Fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar image

Abstract

The present application relates to the field of image processing, especially to a fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images, comprising: extracting a sonar image water body background, dividing regions and calculating a signal-to-noise ratio, dynamically adjusting a noise threshold, and obtaining a saliency image through denoising. Based on size statistics, preset ellipse parameters are constructed, and a multi-rotation angle convolution kernel set is constructed. The saliency image pixel points are clustered, and the circumscribed rectangle is set as a suspected target area. The consistency of the target in the continuous frame is judged through an optical flow algorithm. The convolution kernel weight is dynamically adjusted to enhance the image, the YOLO model is trained to identify the target, the identification result is superimposed, and statistical information is provided. Through the use of dynamic noise threshold adjustment and multi-angle convolution kernel set, combined with the denoising enhancement technology based on the signal-to-noise ratio and the optical flow algorithm, the fish and shrimp targets with variable size and direction in the sonar image are accurately identified and tracked, and the limitations of traditional methods are effectively overcome, providing reliable technical support for target detection in complex underwater environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing. In particular, it relates to a fish and shrimp denoising enhancement and intelligent recognition method based on two-dimensional sonar images. BACKGROUND

[0002] Sonar images are images generated through the principle of underwater sound wave reflection, used to detect underwater objects and environment. In the field of underwater fishery resource investigation, aquaculture monitoring, etc., it has important application value, can provide visual information of underwater environment, help researchers and fishery workers to evaluate fish population, distribution and behavior pattern, so as to make more intelligent management and protection decisions.

[0003] However, the quality of sonar images is often severely affected by environmental noise, which limits its effectiveness in practical applications. The existence of environmental noise significantly reduces the clarity and usability of sonar images. These noises not only obscure the details in the image, but also can cause important target features to be lost or become blurred. This makes it more difficult to accurately identify and analyze underwater targets from the image, which further affects the accuracy of monitoring and evaluation.

[0004] Traditional denoising methods, such as spatial filtering and transform domain filtering, while able to reduce noise to some extent, often have difficulty in preserving both high-frequency details and low-frequency contour information, and are mostly aimed at single noise distribution, lacking adaptability to different noise region characteristics. And the fish and shrimp targets in sonar images are of various sizes and directions, and often appear in groups, making direct identification difficult, further increasing the complexity and difficulty of identification. SUMMARY

[0005] To solve the problem that traditional denoising methods are difficult to balance noise reduction and detail preservation, and lack of adaptability, leading to more complex identification of fish and shrimp targets of varying size and direction in sonar images, the present application provides solutions in the following aspects.

[0006] The fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images comprises the following steps: extracting a water body background area without targets in a sonar image and dividing the area; obtaining a target-containing image to be processed at the same position, dividing the image according to the division mode of the water body background area, calculating a signal-to-noise ratio value, dynamically adjusting the noise threshold value of each area according to the signal-to-noise ratio value, extracting a main noise frequency set as a mask, denoising the image to be processed, and obtaining a saliency image; presetting an ellipse parameter according to the target size statistical result in the saliency image, constructing a convolution kernel set with multiple rotation angles based on the signal-to-noise ratio value; clustering the pixel point coordinates and pixel values of the saliency image to obtain a plurality of clustering clusters and set an enclosing rectangle as a suspected target area, judging whether the suspected target areas of adjacent two frames are the same suspected target area based on the intersection of the suspected target areas in a plurality of continuous frames of saliency images; inputting the continuous multiple frames of saliency images of the same suspected target area into an optical flow algorithm to output an optical flow vector, taking the average angle value of all optical flow vectors of the suspected target area as a dominant motion direction, dynamically adjusting the convolution kernel weight of each rotation angle by calculating the consistency of each rotation angle and the dominant motion direction, and performing image enhancement on the saliency image by using the adjusted convolution kernel set; manually labeling the enhanced saliency image to form a data set and train a YOLO model, identifying the target by using the trained YOLO model, superimposing the identification result on the saliency image, and providing statistical information for real-time observation by a user.

[0007] Preferably, the step of dividing the area comprises:

[0008] The water body background area data is subjected to median filtering processing to remove abnormal values, the processed water body background area is subjected to Fourier transform to calculate the amplitude spectrum noise intensity value of each pixel point, a feature matrix is constructed based on the spatial coordinates and the amplitude spectrum noise intensity value of each pixel point, the feature matrix is clustered, the Euclidean distance between each pixel point is calculated, and different clustering clusters are obtained, wherein one clustering cluster represents an area.

[0009] Preferably, the signal-to-noise ratio value comprises:

[0010] The ratio between the mean value of the amplitude spectrum of each pixel point in each region of the image to be processed and the mean value of the amplitude spectrum of each pixel point in each region of the water body background area is calculated to obtain the signal-to-noise ratio value of the corresponding region.

[0011] Preferably, the step of dynamically adjusting the noise threshold value of each area according to the signal-to-noise ratio value comprises:

[0012] The signal-to-noise ratio value is subjected to normalization processing, the result of 2 minus the normalized result is calculated, and the ratio between the signal-to-noise ratio value and the calculation result is taken as the noise threshold value.

[0013] Preferably, the region corresponding to the minimum signal-to-noise ratio value is the region with the strongest noise signal, and the region corresponding to the maximum signal-to-noise ratio value is the region with the strongest target signal.

[0014] Preferably, the step of constructing the multi-rotation angle convolution kernel set based on the signal-to-noise ratio value comprises:

[0015] The dynamic range of the preset angle interval is normalized by the signal-to-noise ratio value, the difference between the maximum angle interval and the minimum angle interval in the dynamic range is calculated to obtain the maximum adjustable range, the product of the normalized signal-to-noise ratio value and the maximum adjustable range is taken as the dynamic adjustment amount, the maximum angle interval is subtracted by the dynamic adjustment amount to obtain the angle interval, the angle value is gradually increased according to the angle interval until 360 degrees, and the corresponding rotation angle sequence is generated for each region, and the corresponding multi-rotation angle convolution kernel set is generated based on the rotation angle sequence of each region.

[0016] Preferably, the step of judging whether the suspected target regions of the adjacent two frames are the same suspected target region comprises:

[0017] Taking the current frame as a target frame, selecting any suspected target region in the target frame as an analysis region, taking all suspected target regions in the previous frame of the target frame as a matching set, calculating the intersection area of the analysis region in the target frame and the matching set of the previous frame, and selecting the maximum intersection area in all suspected target regions.

[0018] In response to the maximum intersection area being greater than a preset intersection threshold, it is considered that the analysis region of the target frame and the suspected target region corresponding to the maximum intersection area in the previous frame are the same suspected target region, otherwise, it is considered that they are not the same suspected target region.

[0019] Preferably, the step of dynamically adjusting the convolution kernel weight of each rotation angle comprises:

[0020] The absolute angle difference and the complementary angle difference between each rotation angle and the dominant motion direction are calculated, and the minimum angle in the absolute angle difference and the complementary angle difference is selected, the minimum angle is normalized, and the normalized result is exponentially attenuated by using a negative exponential function to obtain the convolution kernel weight of each rotation angle.

[0021] The present application has the following effects:

[0022] 1. The present application can more accurately identify the fish and shrimp targets with variable size and direction in the sonar image by dynamically adjusting the noise threshold and using the multi-rotation angle convolution kernel set, accurately matches the anisotropy characteristics of the target, effectively overcomes the limitations of the traditional denoising technology in detail preservation and noise suppression, and thus realizes more reliable target detection in complex underwater environment.

[0023] 2、The fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images can effectively extract a main noise frequency set and use it as a mask, so that a clearer saliency image is obtained, which helps to retain high-frequency details and low-frequency contour information of the image, and solves the shortcomings of traditional denoising methods.

[0024] 3、The method can more effectively process suspected target regions in continuous multiple frames of images by using the optical flow algorithm and dynamically adjusting the convolution kernel weight, accurately tracks the same suspected target region, and provides strong technical support for fishery resource investigation and aquaculture monitoring applications. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a method flowchart of steps S1-S5 in the fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images according to the embodiment of the present application.

[0026] Figure 2 is a fish and shrimp identification effect diagram in the fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images according to the embodiment of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments.

[0028] Referring to Figure 1 The fish and shrimp denoising enhancement and intelligent identification method based on two-dimensional sonar images includes steps S1-S5, and specifically as follows:

[0029] S1: Extract the water body background region without target in the sonar image and divide the region; obtain the image containing target at the same position to be processed, divide according to the division mode of the water body background region, calculate the signal-to-noise ratio, dynamically adjust the noise threshold value of each region according to the signal-to-noise ratio, extract the main noise frequency set as a mask, denoise the image to be processed, and obtain a saliency image.

[0030] Specifically, the step of dividing the region includes:

[0031] The water body background region data is subjected to median filtering processing to remove outliers, the processed water body background region is subjected to Fourier transform, the amplitude spectrum noise intensity value of each pixel point is calculated, a feature matrix is constructed based on the spatial coordinates and amplitude spectrum noise intensity value of each pixel point, the feature matrix is clustered, the Euclidean distance between each pixel point is calculated, and different clusters are obtained, wherein one cluster represents a region.

[0032] In this embodiment, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm is applied to partition the pure noise image. By calculating the distance and density relationship between the pixel points, the pixel point set with similar noise characteristics is identified, so as to divide the image into multiple continuous regions.

[0033] Each region represents a subspace with similar noise characteristics, such as a high noise intensity region, a low noise intensity region, and a transition region. The division of these regions is based on the comprehensive consideration of the amplitude spectrum noise intensity value and the spatial position of the pixel points.

[0034] In addition, the regions can also be divided in proportion based on the depth-of-field characteristics of the sonar image noise, including: near-field region, middle-field region, far-field region; for example, the near-field accounts for 30% of the total distance, the middle-field accounts for 40%, and the far-field accounts for 30%. Specifically, it can be adjusted according to the specific situation.

[0035] In sonar image processing, the noise energy distribution of different regions has obvious regional characteristics. However, the significance of these characteristics may be affected by the target signal intensity in the actual scene. The traditional fixed threshold method cannot adapt to this change, resulting in the possible loss of important details in the high signal-to-noise ratio region, and the possible ineffective filtering of noise in the low signal-to-noise ratio region. Therefore, by estimating the signal-to-noise ratio of each region, the energy threshold is dynamically adjusted, thereby improving the adaptability and accuracy of noise frequency extraction. The specific steps are as follows:

[0036] The calculation method of the signal-to-noise ratio includes:

[0037] The ratio between the mean value of the amplitude spectrum of each pixel point in each region in the image to be processed and the mean value of the amplitude spectrum of each pixel point in each region in the water background region is calculated to obtain the signal-to-noise ratio of the corresponding region. Among them, the region corresponding to the minimum signal-to-noise ratio is the region with the strongest noise signal, and the region corresponding to the maximum signal-to-noise ratio is the region with the strongest target signal.

[0038] The signal-to-noise ratio is normalized, and the result of 2 minus the normalized result is calculated. The ratio between the signal-to-noise ratio and the calculation result is taken as the noise threshold.

[0039] Specifically, the noise threshold of each region satisfies the following relationship:

[0040] ;

[0041] In the formula, represents the noise threshold of region , and represents region The mean of the amplitude spectrum, Indicates the area The signal-to-noise ratio value, This represents the normalization function.

[0042] It should be noted that, The purpose is to expand the range of normalized values ​​from Change to Between these points, the effect of the amplification adjustment factor is amplified, and the noise threshold is close to the mean of the regional amplitude spectrum. This means that in high signal-to-noise ratio regions, we use a higher noise threshold to retain more signal details. Because the target signal is strong enough to mask noise, a higher threshold will not mistakenly filter out the target signal; the noise threshold is close to... This means that in low signal-to-noise ratio (SNR) regions, a lower noise threshold is used to filter out noise more rigorously. Because the target signal is weak, the impact of noise is significant, so a lower threshold is needed to ensure that noise is effectively filtered out.

[0043] The set of frequency points whose noise amplitude spectrum is greater than or equal to the noise threshold is extracted as the main noise frequency set, and a mask function is defined to mark the noise frequency points. That is, the frequency points of the main noise frequency set are marked as 0, and other frequency points are marked as 1. The mask function is applied to the image to be processed containing the target to obtain the denoised salient image.

[0044] In sonar images, targets such as fish and shrimp typically appear as elliptical bright spots, the size and orientation of which vary depending on the target's species, size, and swimming posture. To accurately enhance these targets, convolutional kernels capable of matching elliptical features of different sizes and orientations are needed. However, fixed-size circular or square convolutional kernels, due to their shape limitations, cannot efficiently match the anisotropic shapes of these targets. Therefore, it is necessary to generate a series of convolutional kernel "blanks" of varying scales that cover the common size range of targets and are elliptical in shape, serving as the base template for subsequent rotation operations to match the target's orientation.

[0045] The principal axis direction of an elliptical operator reflects the main direction of extension of the target. For example, a fish in an image might appear as an elliptical bright spot, and its principal axis direction might be consistent with the direction the fish is swimming. Therefore, a horizontally aligned elliptical convolution kernel can better match these targets. For instance, if fish and shrimp mainly swim horizontally in an image, using a horizontally aligned elliptical convolution kernel can more effectively enhance the features of these targets.

[0046] Typically, the principal axis of the ellipse is set to the horizontal direction to simplify calculations. This allows for the generation of elliptical convolution kernels with a fixed rotation angle and a horizontal principal axis, serving as a base template for generating elliptical convolution kernels in different directions. Through rotation operations, elliptical convolution kernels at different angles can be generated to match the features of the target in different directions. The specific operation steps are as follows:

[0047] S2: Based on the statistical results of target size in the salient image, preset ellipse parameters and construct a set of convolution kernels with multiple rotation angles based on the signal-to-noise ratio value.

[0048] Based on statistical analysis of fish and shrimp targets in a large amount of sonar image data, the minimum, maximum, and common pixel pair sizes of the target fish and shrimp in the images were determined. The major axis of the ellipse was then defined. The range of values ​​is Pixels. For each major axis element Match a set of minor axis pixels that conform to the aspect ratio of common fish and shrimp. This generates a series of scale parameter pairs. .

[0049] Specifically, the values ​​of the convolution kernel satisfy the following relationship:

[0050] ;

[0051] In the formula, Indicates position The length of the major axis of the elliptical convolution kernel at that location and minor axis length The value, This represents an exponential function with base to natural numbers. , Indicates the coordinate position on the convolution kernel grid. The cosine value of the rotation angle of the ellipse. The sine value represents the rotation angle of the ellipse. This represents the length of the major axis of the elliptical convolution kernel. This represents the length of the minor axis of the elliptical convolution kernel.

[0052] Formula Core The standard ellipse equation is obtained through parameters. and Controlling the scale and aspect ratio of the ellipse to match targets of different sizes; through coordinate transformation. and Rotate the major axis of the ellipse until it is aligned with the coordinate axes. For example, in this step... , This means that the generated core is a "blank" at the main axis level.

[0053] The step of adjusting the convolution kernel set to enhance the saliency image is a technique known to those skilled in the art, and will not be described in detail.

[0054] The swimming direction of fish and shrimp in the sonar image is random, and the orientation of the elliptical bright spot is also different. By rotating the elliptical kernel, it is ensured that any direction target can be well matched with a certain orientation kernel, thereby obtaining effective enhancement. The area with high signal strength uses smaller angle interval to generate more dense rotation kernel sequence to finely match the anisotropic shape of fish and shrimp target; the area with low signal strength uses larger angle interval to reduce the calculation overhead while maintaining the enhancement effect. The specific steps are as follows:

[0055] The dynamic range of the preset angle interval is normalized by the signal-to-noise ratio value, the difference between the maximum angle interval and the minimum angle interval in the dynamic range is calculated to obtain the maximum adjustable range, the product of the normalized signal-to-noise ratio value and the maximum adjustable range is taken as the dynamic adjustment amount, the maximum angle interval is subtracted by the dynamic adjustment amount to obtain the angle interval, and the angle value is gradually increased according to the angle interval until 360 degrees, and the corresponding rotation angle sequence is generated for each region. Based on the rotation angle sequence of each region, a convolution kernel set with multiple rotation angles is generated.

[0056] For example, the dynamic range of the angle interval is 5 degrees to 20 degrees, which can be adjusted according to specific conditions. The angle interval of the region with the smallest signal-to-noise ratio value is set to 20 degrees, and the angle interval of the region with the largest signal-to-noise ratio value is set to 5 degrees. For example, 18 kernels are generated for the region with an angle interval of 20 degrees, and 36 kernels are generated for the region with an angle interval of 10 degrees.

[0057] Further, small fish and shrimp often appear in groups and move in the same direction. By analyzing the motion direction of the group between consecutive frames, the current frame convolution enhancement orientation is obtained, so that the enhancement operation is more targeted. The kernel with the same direction as the motion direction is given a higher weight, and the group target is significantly enhanced; the irrelevant direction response is suppressed, thereby improving the signal-to-noise ratio. The specific steps are as follows:

[0058] S3: Cluster the saliency image pixel point coordinates and pixel values to obtain a plurality of clustering clusters and set the circumscribed rectangle as a suspected target area, and judge whether the suspected target areas of adjacent two frames are the same suspected target area based on the intersection of the suspected target areas in consecutive multiple frames.

[0059] For example, the DBSCAN algorithm is used to cluster the denoised saliency image, and the K-means algorithm can also be used, which can be adjusted according to specific conditions.

[0060] Taking the current frame as a target frame, selecting any suspected target region in the target frame as an analysis region, taking all suspected target regions in the previous frame of the target frame as a matching set, calculating the intersection area of the analysis region in the target frame and the matching set of the previous frame, and selecting the maximum intersection area in all suspected target regions;

[0061] In response to the maximum intersection area being greater than a preset intersection threshold, it is considered that the analysis region of the target frame and the suspected target region corresponding to the maximum intersection area in the previous frame are the same suspected target region, otherwise, less than or equal to the preset intersection threshold, it is considered that they are not the same suspected target region.

[0062] S4: inputting the continuous multiple frames of the same suspected target region into an optical flow algorithm, outputting an optical flow vector, taking the average angle value of all optical flow vectors of the suspected target region as a dominant motion direction, calculating the consistency of each rotation angle and the dominant motion direction, dynamically adjusting the convolution kernel weight of each rotation angle, and performing image enhancement on the saliency map by using the adjusted convolution kernel set.

[0063] It should be noted that the optical flow algorithm is a known technology in the art and will not be described in detail.

[0064] The absolute angle difference and the complementary angle difference between each rotation angle and the dominant motion direction are calculated, and the minimum angle in the absolute angle difference and the complementary angle difference is selected, the minimum angle is normalized, the normalized result is exponentially attenuated by using a negative exponential function, and the convolution kernel weight of each rotation angle is obtained.

[0065] Specifically, the convolution kernel weight satisfies the following relationship:

[0066] ;

[0067] In the formula, represents the weight of the first rotation angle convolution kernel of the i-th suspected target region, represents an exponential function with a natural number n as the base, represents a normalization function, represents a minimum value function, represents the i-th rotation angle, represents the dominant motion direction of the i-th suspected target region.

[0068] Further explanation, is to deal with the periodicity of the angle. The angle is periodic, for example, the actual difference between 350 degrees and 10 degrees should be 20 degrees, not 340 degrees. By calculating the minimum value of the angle difference and its complementary angle difference, the formula correctly reflects the actual distance between two angles.​​​​​

[0069] Obtain the exclusive weight set of each suspected target region For each scale pre-generated , the following operations are performed: first use the exclusive angle weight set of the suspected target region , weight convolution is performed on all rotation kernels at this scale, and then pixel-level averaging is performed on the enhancement results at all scales to obtain the final enhancement result of the suspected target region.

[0070] The exponential function is used to map the angle difference, ensuring that the rotation angle convolution kernel consistent with the dominant motion direction obtains the highest weight, and the greater the direction deviation, the faster the weight decay. This weight distribution emphasizes features consistent with the dominant motion direction, which helps to more accurately identify and locate targets in image processing or target detection.

[0071] S5: Manually annotate the enhanced saliency map to form a dataset and train the YOLO model. Identify the target using the trained YOLO model and superimpose the identification result on the saliency map to provide statistical information for users to observe in real time.

[0072] Use the open source tool LabelImg (image annotation tool) to manually annotate the fish and shrimp targets in the enhanced image to ensure the accuracy and consistency of the annotations. Build a large-scale dataset covering fish and shrimp images under different scenarios and conditions to improve the model's generalization ability.

[0073] Randomly divide the dataset into training set and test set according to a certain proportion (for example, 6:4) to ensure that the distribution of the training set and test set is representative.

[0074] In this embodiment, the YOLOv8 (You Only Look Once version 8) framework is selected as the target detection model, and the network parameters are configured to adapt to the characteristics of sonar images. During training, monitor key indicators such as loss function (Loss) and mean average precision (mAP), and optimize model performance by adjusting hyperparameters. Through multiple iterations and verification, the mAP of the model on the test set finally reaches more than 95%, ensuring that the model has high precision and high recall rate. The trained YOLOv8 model is coupled and integrated into the data processing backend of the sonar system to ensure seamless integration of the model with the sonar system.

[0075] The sonar device continuously collects raw echo data, processes the echo data, converts it into a two-dimensional sonar image, processes the current frame image, removes noise interference, improves image quality, processes the denoised image, enhances image features, improves target detection accuracy, inputs the enhanced image into the loaded YOLOv8 model for inference, and quickly and accurately detects targets in the image. The model outputs all detected target bounding box coordinates, class labels and confidence probabilities in the image, providing users with detailed detection results.

[0076] Reference Figure 2 , the sonar imaging in the figure is usually a grayscale or pseudo-color image in an underwater environment, and there are a lot of noise and clutter. The fish and shrimp targets in the figure are labeled by a rectangular frame. Class label: such as "fish" or "shrimp". Confidence score: such as "0.99", indicating the confidence of the YOLOv8 model for the identification result.

[0077] It should be pointed out that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application patent should be subject to the appended claims.

Claims

1. A method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images, characterized in that, include: Extract and segment targetless water background regions from sonar images; The image to be processed containing the target at the same location is acquired, divided according to the water background region division method, and the signal-to-noise ratio value is calculated. The noise threshold of each region is dynamically adjusted according to the signal-to-noise ratio value. The main noise frequency set is extracted as a mask to denoise the image to be processed and obtain a significant image. Based on the statistical results of target size in salient images, ellipse parameters are preset, and a set of convolution kernels with multiple rotation angles is constructed based on the signal-to-noise ratio value; The dynamic range of the preset angle interval is used to normalize the signal-to-noise ratio (SNR) value. The difference between the maximum and minimum angle intervals in the dynamic range is calculated to obtain the maximum adjustable range. The product of the normalized SNR value and the maximum adjustable range is used as the dynamic adjustment amount. The maximum angle interval is subtracted from the dynamic adjustment amount to obtain the angle interval. The angle value is gradually increased according to the angle interval until 360 degrees. A corresponding rotation angle sequence is generated for each region. Based on the rotation angle sequence of each region, a set of convolutional kernels with corresponding multi-rotation angles is generated. Clustering of salient image pixel coordinates and pixel values ​​yields several clusters, and a bounding rectangle is set as the suspected target region. Based on the intersection of suspected target regions in multiple consecutive salient images, it is determined whether the suspected target regions of two adjacent frames are the same suspected target region. The optical flow algorithm takes multiple consecutive frames of salient images of the same suspected target region as input, outputs optical flow vectors, takes the average angle value of all optical flow vectors of the suspected target region as the dominant motion direction, calculates the consistency of the dominant motion direction of each rotation angle, dynamically adjusts the convolution kernel weights of each rotation angle, and uses the adjusted convolution kernel set to enhance the salient image. The enhanced salient images are manually annotated to form a dataset and a YOLO model is trained. The trained YOLO model identifies the target, and the identification results are superimposed on the salient images to provide statistical information for users to observe in real time.

2. The method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images according to claim 1, characterized in that, The steps for dividing the region include: Median filtering is performed on the water background area data to remove outliers. Fourier transform is then performed on the processed water background area to calculate the amplitude spectrum noise intensity value of each pixel. A feature matrix is ​​constructed based on the spatial coordinates and amplitude spectrum noise intensity value of each pixel. The feature matrix is ​​then clustered, and the Euclidean distance between each pixel is calculated to obtain different clusters, with each cluster representing a region.

3. The method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images according to claim 1, characterized in that, The signal-to-noise ratio value is obtained through the following methods: The signal-to-noise ratio (SNR) of the corresponding region is obtained by calculating the ratio between the mean amplitude spectrum of each pixel in each region of the image to be processed and the mean amplitude spectrum of each pixel in each region of the water background region.

4. The method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images according to claim 1, characterized in that, The step of dynamically adjusting the noise threshold of each region based on the signal-to-noise ratio includes the following steps: The signal-to-noise ratio (SNR) value is normalized, and the result after normalization is calculated by subtracting 2 from the normalized value. The ratio between the SNR value and the calculated result is used as the noise threshold.

5. The method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images according to claim 1, characterized in that, The region with the lowest signal-to-noise ratio (SNR) is the region with the strongest noise signal, and the region with the highest SNR is the region with the strongest target signal.

6. The method for fish and shrimp denoising enhancement and intelligent recognition based on two-dimensional sonar images according to claim 1, characterized in that, The step of determining whether the suspected target regions of two adjacent frames are the same suspected target region includes: Taking the current frame as the target frame, select any suspected target region in the target frame as the region to be analyzed, take all suspected target regions in the previous frame of the target frame as the matching set, calculate the intersection area between the region to be analyzed in the target frame and the matching set of the previous frame, and select the largest intersection area among all suspected target regions. If the maximum intersection area is greater than the preset intersection threshold, the region to be analyzed in the target frame is considered to be the same suspected target region as the suspected target region corresponding to the maximum intersection area in the previous frame. Conversely, if the intersection area is less than or equal to the preset intersection threshold, they are considered not to be the same suspected target region.

7. The method for denoising, enhancing, and intelligently recognizing fish and shrimp based on two-dimensional sonar images according to claim 1, characterized in that, The step of dynamically adjusting the weights of the convolution kernels at each rotation angle includes: Calculate the absolute angle difference and supplementary angle difference between each rotation angle and the dominant motion direction, select the angle with the smallest absolute angle difference and supplementary angle difference, normalize the smallest angle, and use a negative exponential function to exponentially decay the normalization result to obtain the convolution kernel weights for each rotation angle.

Images