underwater image recognition technology. SUMMARY

[0003] The present application provides an underwater image recognition method based on deep learning, mainly including: The underwater original image data is acquired, the image brightness distribution and contrast difference are calculated, and the area distribution range affected by the water body turbidity change and the suspended particle density is determined; the original image data is denoised and enhanced, the image texture interference affected by the suspended particle density is separated, and a first processed image is generated; the contour features of the target object in the first processed image are extracted, the edge feature completeness is calculated in combination with the water body turbidity change and the suspended particle density, and when the edge feature completeness is lower than a preset completeness threshold, the edge region is locally enhanced to generate a second processed image; the second processed image is subjected to edge detection and contour extraction, multi-scale analysis is introduced, the feature response values of the image under different water depth levels and distance ranges and the spatial distribution positions of key feature points are determined; based on the multi-scale feature response values and the key feature points, the category and position information of the target object are judged, and a preliminary recognition result is generated; the preliminary recognition result is compared with the marine environment condition, the matching degree is calculated, when the matching degree is lower than a preset matching threshold, the scale weight parameter and the feature point screening threshold are adjusted, and the corrected recognition data is generated; based on the corrected recognition data, the water body transparency, the suspended particle concentration and the seabed illumination intensity are fused, the coordinate accuracy is corrected, and the accurate position and category information of the target object are generated.

[0004] Further, the calculation of the image brightness distribution and the contrast difference, and the determination of the area distribution range affected by the water body turbidity change and the suspended particle density, comprises: The underwater original image data is subjected to block processing, the average brightness value of each image block is calculated, the initial area affected by the turbidity is marked according to the comparison of the average brightness value with a preset threshold value; the gray scale gradient information of each image block in the initial area is extracted, the gradient components in the horizontal and vertical directions are calculated, and a gradient amplitude distribution map is generated; the contrast difference values of adjacent pixels in the initial area are calculated, and the area affected by the suspended particle scattering is determined; based on the gradient amplitude distribution map and the contrast difference values, the transmittance of each area is calculated, and the area distribution range affected by the water body turbidity change and the suspended particle density is determined.

[0005] Further, the denoising and enhancement processing of the original image data, the separation of the image texture interference affected by the suspended particle density, and the generation of the first processed image, comprises: The local variance values of each area of the original image data are calculated, and the high-noise area is identified; a gray scale co-occurrence matrix is constructed, the gray scale value distribution of adjacent pixel pairs in the high-noise area is counted, the contrast feature value is calculated, and the texture interference area caused by the suspended particles is determined; the pixel points in the texture interference area are subjected to bilateral filtering processing, the comprehensive weight of the spatial distance and the gray scale difference is calculated, and the filtered pixel value is generated; the filtered pixel value is subjected to frequency spectrum analysis, the high-frequency noise component is suppressed, and a denoised image block is generated; the denoised image blocks are combined, and the first processed image is generated.

[0006] Further, the extraction of the first processing image object contour features, calculate the edge feature integrity, generating a second processing image, comprising: The first processing image using edge detection operator extraction contour point set, calculate the gradient amplitude and direction of each point in the contour point set, statistics contour breakpoint number and total point ratio, generate initial integrity value; based on the initial integrity value and environmental parameters, calculate the edge feature integrity; when the edge feature integrity is lower than the preset integrity threshold, along the average gradient direction of the breakpoint two sides contour segment filtering enhancement, generating local enhanced edge image; based on the local enhanced edge image, connecting the breakpoint two sides contour segment, update pixel value, generate the second processing image.

[0007] Further, the second processing image edge detection and contour extraction, introduce multi-scale analysis, determine the feature response value and the spatial distribution position of the key feature points of the image under different water depth levels and distance ranges, comprising: The second processing image computing horizontal and vertical gradient values, extract continuous contour lines, calculate the curvature value of the contour line, generate edge feature data; based on the edge feature data, set multi-scale factor for Gaussian smoothing, calculate the feature response value under each scale, generate feature response data set; the feature response data set is corrected for water depth and distance, and the feature intensity value after depth correction is calculated; based on the feature intensity value, scan the local window, select the point corresponding to the maximum intensity value, determine the spatial distribution position of the key feature points.

[0008] Further, the multi-scale feature response value and the key feature points are used to judge the category and location information of the target object, and a preliminary recognition result is generated, comprising: Based on the multi-scale feature response value and the key feature points, a feature vector is constructed, and the feature vector is normalized; the similarity between the feature vector and the category template vector is calculated, and the target object category and the initial confidence are determined; based on the key feature point set of the target object category, the spatial coordinate range is calculated; according to the number of key feature points, the initial confidence is adjusted, and the preliminary recognition result containing the target category label, spatial coordinate position and recognition confidence is generated.

[0009] Further, the multi-scale feature response value and the key feature points are used to judge the category and location information of the target object, and a preliminary recognition result is generated, comprising: Based on the multi-scale feature response value, a target contour point sequence is extracted, a shape descriptor and a texture feature are calculated, and a feature vector is formed; according to the shape descriptor and the texture feature, it is judged whether the target is a fish, a seabed deposit or a man-made object, and a corresponding category code is obtained; based on the pixel coordinates and the depth measurement value of the target, a horizontal distance and an azimuth angle are calculated, and a three-dimensional position vector is generated; based on the three-dimensional position vector, a conversion to a world coordinate system is performed, and the preliminary identification result is generated.

[0010] Further, the comparison between the preliminary identification result and the marine environment condition is performed to calculate a matching degree, adjust a scale weight parameter and a feature point screening threshold, and generate the modified identification data, including: The target category, the feature vector and the spatial position information in the preliminary identification result are extracted, a difference measurement value of the feature vector and a standard feature vector is calculated, and a matching degree is generated; when the matching degree is lower than a preset matching threshold, the scale weight parameter and the feature point screening threshold are adjusted, and the feature vector is re-extracted; based on the feature vector, a classification recognition is performed, and the modified identification data containing the modified target category, the feature vector and the matching degree value is generated.

[0011] Further, based on the modified identification data, the water body transparency, the suspended particle concentration and the seabed illumination intensity are fused, the coordinate accuracy is corrected, and the accurate position and category information of the target object are generated, including: The target category and the initial coordinate information in the modified identification data are extracted, a standard deviation of the environmental parameter is calculated, and an environmental stability index is generated; based on the environmental stability index, the identification confidence is adjusted, and the multi-frame identification results are fused to generate a category determination result; based on the initial coordinate information, the water body refractive index is calculated, the line-of-sight offset is corrected, and the identification output containing the three-dimensional coordinate value and the category code is generated.

[0012] Further, based on the modified identification data, the water body transparency, the suspended particle concentration and the seabed illumination intensity are fused, the coordinate accuracy is corrected, and the accurate position and category information of the target object are generated, including: Based on the modified identification data, a water body transparency change amount is calculated, and a target edge blurring degree is evaluated; the suspended particle concentration data is analyzed, and noise intensity and type are identified; based on the seabed illumination intensity, a target feature contrast is calculated; a compensation vector is constructed, a confidence and a coordinate correction parameter are adjusted, and the accurate position and category information of the target object are generated.

[0013] The technical scheme provided by the embodiments of the present application can include the following beneficial effects: The application discloses an underwater image recognition method based on deep learning, which realizes accurate target recognition in complex marine environment through multi-stage processing and analysis of underwater original images. Firstly, the original image is preprocessed to remove the interference of water turbidity and suspended particles and extract target contour features. Then, an adaptive feature extraction framework is constructed, a multi-scale analysis method is introduced, and the distribution of key feature points is determined. Based on this, deep analysis is carried out to obtain the target category, position and recognition confidence. Through comparison with the actual environment, the parameters are dynamically adjusted to improve the recognition accuracy. Finally, the influence of environmental condition fluctuation is analyzed, a compensation mechanism is established, and stable recognition results are generated. The application can effectively overcome the interference of complex underwater environment on target recognition, improve the accuracy and reliability of underwater target detection and recognition, and provide important technical support for marine resource exploration, underwater operation and other fields. BRIEF DESCRIPTION OF DRAWINGS

[0014] Fig. 1 A flowchart of the underwater image recognition method based on deep learning of the application.

[0015] Fig. 2 A schematic diagram of the underwater image recognition method based on deep learning of the application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be described in detail below in combination with the drawings and specific embodiments.

[0017] As Figs. 1-2 , the underwater image recognition method based on deep learning of the embodiment can specifically include: Step S101, acquire underwater original image data, preliminarily analyze the collected image, extract the regional features affected by water turbidity change and suspended particle density, calculate the image brightness distribution and contrast difference, and determine the distribution range of the interference region.

[0018] The process involves acquiring raw underwater image data, segmenting the images into blocks, calculating the average brightness value of each block, and marking initial regions affected by turbidity based on a comparison of the brightness value with a preset threshold. Simultaneously, grayscale gradient information for each block is extracted, and the horizontal and vertical gradient components are calculated using the Sobel operator to obtain a gradient magnitude distribution map. For the marked initial regions, the contrast difference between adjacent pixels is calculated. The contrast difference is defined as the absolute value of the difference between the grayscale values ​​of adjacent pixels. If the contrast difference is lower than a preset contrast threshold, the region is determined to be affected by suspended particle scattering. The spatial density distribution of suspended particles is estimated based on the inverse relationship of the contrast difference. A scattering influence matrix is ​​constructed based on the particle density distribution and the gradient magnitude distribution map. Each element in the matrix is ​​obtained by multiplying the particle density by the corresponding gradient magnitude and normalizing. Based on the Beer-Lambert law, the transmittance of each region is calculated; the transmittance is equal to the natural exponent raised to the power of the negative scattering influence. Based on the transmittance distribution, a circular structuring element is used to perform morphological closure operations, first expanding and then corroding to connect adjacent low transmittance regions. Based on the boundaries of connected regions with transmittance below the threshold, the distribution range of regions affected by changes in water turbidity and suspended particle density is determined.

[0019] For example, during underwater image imaging, light is scattered and absorbed by water molecules and suspended particles as it propagates through the water, leading to a decrease in image quality. When the turbidity of the water increases, the density of suspended particles increases, and the scattering effect of light becomes more pronounced, resulting in an image that is generally brighter but has lower contrast.

[0020] In one possible implementation, when processing the raw underwater image in blocks, a fixed-size sliding window traverses the entire image. The average brightness value of each image block reflects the degree to which that area is affected by turbidity. When turbidity is high, the increased scattered light leads to a higher average brightness value for the image block. By setting a reasonable brightness threshold, areas affected by turbidity can be preliminarily identified. The Sobel operator, as a classic edge detection operator, can effectively extract gradient information from the image by calculating the gray-level change rate of pixels in the horizontal and vertical directions. The generation of the gradient magnitude distribution map provides basic data for image texture features in subsequent analysis.

[0021] It's important to note that the contrast difference between adjacent pixels directly reflects the local sharpness of an image. In clear water, the edges of objects exhibit high contrast; however, in turbid water, due to the scattering effect of suspended particles, object edges become blurred, and contrast is significantly reduced. By statistically analyzing the distribution pattern of contrast differences, an inverse relationship model between contrast difference and suspended particle density was established. The physical basis of this relationship is that the higher the particle density, the more severe the forward scattering, and the more pronounced the decrease in image contrast.

[0022] Specifically, the construction of the scattering influence matrix comprehensively considers two factors: particle density and image gradient. Particle density reflects the concentration of the scattering medium, while the gradient magnitude characterizes the image's detail information. The product of these two factors, after normalization, quantifies the degree of scattering influence at each pixel location. Beer-Lambert's law describes the attenuation of light in a medium, with transmittance exhibiting an exponential relationship with the scattering coefficient. Using this physical model, the transmittance distribution of each region can be deduced from the degree of scattering influence.

[0023] In one embodiment, the application of morphological closing operations effectively solves the noise problem in the preliminary detection results. The selection of circular structural elements takes into account the isotropic characteristics of underwater scattering. The expansion operation fills the small holes inside the low-transmittance region, and the corrosion operation restores the original size of the region; the combination of the two achieves the connection of adjacent low-transmittance regions. Through transmittance threshold determination and connected component analysis, the boundary range of the disturbed region is accurately delineated.

[0024] Step S102: Preprocess the original image, identify image texture interference caused by the density of suspended particles in the disturbed area, separate the noise signal, and obtain the first processed image.

[0025] Noise detection is performed on the original image, and the local variance value of each region is calculated. High-noise regions are identified based on the variance value. Simultaneously, a gray-level co-occurrence matrix is ​​constructed, and the gray-level value distribution of adjacent pixel pairs in different directions is statistically analyzed. The contrast feature value of the matrix is ​​calculated. When the contrast feature value exceeds a preset threshold, the region is identified as a texture interference region caused by suspended particles. For pixels in texture interference regions, a bilateral filter is used for processing. The spatial distance and gray-level difference between the current pixel and its neighbors are calculated. The spatial weight is determined using a Gaussian function based on the spatial distance, and the range weight is calculated using a Gaussian function based on the gray-level difference. The two weights are multiplied to obtain a comprehensive weight. A weighted average of the neighboring pixels is then applied according to this comprehensive weight to obtain the filtered pixel value. A two-dimensional discrete Fourier transform is performed on the local image patch composed of the filtered pixel values ​​to obtain the spectral distribution. The amplitude of components in the spectrum higher than a preset frequency threshold is analyzed. When the amplitude of a high-frequency component exceeds a noise judgment threshold, noise suppression is achieved by setting that frequency component to zero. An inverse Fourier transform is then performed on the processed spectrum to obtain a denoised image patch in the spatial domain. The signal-to-noise ratio (SNR) is calculated based on the mean square error between the denoised image block and the original corresponding region. If the SNR is lower than a preset threshold, the image block is subjected to contrast stretching. The grayscale range is expanded through linear transformation to improve image clarity. All processed image blocks are combined to form the first processed image with the texture interference of suspended particles removed.

[0026] For example, the calculation of local variance is a key indicator for noise detection.

[0027] In one possible implementation, the image is divided into 16×16 pixel windows. The sum of squared deviations of the grayscale values ​​of each pixel within the window from the mean is calculated and then divided by the total number of pixels. When underwater images are disturbed by suspended particles, the random distribution of these particles causes increased fluctuations in grayscale values ​​in local areas, resulting in a significant increase in variance. Regions with variance values ​​exceeding 30 are typically considered high-noise regions. The gray-level co-occurrence matrix reflects the statistical characteristics of image texture.

[0028] Specifically, the matrix records the frequency of occurrence of pixel pairs with specific grayscale values ​​at a specific direction and distance.

[0029] For example, if a pair of pixels with grayscale values ​​of 50 and 55 appear 120 times at a horizontal distance of one pixel, this information is recorded in the corresponding position in the matrix. The contrast feature value is obtained by calculating the weighted sum of the squared differences between each element in the matrix and its row and column differences. Texture interference caused by floating particles will cause the contrast feature value to increase abnormally; when it exceeds the threshold of 80, it can be identified as an interference area.

[0030] It's important to note that the unique advantage of bilateral filters lies in their edge-preserving properties. Spatial weights are calculated based on the physical distance between pixels; closer pixels receive greater weights, following a Gaussian distribution. Range weights consider the similarity of pixel grayscale values, with pixels exhibiting smaller grayscale differences receiving higher weights. This dual-weighting mechanism allows the filter to smooth noise while protecting image edges from blurring. For underwater images, the contour information of target objects is preserved, while random noise from suspended particles is effectively suppressed.

[0031] In one embodiment, a two-dimensional discrete Fourier transform converts the spatial domain image to the frequency domain for analysis. Noise generated by underwater suspended particles is mainly concentrated in the high-frequency range, appearing as bright spots far from the center in the spectrum. By setting a frequency threshold, components with frequencies higher than 0.3 times the sampling frequency are identified. When the amplitude of these high-frequency components exceeds twice the average amplitude, they are determined to be noise components and set to zero. This frequency domain filtering method is particularly suitable for handling periodic or quasi-periodic particle interference patterns. The signal-to-noise ratio (SNR) calculation provides a quantitative evaluation standard for image quality. The SNR value is obtained by calculating the mean square error between the denoised image patch and the original image patch, and then comparing it with the signal power. When the SNR is below 15 dB, it indicates that the image quality still needs improvement. Contrast stretching extends the original grayscale range to a wider interval through linear transformation, enhancing both dark and bright details. The resulting first processed image effectively removes texture interference caused by suspended particles while preserving the original structural information, significantly improving image clarity.

[0032] Step S103: Extract the contour features of the target object in the first processed image, combine the water turbidity change and suspended particle density to obtain the interference degree, and analyze the integrity of the edge features based on the edge continuity and gradient change. If the integrity is lower than the preset integrity threshold, the edge region is locally enhanced to obtain the second processed image.

[0033] The Canny edge detection operator is used to extract the contour points of the target object from the first processed image. The gradient magnitude and gradient direction of each contour point are calculated. When the distance between adjacent contour points exceeds a preset distance threshold or the difference in gradient direction exceeds a preset angle threshold, it is marked as a contour breakpoint. The ratio of the number of breakpoints to the total number of contour points is used as the initial integrity value. A correction coefficient is calculated based on the initial integrity value and pre-acquired environmental parameters. The correction coefficient is equal to the sum of the turbidity normalized value and the particle density normalized value divided by two. The difference between the initial integrity value and the correction coefficient is multiplied by one to obtain the edge feature integrity considering environmental interference. If the edge feature integrity is lower than the preset integrity threshold, the coordinates and corresponding gradient direction information of all breakpoint locations are extracted. For the neighborhood region of each breakpoint, Gaussian filtering is performed along the average gradient direction of the contour segments on both sides of the breakpoint to enhance the edge response in that direction, resulting in a locally enhanced edge image. Based on the locally enhanced edge image, the endpoint positions of the contour segments on both sides of the breakpoint are detected. When the distance between the endpoints is less than a preset connection threshold and the gradient direction difference is less than a preset angle threshold, a connected pixel sequence is generated by linear interpolation. The connected pixel sequence is merged with the original contour points, and the pixel values ​​at the corresponding positions in the image are updated to obtain a second processed image with improved contour continuity.

[0034] For example, the Canny edge detection operator achieves accurate edge extraction through multi-stage processing.

[0035] In one possible implementation, the operator first applies a Gaussian filter to the image to reduce noise interference, then calculates the gradient values ​​for each pixel in the horizontal and vertical directions. The gradient magnitude reflects the drastic change in grayscale at that point, while the gradient direction indicates the direction of the fastest grayscale change. Through non-maximum suppression and dual threshold detection, a contour line with a single pixel width is finally obtained. The identification of contour breakpoints is based on a dual criterion of spatial distance and directional consistency.

[0036] Specifically, when the Euclidean distance between adjacent contour points exceeds 5 pixels, it indicates a spatial break in the contour at that location. Simultaneously, even if the spatial distance is relatively short, a difference in gradient direction exceeding 45 degrees between adjacent points also indicates a sharp turn or break in the contour at that position. This dual criterion can accurately locate various types of contour discontinuities.

[0037] It should be noted that the impact of environmental parameters on edge integrity is quantified using correction coefficients. The turbidity normalized value is obtained by dividing the measured turbidity value by the reference turbidity value, and its value ranges from 0 to 1. The particle density normalized value is calculated similarly. The correction coefficient reflects the overall impact of environmental disturbances; the more turbid the water and the denser the suspended particles, the larger the correction coefficient, and correspondingly, the lower the calculated edge feature integrity value. This correction mechanism makes the integrity assessment more consistent with actual imaging conditions.

[0038] In one embodiment, local enhancement is achieved using directional Gaussian filtering. For each breakpoint location, the average gradient direction of the contour segments on both sides is first calculated; this direction represents the main orientation of the contour. Then, an elliptical Gaussian kernel is constructed, with its major axis along the average gradient direction and its minor axis perpendicular to this direction. This anisotropic filtering process enhances the edge response in the contour direction while avoiding blurring in the vertical direction. The processed edges exhibit a stronger tendency for continuity near the breakpoint. The implementation of contour connection relies on two key steps: endpoint matching and interpolation generation. During endpoint matching, not only spatial distance is considered, but the consistency of gradient directions is also verified.

[0039] For example, if two endpoints are 8 pixels apart and their gradient directions differ by only 15 degrees, the connection condition is met. During linear interpolation, the coordinates of intermediate points are calculated based on the position coordinates of the two endpoints. Simultaneously, the gray values ​​of intermediate points are assigned according to distance weights based on the gray values ​​of the endpoints. The generated connected pixel sequence naturally fills in contour breaks, making the boundary of the target object appear complete and continuous in the second processed image, significantly improving the accuracy of subsequent image analysis and target recognition.

[0040] Step S104: Perform edge detection and contour extraction on the second processed image to obtain the edge features of the second image. Based on the edge features of the second processed image, analyze the feature response values ​​of the image under different water depths and distance ranges to determine the spatial distribution of key feature points.

[0041] The Sobel operator is used for edge detection on the second processed image. Gradient values ​​in the horizontal and vertical directions are calculated, and continuous contour lines are extracted using gradient magnitude and direction information. The rate of change of the angle formed by three adjacent points along the contour line is calculated as the curvature value, resulting in edge feature data containing position coordinates, gradient intensity, and curvature value. A feature extraction processing flow is constructed based on the edge feature data. Gaussian smoothing is applied to the image using three different scale factors. The gradient response of each point in the edge feature data is recalculated at each scale. The original gradient intensity is convolved with the corresponding Gaussian kernel to obtain the feature response value at that scale, forming a feature response dataset containing three scale layers. Depth and distance-related corrections are performed based on the feature response dataset. According to underwater imaging principles, an exponential decay function is used to calculate the decay factor at different depths. The decay factor is equal to the product of the negative depth and the decay constant raised to the power of the natural exponent. The feature response value at each scale is divided by the corresponding decay factor to obtain the depth-corrected feature intensity value. Spatial distribution analysis is performed on the depth-corrected feature intensity values. Local window scanning is performed on each scale layer. When the feature intensity value of the center point of the window is greater than the intensity values ​​of all other points in the window, the coordinates and intensity value of the point are recorded. By comparing the intensity values ​​of the same spatial location on different scale layers, the point corresponding to the maximum value is selected as the key feature point, and the spatial distribution location of the key feature points is determined.

[0042] For example, the Sobel operator, as a classic edge detection operator, uses two 3×3 convolution kernels to detect grayscale changes in the horizontal and vertical directions of an image, respectively.

[0043] In one possible implementation, the weight matrix of the horizontal convolution kernel is in the form of -1 in the left column, 0 in the middle column, and 1 in the right column, while the vertical weight matrix is ​​-1 in the top row, 0 in the middle row, and 1 in the bottom row. After performing convolution on the image, each pixel obtains a horizontal gradient value and a vertical gradient value. The square root of the sum of their squares is the gradient magnitude at that point, and the arctangent operation yields the gradient direction. Curvature calculation reflects the degree of curvature of the contour.

[0044] Specifically, three consecutive points on the contour are selected, forming two line segments. The angle between these two segments is then calculated. When the contour is straight, the angle is close to 180 degrees; when the contour makes a sharp turn, the angle deviates significantly from 180 degrees. The rate of change of the angle is a measure of curvature. By calculating point by point along the contour line using a sliding window, the inflection points and corners of the contour can be accurately identified. These feature points often correspond to key structural information of the target object.

[0045] It should be noted that multi-scale analysis is achieved through Gaussian smoothing of varying degrees. The three scale factors correspond to Gaussian kernels of different sizes; smaller kernels retain more detailed features, while larger kernels extract overall contour features. At each scale, the original edge gradient intensity is convolved with the corresponding Gaussian kernel to generate the feature response value at that scale. This multi-scale processing method allows the algorithm to simultaneously capture both local details and the overall structure of the target. The unique characteristic of underwater imaging lies in the exponential decay of light.

[0046] In one embodiment, the intensity of light decreases exponentially as it travels through water. The attenuation factor is calculated based on Beer-Lambert's law, which states that the light intensity decreases to a fixed percentage of its original value for every unit increase in depth.

[0047] For example, at a depth of 5 meters, if the attenuation constant is 0.2, the attenuation factor is approximately 0.37, meaning only 37% of the light can reach that depth. By dividing the feature response value by the corresponding attenuation factor, the signal attenuation caused by depth can be compensated. Local window scanning is a crucial step in feature point localization. The window size is typically set to 5×5 or 7×7 pixels, and the intensity values ​​of the center point and surrounding points are compared at each location. Only when the intensity value of the center point is strictly greater than that of all neighboring points is it marked as a local maximum. Cross-scale comparison further filters feature points; the same spatial location may have local maxima at different scales, and the scale with the highest intensity is selected as the feature scale for that point. The identified key feature points not only have accurate spatial coordinates but also carry scale and intensity information, providing a rich and reliable feature foundation for subsequent image analysis.

[0048] Step S105: In conjunction with the underwater target recognition task, perform in-depth analysis on multi-scale feature response values ​​and key feature points to determine the category and location information of the target object and obtain preliminary recognition results.

[0049] Feature vectors are constructed based on multi-scale feature response values ​​and key feature points. The response intensity value, scale level, and gradient direction histogram of surrounding pixels for each key feature point are concatenated into a one-dimensional vector. This vector is then normalized to a magnitude of 1, resulting in a standardized feature vector. Similarity is calculated between the feature vector and pre-stored template vectors for each category. The similarity value is obtained by calculating the inner product of the feature vector and each category template vector; a similarity value closer to 1 indicates greater similarity. When the maximum similarity value exceeds a preset category determination threshold, the corresponding category is identified as the target object category, and this maximum similarity value is used as the initial confidence level for category determination. Target location information is calculated based on the set of key feature points belonging to the determined category. The maximum and minimum values ​​of the x-coordinate and y-coordinate of these feature points are calculated. The minimum x-coordinate and minimum y-coordinate form the top-left corner coordinate, and the maximum x-coordinate and maximum y-coordinate form the bottom-right corner coordinate. These two sets of coordinates together determine the spatial location range of the target. The initial confidence level is adjusted based on the number of valid feature points involved in the category determination. The ratio of the number of valid feature points to the number of standard feature points for that category is calculated. The initial confidence level is multiplied by this ratio to obtain the corrected recognition confidence level. The target category label, the spatial coordinates of the upper left and lower right corners, and the recognition confidence level are summarized to generate a preliminary recognition result containing complete recognition information.

[0050] For example, the process of constructing feature vectors embodies the organic integration of multi-dimensional information.

[0051] In one possible implementation, each key feature point carries three core types of information: the response intensity value reflects the significance of the point in multi-scale analysis, typically ranging from 0 to 255; the scale level indicates the Gaussian smoothing level at which the feature point was detected, usually represented by integers such as 1, 2, and 3; and the gradient orientation histogram describes the gradient distribution characteristics within an 8×8 pixel region around the point, usually divided into 8 directional intervals, with each interval summing the gradient magnitudes. The generation process of the gradient orientation histogram deserves detailed explanation.

[0052] Specifically, gradient information is extracted from 64 pixels surrounding a key feature point. The 360-degree directional range is evenly divided into 8 intervals, each covering 45 degrees. The sum of gradient magnitudes within each directional interval is calculated to form an 8-dimensional directional feature. This representation method is robust to image rotation; even with slight rotation of the target, the overall pattern of its gradient direction distribution remains relatively stable.

[0053] It should be noted that vector normalization ensures the comparability between different feature points. After concatenating the response intensity value, scale level, and gradient statistics in eight directions into an 11-dimensional vector, the Euclidean norm of this vector is calculated, and then each component of the vector is divided by this norm. The normalized vector has a magnitude of 1, eliminating the influence of absolute numerical magnitude and making subsequent similarity calculations more reasonable.

[0054] In one embodiment, similarity calculation uses vector dot products for fast matching. Pre-stored category template vectors are standardized representations obtained by averaging feature vectors from a large number of similar targets. When two normalized vectors are identical, their dot product is 1; when two vectors are orthogonal, the dot product is 0; and when their directions are opposite, the dot product is -1. In underwater target identification, a similarity value of 0.7 is typically set as the judgment threshold; exceeding this value indicates a successful match. The determination of the target's spatial location is based on the statistical distribution of feature points. Once the target category is determined, all feature points involved in the category determination constitute a point set. By traversing the coordinates of these points, the minimum and maximum values ​​of the horizontal coordinate (xmin and xmax) and the minimum and maximum values ​​of the vertical coordinate (ymin and ymax) are found. These four values ​​determine the smallest rectangular region containing the target, providing an accurate spatial positioning basis for subsequent target tracking and analysis. The confidence correction mechanism reflects the impact of feature completeness on recognition reliability. The standard feature point count is the number of typical feature points for this category of targets, derived from a large number of sample statistics.

[0055] For example, around 50 stable feature points can typically be extracted from a certain type of fish target. If only 30 effective feature points are obtained in actual detection, the ratio is 0.6. Multiplying the initial confidence score of 0.8 by 0.6 yields a corrected confidence score of 0.48. This correction method reasonably reduces the recognition confidence score under conditions of partial occlusion or degraded image quality, thereby improving the reliability of the system.

[0056] Based on multi-scale feature response values, the shape contour and texture features of the target object are extracted, and key feature points are mapped to marine organism category identifiers and artificial structure identifiers. Target category labels containing fish species codes, seabed sediment type codes, and artificial object codes are generated. At the same time, combined with image pixel coordinate system and underwater depth measurement data, a three-dimensional spatial coordinate positioning system is established to generate spatial location information containing horizontal position coordinates and vertical depth coordinates.

[0057] The target contour point sequence is extracted based on multi-scale feature response values. The contour is sampled at equal intervals to obtain a fixed number of sampling points. The distance sequence from each sampling point to the contour centroid is calculated. A discrete Fourier transform is performed on the distance sequence to obtain frequency domain coefficients as shape descriptors. Simultaneously, the energy value, entropy value, and correlation value of the texture are calculated using a gray-level co-occurrence matrix, forming a feature vector containing shape and texture information. Classification is performed based on the feature vector using preset classification rules. The regularity of the target contour is judged based on the proportion of low-frequency coefficients in the shape descriptor. If the proportion of low-frequency coefficients is high and the texture correlation value is greater than a preset threshold, it is identified as a fish with regular scales, and the corresponding code is obtained by querying a predefined fish species coding table. If the texture entropy value is lower than a preset threshold and the shape is irregular, it is identified as seabed sediment and the corresponding type code is obtained. If the shape descriptor shows straight line features, it is identified as a man-made object and the corresponding code is obtained. The system reads the pixel coordinates of the target in the image, calculates the normalized planar coordinates of that pixel based on the camera's intrinsic parameter matrix, and combines this with the simultaneously acquired depth measurement. Using triangulation, it calculates the target's horizontal distance and azimuth relative to the camera, combining these values ​​into a three-dimensional position vector. Based on this 3D position vector, a coordinate transformation is performed. According to the camera's installation depth and pitch angle relative to the water surface, a rotation matrix and translation vector are used to transform the position from the camera coordinate system to a world coordinate system based on the water surface. This yields spatial coordinates including eastward position, northward position, and underwater depth. The target category code is then combined with these spatial coordinates to generate a complete recognition result containing category labels and location information.

[0058] For example, equally spaced sampling of the contour point sequence ensures consistency in shape description.

[0059] In one possible implementation, for the extracted target contour, a sampling point is selected at fixed pixel intervals along the contour line, typically 64 or 128 points. This uniform sampling method ensures that similar targets of different sizes can obtain the same number of descriptive points, facilitating subsequent feature comparison. The distance sequence from the sampling point to the centroid reflects the radial variation characteristics of the contour; for example, the distance sequence of a circular target is basically constant, while the distance sequence of a star-shaped target exhibits periodic fluctuations. The Fourier transform converts the distance sequence from the spatial domain to the frequency domain for analysis.

[0060] Specifically, low-frequency coefficients reflect the overall shape characteristics of the contour, while high-frequency coefficients contain the detailed variations of the contour. For the contour of a smooth fish, its Fourier coefficients are mainly concentrated in the low-frequency part, and the first 10 coefficients are sufficient to reconstruct the basic shape of the contour. However, the contours of irregular objects such as rocks contain more high-frequency components, requiring more coefficients for accurate description. By analyzing the distribution characteristics of the frequency domain coefficients, different types of underwater targets can be effectively distinguished.

[0061] It should be noted that the gray-level co-occurrence matrix (GLCM) statistically characterizes the texture properties of an image. This matrix records the frequency of different gray-level value pairs at specific directions and distances. Energy values ​​reflect the uniformity of the texture; uniform textures have higher energy values. Entropy values ​​measure the complexity of the texture; complex textures have higher entropy values. Correlation values ​​characterize the directionality of the texture; regions with a clear texture direction have higher correlation. Fish scales exhibit a regular arrangement, and their texture correlation typically exceeds 0.8; while the texture of seabed sediment is more random, with a correlation below 0.3.

[0062] In one embodiment, the classification discriminant function considers both shape and texture features for target recognition. The proportion of low-frequency coefficients is obtained by calculating the ratio of the energy of the first 10 Fourier coefficients to the total energy. When this ratio exceeds 0.85, it indicates a relatively regular contour. Combined with a texture correlation greater than 0.7, the target can be identified as a fish with a regular scale structure. A predefined encoding table assigns a unique numerical code to each common fish species, such as 101 for grouper and 102 for yellow croaker. The camera intrinsic parameter matrix contains inherent parameters such as focal length and principal point coordinates, which are obtained in advance through the calibration process. Pixel coordinates are converted into normalized planar coordinates through the inverse operation of the intrinsic parameter matrix. These coordinates represent the target's orientation vector in the camera coordinate system. Combined with depth measurements, the target's three-dimensional position can be determined.

[0063] For example, with normalized coordinates of 0.5 and 0.3, and a depth of 10 meters, the horizontal distance is approximately 5.8 meters, and the azimuth angle is approximately 31 degrees. Coordinate transformation maps the camera's viewpoint to its real-world location. The camera's mounting parameters include its depth and pitch angle relative to the water surface. By constructing a rotation matrix and translation vector, the position in the camera coordinate system is transformed to the world coordinate system. The resulting east and north positions can be directly mapped to geographic coordinates, and combined with GPS positioning information, precise spatial positioning of underwater targets can be achieved.

[0064] Step S106: By comparing the preliminary identification results with the actual marine environmental conditions, the matching degree between the identification results and the expected target features is obtained. If the matching degree is lower than the preset matching threshold, the scale weight parameter and feature point screening threshold are adjusted to obtain the corrected identification data.

[0065] The system acquires the target category, feature vector, and spatial location information from the preliminary identification results. Simultaneously, it reads current marine environmental conditions, including water temperature, salinity, and light intensity measurements. Based on these environmental conditions, it searches for the corresponding standard feature vector in a pre-established environment-feature mapping relationship. The system calculates the sum of squares of the differences between the identified feature vector and the standard feature vector across each dimension, obtaining a difference metric. This difference metric is normalized by dividing by the number of feature vector dimensions. Subtracting the normalized value from 1 yields the matching degree. If the matching degree is lower than a preset matching threshold, the system analyzes the dimension contributing the largest difference in the feature vector to determine whether the main deviation is caused by the shape descriptor, texture parameters, or scale information. Based on the main source of deviation, the feature extraction parameters are adjusted. When the shape descriptor deviation exceeds the threshold, the weight coefficient of the corresponding scale layer is increased, with the weight increment equal to the deviation value multiplied by a preset adjustment coefficient. When the scale information deviation exceeds the threshold, the filtering threshold for feature point response intensity is decreased, with the decrease proportional to the deviation value. Using adjusted weighting coefficients and screening thresholds, the original multi-scale feature response data is re-weighted and feature points are screened to extract new feature vectors and perform classification recognition. The matching degree between the new recognition results and standard features is calculated, and the output includes the corrected target category, the new feature vector, and the recalculated matching degree value as the corrected recognition data.

[0066] For example, the establishment of environment-feature mapping relationships is based on a large amount of field observation data.

[0067] One possible approach involves long-term monitoring of marine life at different depths and in different sea areas, recording their appearance characteristics under specific environmental conditions.

[0068] For example, when the water temperature is 15 degrees Celsius and the salinity is 35‰, a certain type of fish has a darker body color and a texture contrast of approximately 0.6; however, in an environment with a water temperature of 25 degrees Celsius, the same type of fish has a lighter body color and a texture contrast that drops to 0.4. This mapping relationship forms a lookup table structure, where each set of environmental parameters corresponds to a standard feature vector. The difference between the feature vectors is measured using the Euclidean distance method.

[0069] Specifically, assume the feature vector contains 64 dimensions, covering information such as shape, texture, and color. The identified feature vector differs from the standard feature vector in each dimension. These differences are squared and summed to obtain the overall difference measure. Normalization is achieved by dividing by the number of dimensions (64), ensuring the difference measure falls within a reasonable range for subsequent matching degree calculations.

[0070] It should be noted that the analysis of deviation sources is instructive for parameter adjustment. The shape descriptor typically occupies the first 20 dimensions of the feature vector, mainly reflecting the geometric characteristics of the target contour; texture parameters occupy the middle 30 dimensions, containing statistical features extracted from the gray-level co-occurrence matrix; scale information occupies the last 14 dimensions, recording the target's response intensity at different scales. By calculating the deviation contribution rate of each part, the main sources of identification error can be accurately located.

[0071] In one embodiment, the weight coefficients are adjusted according to an adaptive principle. When a large deviation in the shape descriptor is detected, it indicates that the current scale weight setting has failed to adequately capture the shape features of the target. In this case, the weight coefficients of the corresponding scale layer are increased, so that the features extracted at that scale occupy a larger proportion in the final feature vector. The adjustment coefficients are typically set between 0.1 and 0.3 to ensure the stability of parameter updates.

[0072] For example, if the original weight is 0.3 and the deviation is 0.2, the new weight is updated to 0.3 plus 0.2 multiplied by 0.2, which is 0.34. Adjusting the feature point filtering threshold directly affects the number of detected feature points. A higher threshold filters out weaker feature points, potentially missing important details; a lower threshold retains more feature points but may introduce noise. When the scale information deviation exceeds the allowable range, the feature point density is increased by lowering the filtering threshold. The relationship between the reduction magnitude and the deviation value is determined by a linear function to ensure the rationality of the adjustment. The re-identification process after parameter adjustment embodies a feedback optimization mechanism. The original multi-scale response data is weighted and fused using the new weight coefficients to form the adjusted feature representation. Simultaneously, the lowered filtering threshold retains more edge feature points, enriching the completeness of the feature description. This iterative adjustment method can gradually improve the recognition accuracy, especially in complex and variable marine environments, significantly improving the system's adaptability and robustness.

[0073] Step S107: Analyze the changes caused by environmental condition fluctuations based on the corrected identification data, generate target identification output, and determine the precise location and category information of the underwater target by weighted fusion of confidence and coordinate accuracy correction of the identification data.

[0074] Based on the corrected identification data, target category and initial coordinate information are extracted. Simultaneously, water transparency measurements, suspended particle concentration data, and seabed illumination intensity readings collected by environmental monitoring sensors are acquired. The standard deviation of each environmental parameter within a preset time window is calculated, and the normalized sum of the three standard deviations is used as the environmental stability index. The identification confidence level is adjusted based on the environmental stability index. When the index is below a preset threshold, the original confidence level is multiplied by the index value to obtain a reduced confidence level; when the index is above the threshold, the confidence level remains unchanged. For multiple frames of identification results for the same target, a weighted average is calculated according to the proportion of each frame's confidence level to the total confidence level to obtain the fused category determination result. Refraction correction is performed on the fused coordinate data. The water refractive index is calculated based on Snell's law and the current transparency value. The line-of-sight offset is calculated using the relationship between the incident angle and the refraction angle. The initial coordinates are shifted along the line-of-sight direction by this offset to obtain the corrected coordinates. The coordinate error range is determined based on the ratio of suspended particle concentration to a preset concentration. The corrected coordinate information, fused category labels, and adjusted confidence values ​​are integrated to generate a recognition output. The precise location information includes the corrected 3D coordinate values ​​and coordinate error range, and the category information includes the target category code and the corresponding confidence value.

[0075] For example, the calculation of the environmental stability index embodies the idea of ​​multi-parameter comprehensive evaluation.

[0076] In one possible implementation, the time window is set to 5 minutes, during which environmental data is collected every 10 seconds. Water transparency may fluctuate between 2.5 meters and 3.2 meters, with a calculated standard deviation of 0.21 meters; suspended particle concentration varies between 15 mg / L and 25 mg / L, with a standard deviation of 3.2 mg / L; seabed light intensity ranges from 200 lux to 280 lux, with a standard deviation of 26 lux. Each standard deviation is normalized by dividing by its mean, and then summed to obtain the environmental stability index. The smaller the index, the more stable the environment and the higher the reliability of the identification results. A confidence adjustment mechanism ensures the matching of the identification results with environmental conditions.

[0077] Specifically, when the environmental stability index is 0.3, which is below the preset threshold of 0.5, it indicates that the environment is relatively stable. At this point, the original confidence level of 0.85 remains unchanged. However, when the index reaches 0.7, the environmental fluctuation is significant, and the confidence level is adjusted by multiplying 0.85 by 0.7, resulting in 0.595. This adjustment reflects the actual impact of environmental changes on recognition accuracy, avoiding over-reliance on recognition results under adverse conditions.

[0078] It should be noted that multi-frame fusion improves the robustness of the recognition. Underwater targets are detected in consecutive multi-frame images, with each frame generating a recognition result and a corresponding confidence level.

[0079] For example, in five consecutive frames, three frames were identified as grouper with confidence levels of 0.8, 0.75, and 0.82, respectively; and two frames were identified as sea bass with confidence levels of 0.6 and 0.65. The total confidence level for grouper was calculated to be 2.37, and for sea bass, it was 1.25. After proportional allocation, the fusion confidence level for grouper was approximately 0.65, significantly higher than that for sea bass, therefore it was ultimately identified as grouper.

[0080] In one embodiment, refraction correction is based on the application of Snell's law. When light travels from water to the camera, it refracts at the water-glass interface. The refractive index of water varies with transparency; it is approximately 1.34 at 3 meters of transparency and increases to 1.36 at 2 meters. Assuming the line connecting the actual target position to the camera forms a 30-degree angle with the vertical, the observation angle calculated according to the law of refraction is approximately 22 degrees, resulting in an 8-degree angular deviation. This deviation causes the visual position of the target to deviate from its actual position, with the offset proportional to the target depth. The coordinate error range is determined by considering the scattering effect of suspended particles. Higher particle concentrations result in more severe light scattering and lower ranging accuracy. When the particle concentration is 1.5 times the standard concentration, the coordinate error range expands accordingly to 1.5 times the original value.

[0081] For example, the positioning error is ±0.5 meters under normal circumstances, expanding to ±0.75 meters at high concentrations. This dynamic adjustment of the error range provides important reference for subsequent path planning and target tracking. The identification output integrates all correction and optimization results. The precise location of the target is no longer a single coordinate point, but a region encompassing a range of uncertainties. Category information is accompanied by a confidence score, allowing users to set acceptance thresholds according to actual needs. This complete output format ensures both comprehensive information and flexibility in decision-making, significantly enhancing the practical value of underwater target identification systems in complex marine environments.

[0082] Based on the changing trends of environmental parameters in the corrected identification data, the impact of reduced water transparency on target edge blurring is analyzed, the image noise interference pattern caused by increased suspended particle concentration is identified, the target feature contrast decrease caused by the attenuation of seabed light intensity is detected, a compensation mechanism is established to correlate environmental condition fluctuations with identification accuracy, the confidence weight allocation and coordinate correction parameters are adjusted, and a stable identification output result adapted to complex marine environments is generated.

[0083] Based on the corrected identification data, a numerical sequence of water transparency over a continuous time period is extracted. The difference in transparency between adjacent time points is calculated and accumulated to obtain the cumulative change. The contour sharpness is evaluated by calculating the gray-level gradient values ​​of target edge pixels. The degree of blurring is obtained by dividing the current sharpness value by the initial sharpness value. The relationship between the change in transparency and the degree of blurring is fitted using the least squares method to obtain the fitting slope. Based on the fitting slope and the time-series data of suspended particle concentration, the rate of environmental degradation is calculated. A two-dimensional fast Fourier transform is performed on the image to obtain a frequency domain representation. The energy proportion of high-frequency components is statistically analyzed to identify noise intensity. The noise type is determined based on the spatial distribution pattern of noise. The Pearson correlation coefficient between noise intensity and particle concentration is calculated. Simultaneously, the contrast value is obtained by subtracting the average gray-level value of the background area from the average gray-level value of the target area. A three-dimensional compensation vector is constructed based on the degree of blurring, noise intensity, and contrast values. When the degree of blurring exceeds a preset threshold, the first dimension of the compensation vector is set to the difference between the degree of blurring and the threshold; when the noise intensity exceeds the threshold, the second dimension is set to the ratio of the noise intensity to the threshold; and the third dimension is set to the ratio of the initial contrast to the current contrast. A confidence reduction coefficient is calculated based on the values ​​of each dimension of the compensation vector. The original recognition confidence is corrected using the confidence reduction coefficient. The correction value for the light path refraction angle is calculated based on the change in transparency. The error range of depth estimation is adjusted based on the contrast reduction ratio. The corrected target category, updated three-dimensional coordinates, and adjusted confidence are combined to output a stable recognition result adapted to the current fluctuations in the marine environment.

[0084] For example, the assessment of contour sharpness reflects the clarity of image edges.

[0085] In one possible implementation, sharpness is quantified by calculating the grayscale gradient values ​​of edge pixels. For sharp edges, the grayscale differences between adjacent pixels are significant, resulting in large gradient values, potentially exceeding 100; while the gradient values ​​for blurred edges may only be 20-30. Initially in clear water, the average gradient value of the target outline is 85. As the transparency decreases from 3 meters to 2 meters, the gradient value drops to 60. The degree of blurring is 60 divided by 85, approximately 0.71, indicating a 29% decrease in edge sharpness. Least squares fitting reveals the quantitative relationship between environmental parameters and image quality.

[0086] Specifically, multiple sets of data points were collected showing changes in transparency and their corresponding degrees of blurring. For example, a 0.5-meter decrease in transparency corresponds to a blurring degree of 0.15, a 1-meter decrease corresponds to 0.29, and a 1.5-meter decrease corresponds to 0.42. Using the least squares method, the slope of the fitted line was found to be approximately 0.28, meaning that for every 1-meter decrease in transparency, the blurring degree increases by 0.28. This slope value provides an important basis for subsequent compensation calculations.

[0087] It should be noted that the two-dimensional Fast Fourier Transform (FFT) transforms the image from the spatial domain to the frequency domain for analysis. In the frequency domain, low-frequency components correspond to the overall structure and slowly changing regions of the image, while high-frequency components reflect details and noise. The noise generated by suspended particles is mainly concentrated in the high-frequency range. The noise intensity index is obtained by calculating the ratio of the total energy in the high-frequency region to the total energy of the entire spectrum. When the particle concentration is 20 mg / L, the high-frequency energy may account for 15%; as the concentration increases to 40 mg / L, the proportion rises to 28%.

[0088] In one embodiment, the Pearson correlation coefficient quantifies the correlation between noise and particle concentration. Statistical analysis of a series of sample data is performed to calculate the covariance of the noise intensity sequence and the particle concentration sequence, then divided by the product of their standard deviations. A correlation coefficient of 0.85 indicates a strong positive correlation, meaning that higher particle concentration leads to more severe image noise. This quantitative relationship supports the rationality of noise compensation based on particle concentration. The construction of the three-dimensional compensation vector comprehensively considers various environmental influencing factors. The first dimension reflects the compensation requirement for edge blurring; when the blurring degree of 0.4 exceeds the threshold of 0.3, the compensation value is 0.1. The second dimension addresses noise interference; the ratio of noise intensity of 0.25 to the threshold of 0.2 is 1.25. The third dimension handles contrast reduction; the ratio of an initial contrast of 80 to 50 is 1.6. The compensation values ​​of these three dimensions collectively determine the adjustment range of the confidence level. The correction of the light path refraction angle is based on the application of Snell's law. Changes in water transparency directly affect the refractive index, thereby altering the light propagation path. When transparency decreases by 1 meter, the refractive index increases by approximately 0.02. For light with an incident angle of 30 degrees, the change in the angle of refraction is approximately 0.5 degrees. This angular deviation will result in a positional shift of approximately 8.7 centimeters over a propagation distance of 10 meters. By calculating and compensating for this shift in real time, the accuracy of target positioning is improved, enabling the system to maintain stable recognition performance even in harsh underwater environments.

[0089] The above are only some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and modifications can be made. Any improvements and modifications made based on the basic principles of the present invention should be considered to fall within the protection scope of the present invention.

Claims

1. A deep learning-based underwater image recognition method, characterized in that, The method includes: Acquire raw underwater image data, calculate image brightness distribution and contrast differences, and determine the distribution range of areas affected by water turbidity changes and suspended particle density; perform denoising and enhancement processing on the raw image data, separate image texture interference affected by suspended particle density, and generate a first processed image; extract the contour features of target objects in the first processed image, and calculate the edge feature integrity based on water turbidity changes and suspended particle density; when the edge feature integrity is lower than a preset integrity threshold, perform local enhancement on the edge region to generate a second processed image; perform edge detection and contour extraction on the second processed image, and introduce multi-scale... Multi-scale analysis is performed to determine the feature response values ​​and spatial distribution of key feature points of the image at different water depths and distances. Based on the multi-scale feature response values ​​and key feature points, the category and location information of the target object are determined, and a preliminary recognition result is generated. The preliminary recognition result is compared with marine environmental conditions to calculate the matching degree. When the matching degree is lower than a preset matching threshold, the scale weight parameter and feature point screening threshold are adjusted to generate corrected recognition data. Based on the corrected recognition data, water transparency, suspended particle concentration, and seabed illumination intensity are fused to correct the coordinate accuracy and generate the precise location and category information of the target object.

2. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The calculation of image brightness distribution and contrast differences determines the regional distribution range affected by changes in water turbidity and suspended particle density, including: The underwater raw image data is segmented into blocks, and the average brightness value of each image block is calculated. Based on the comparison of the average brightness value with a preset threshold, the initial region affected by turbidity is marked. The grayscale gradient information of each image block in the initial region is extracted, and the gradient components in the horizontal and vertical directions are calculated to generate a gradient amplitude distribution map. The contrast difference between adjacent pixels in the initial region is calculated to determine the region affected by suspended particle scattering. Based on the gradient amplitude distribution map and the contrast difference, the transmittance of each region is calculated to determine the distribution range of regions affected by changes in water turbidity and suspended particle density.

3. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The step of denoising and enhancing the original image data, separating image texture interference affected by the density of suspended particles, and generating a first processed image includes: Calculate the local variance of each region in the original image data to identify high-noise regions; construct a gray-level co-occurrence matrix, statistically analyze the gray-level distribution of adjacent pixel pairs within the high-noise regions, calculate contrast feature values, and determine the texture interference regions caused by suspended particles; perform bilateral filtering on the pixels in the texture interference regions, calculate the combined weight of spatial distance and gray-level difference, and generate filtered pixel values; perform spectral analysis on the filtered pixel values ​​to suppress high-frequency noise components and generate denoised image blocks; combine the denoised image blocks to generate the first processed image.

4. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The step of extracting the contour features of the target object in the first processed image, calculating the edge feature integrity, and generating the second processed image includes: An edge detection operator is used to extract a set of contour points from the first processed image. The gradient magnitude and direction of each point in the set of contour points are calculated, and the ratio of the number of contour breakpoints to the total number of points is calculated to generate an initial integrity value. Based on the initial integrity value and environmental parameters, the edge feature integrity is calculated. When the edge feature integrity is lower than a preset integrity threshold, filtering and enhancement are performed along the average gradient direction of the contour segments on both sides of the breakpoint to generate a locally enhanced edge image. Based on the locally enhanced edge image, the contour segments on both sides of the breakpoint are connected, the pixel values ​​are updated, and the second processed image is generated.

5. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The process of edge detection and contour extraction on the second processed image, incorporating multi-scale analysis to determine the feature response values ​​and spatial distribution of key feature points at different water depths and distances, includes: The second processed image is used to calculate the gradient values ​​in the horizontal and vertical directions, extract continuous contour lines, calculate the curvature values ​​of the contour lines, and generate edge feature data. Based on the edge feature data, Gaussian smoothing is performed by setting a multi-scale factor, and feature response values ​​at each scale are calculated to generate a feature response dataset. The feature response dataset is corrected for water depth and distance, and the feature intensity value after depth correction is calculated. Based on the feature intensity value, a local window is scanned, and the point corresponding to the maximum intensity value is selected to determine the spatial distribution location of the key feature point.

6. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The step of determining the category and location information of the target object based on the multi-scale feature response values ​​and the key feature points, and generating preliminary recognition results, includes: Based on the multi-scale feature response values ​​and the key feature points, a feature vector is constructed and normalized; the similarity between the feature vector and the category template vector is calculated to determine the target object category and initial confidence; based on the set of key feature points of the target object category, the spatial coordinate range is calculated; the initial confidence is adjusted according to the number of key feature points to generate the preliminary recognition result containing the target category label, spatial coordinate position, and recognition confidence.

7. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The step of determining the category and location information of the target object based on the multi-scale feature response values ​​and the key feature points, and generating preliminary recognition results, includes: Based on the multi-scale feature response values, the target contour point sequence is extracted, and shape descriptors and texture features are calculated to form a feature vector. Based on the shape descriptors and texture features, the target is determined to be a fish, seabed sediment, or man-made object, and the corresponding category code is obtained. Based on the target's pixel coordinates and depth measurement values, the horizontal distance and azimuth are calculated to generate a three-dimensional position vector. Based on the three-dimensional position vector, it is converted to the world coordinate system to generate the preliminary recognition result.

8. The underwater image recognition method based on deep learning according to claim 1, characterized in that, The process of comparing the preliminary identification results with marine environmental conditions, calculating the matching degree, adjusting the scale weight parameters and feature point screening thresholds, and generating corrected identification data includes: Extract the target category, feature vector, and spatial location information from the preliminary identification results, calculate the difference metric between the feature vector and the standard feature vector, and generate a matching degree; when the matching degree is lower than a preset matching threshold, adjust the scale weight parameter and feature point screening threshold, and re-extract the feature vector; based on the feature vector, perform classification and identification to generate the corrected identification data containing the corrected target category, feature vector, and matching degree value.

9. The underwater image recognition method based on deep learning according to claim 1, characterized in that, Based on the corrected identification data, the system integrates water transparency, suspended particle concentration, and seabed illumination intensity to correct coordinate accuracy and generate precise location and category information for the target object, including: Extract the target category and initial coordinate information from the corrected recognition data, calculate the standard deviation of environmental parameters, and generate an environmental stability index; based on the environmental stability index, adjust the recognition confidence, fuse multi-frame recognition results, and generate a category determination result; based on the initial coordinate information, calculate the water refractive index, correct the line-of-sight offset, and generate a recognition output containing three-dimensional coordinate values ​​and category codes.

10. The underwater image recognition method based on deep learning according to claim 1, characterized in that, Based on the corrected identification data, the system integrates water transparency, suspended particle concentration, and seabed illumination intensity to correct coordinate accuracy and generate precise location and category information for the target object, including: Based on the corrected identification data, the change in water transparency is calculated, and the degree of target edge blurring is assessed; the suspended particle concentration data is analyzed to identify noise intensity and type; the target feature contrast is calculated based on seabed illumination intensity; a compensation vector is constructed, and the confidence level and coordinate correction parameters are adjusted to generate the precise location and category information of the target object.

Images