Intelligent identification and screening system for male and female blue crabs and multi-view image recognition method

Abstract

The present invention relates to the field of image recognition technology, and in particular to an intelligent sex identification and screening system for blue crabs and a multi-view image recognition method. The system comprises a multi-view image acquisition module, an image preprocessing unit, a feature fusion and identification module, an intelligent decision-making control module, and a dynamic screening actuator. The multi-view image acquisition module includes at least three high-definition cameras arranged at different angles to capture multi-angle images of the carapace, plastron, and lateral carapace of blue crabs. The image preprocessing unit performs grayscale normalization, noise removal, and region of interest extraction on the captured images. The feature fusion and identification module includes a multimodal feature extraction network based on deep learning and a sex identification classifier. The present invention can comprehensively acquire multi-view image information of blue crabs and perform automated identification and screening by efficiently fusion-analyzing multimodal features.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to an intelligent identification and screening system for male and female blue crabs and a multi-view image recognition method. Background Art

[0002] In the mud crab aquaculture industry, sexing during the clashing stage is a critical step in achieving precision farming and improving economic efficiency. Previous studies have found that a sex ratio close to 2:1 during clashing crab farming yields the best growth performance and survival rate (Luo et al., Crustaceana. 2025, 98:271-281). Currently, sexing mud crabs relies primarily on manual inspection, using visual observation of carapace morphology and plastron characteristics. However, this manual identification method has significant drawbacks. Firstly, individual characteristics of mud crabs vary slightly, requiring a high level of operator experience, resulting in low identification efficiency and difficulty meeting the high-throughput screening requirements of large-scale aquaculture. Secondly, manual operation is subject to significant subjective judgment, making it prone to misjudgment and resulting in unstable sexing accuracy. With the application of machine vision technology in aquaculture, some existing technologies have attempted to achieve sexing of mud crabs through single-view image acquisition and simple image analysis algorithms. However, this type of technology generally suffers from incomplete visual coverage, capturing only single-view image information of blue crabs, failing to fully capture the comprehensive morphological differences between male and female individuals across multiple areas, such as the carapace, plastron, and lateral carapaces. Because sexing blue crabs requires a comprehensive assessment of texture, geometry, and color features across multiple areas, single-view image acquisition results in the loss of key identifying features, hindering the accuracy of existing automated identification technologies and making them unable to meet the demand for high-precision sexing in actual production.

[0003] Based on the above problems, there is an urgent need for an automated identification and screening technology that can comprehensively obtain multi-perspective image information of blue crabs and perform efficient fusion analysis of multimodal features, so as to solve the problems of low identification accuracy and poor efficiency caused by a single perspective and incomplete feature extraction in the existing technology, and realize the intelligent and precise identification of male and female blue crabs. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings existing in the prior art, and a blue crab male and female intelligent identification and screening system is proposed, which includes a multi-view image acquisition module, an image preprocessing unit, a feature fusion identification module, an intelligent decision control module and a dynamic screening execution mechanism; the multi-view image acquisition module includes at least three high-definition cameras arranged at different angles, which are used to obtain multi-angle images of the carapace, plastron and lateral carapace of the blue crab; the image preprocessing unit performs grayscale normalization, noise elimination and region of interest extraction on the collected images; the feature fusion identification module includes a multimodal feature extraction network based on deep learning and a male and female identification classifier, which can fuse the texture, geometry and color features of multi-view images and output identification results; the intelligent decision control module generates a corresponding screening control signal according to the identification result; the dynamic screening execution mechanism performs male and female classification and screening of blue crabs according to the screening control signal.

[0005] Preferably, the multi-view image acquisition module has four high-definition cameras, which are respectively arranged directly above, directly below, to the left and to the right of the blue crab support platform, and the angle between the optical axis of each camera and the horizontal plane of the support platform is 30°-60°, which can cover the complete outer surface morphology of the blue crab.

[0006] Further preferably, the noise elimination process of the image preprocessing unit adopts an adaptive median filtering algorithm, and dynamically adjusts the filter window size according to the noise density of the local area of ​​the image. The window size calculation formula is:

[0007] ;

[0008] in, is the side length of the filtering window, is the coordinate in the image The absolute difference between the pixel gray value at and the neighborhood mean, is the number of pixels in the current neighborhood, is the preset noise discrimination threshold.

[0009] Further preferably, the region of interest extraction processing includes carapace contour positioning and plastron genital pore area segmentation. The carapace contour positioning adopts an edge detection algorithm based on the active shape model. By establishing a priori shape model of the blue crab carapace, the carapace edge in the image is iteratively fitted; the plastron genital pore area segmentation adopts a threshold segmentation algorithm based on color space conversion. The image is converted from RGB color space to HSV color space, the V channel component is extracted, and the Otsu algorithm is used to determine the optimal segmentation threshold.

[0010] Further preferably, the intelligent decision control module is integrated with a real-time data monitoring unit and a model update interface. The real-time data monitoring unit can monitor the accuracy and efficiency of identification and screening in real time and generate a status report; the model update interface supports online update training of the male and female identification classifier through external data input to adapt to the identification needs of blue crabs and crab varieties at different growth stages and varieties.

[0011] A multi-view image recognition method, applied to any of the above-mentioned intelligent identification and screening systems for male and female blue crabs, comprises the following steps:

[0012] Multi-view image acquisition: High-definition cameras at at least three different angles are used to capture multi-angle images of the carapace, plastron, and lateral carapace of the blue crab.

[0013] Image preprocessing: grayscale normalization, noise elimination and region of interest extraction of the collected images;

[0014] Multimodal feature extraction: A deep learning network is used to extract the texture, geometric, and color features of each view image. The texture feature uses an improved local binary pattern algorithm to calculate the 8-neighborhood local binary pattern value of each pixel. The formula is:

[0015] ;

[0016] in, is the center pixel coordinate, is the gray value of the center pixel, is the gray value of the neighborhood pixel, is a symbolic function, when hour ,otherwise ;

[0017] Feature fusion and identification: The multimodal features from each perspective are spatially aligned and dimensionally spliced, and then input into the male-female identification classifier for training and classification. The classifier uses a multi-layer perceptron model combined with an attention mechanism, and adjusts the contribution of different perspective features to the identification results through attention weights.

[0018] Screening decision and execution: Generate corresponding screening control signals based on the identification and classification results to drive the dynamic screening actuator to perform sex-classification screening on blue crabs.

[0019] Further preferably, in the multimodal feature extraction, the geometric features include the carapace length-to-width ratio, the position coordinates and shape parameters of the plastron genital pore, and the carapace length-to-width ratio calculation formula is:

[0020] ;

[0021] in, is the carapace length-to-width ratio, is the longest straight-line distance of the carapace, is the widest straight-line distance of the carapace; the shape parameters of the plastron genital pore include area, perimeter and circularity; the circularity calculation formula is:

[0022]

[0023] in, is the circularity, is the area of ​​the genital pore, is the circumference of the genital pore area.

[0024] Further preferably, in the feature fusion process, an adaptive weight allocation algorithm is used to fuse features from different perspectives, and the weight calculation method is:

[0025] ;

[0026] in, For the The weight of the view feature, For the The identification confidence of each view feature, is the Sigmoid activation function, is the number of viewing angles.

[0027] Further preferably, the training process of the male-female discrimination classifier includes: collecting male and female sample images of blue crabs, performing multi-view acquisition and preprocessing on the sample images, extracting multimodal features and constructing a training data set; using a cross entropy loss function as a training objective function, the formula is:

[0028] ;

[0029] in, is the number of training samples, is the true label of the sample, female is 1, male is 0, Predict probability values ​​for the classifier; optimize the classifier parameters using the stochastic gradient descent algorithm until the loss function converges.

[0030] Further preferably, the dynamic screening actuator includes a vibrating sorting conveyor belt and a pneumatic sorting device. The vibrating sorting conveyor belt can separate the blue crabs individually and transport them to the sorting area. The pneumatic sorting device blows the corresponding blue crabs into female or male collection containers through high-pressure airflow according to the screening control signal of the intelligent decision-making control module. The airflow injection angle is 45°-60° to the conveyor belt plane, and the injection pressure is dynamically adjusted according to the size of the blue crabs.

[0031] Technical effects:

[0032] The present invention addresses the problems of low efficiency of manual identification in the background technology and poor identification accuracy due to the single perspective of existing automation technology. The system uses at least three high-definition cameras of a multi-perspective image acquisition module to collect images of blue crabs in all directions to avoid missing key features; and uses image preprocessing units, feature fusion identification modules, etc. to realize multimodal feature intelligent analysis. Compared with traditional methods, the system realizes the automation of the entire process from image acquisition to screening, greatly improving the efficiency of male and female identification of blue crabs, reducing labor costs and subjective errors; through multi-perspective information fusion and deep learning technology, it significantly improves the identification accuracy, meets the needs of high-throughput and precise screening in large-scale farming, and provides effective technical support for the intelligent development of the blue crab farming industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a block diagram of the intelligent identification and screening system for male and female blue crabs in this application;

[0034] Figure 2 This is a flow chart of the multi-view image recognition method of this application. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] See also Figure 1 Traditional technical solutions have the following technical problems: Traditional mud crab sex identification relies on manual operation, resulting in low efficiency, poor accuracy, and high labor intensity. Furthermore, existing automatic identification technologies often have a single perspective and incomplete feature extraction, making it difficult to accurately identify the sex characteristics of mud crabs. Based on this, this embodiment provides an intelligent mud crab sex identification and screening system, comprising a multi-perspective image acquisition module, an image preprocessing unit, a feature fusion identification module, an intelligent decision-making control module, and a dynamic screening actuator. The multi-perspective image acquisition module includes at least three high-definition cameras arranged at different angles to capture multi-angle images of the mud crab's carapace, plastron, and lateral carapace. The image preprocessing unit performs grayscale normalization, noise removal, and region of interest extraction on the captured images. The feature fusion identification module includes a multimodal feature extraction network based on deep learning and a sex identification classifier, capable of fusing texture, geometry, and color features from multiple perspectives and outputting an identification result. The intelligent decision-making control module generates a corresponding screening control signal based on the identification result. The dynamic screening actuator performs sex identification and screening of mud crabs based on the screening control signal.

[0037] This solution aims to solve the problem of how to efficiently and accurately collect multi-view images of blue crabs, and realize automated and high-precision sex identification and screening through multimodal feature fusion and intelligent decision-making, overcoming the shortcomings of manual and existing automatic technologies.

[0038] It is worth mentioning that this embodiment builds the basic architecture of the intelligent identification and screening system for male and female blue crabs, covering a multi-view image acquisition module, an image preprocessing unit, a feature fusion identification module, an intelligent decision-making control module, and a dynamic screening actuator. The multi-view image acquisition module uses at least three high-definition cameras at different angles to capture all-round images of the carapace, plastron, and lateral carapace of blue crabs, providing sufficient data for subsequent analysis; the image preprocessing unit performs grayscale normalization, noise elimination, and region of interest extraction on the original image to improve image quality and analysis targeting; the feature fusion identification module uses deep learning technology to fuse multimodal features and achieve male and female identification; the intelligent decision-making control module generates a screening signal based on the identification results; and the dynamic screening actuator completes the final male and female classification and screening operation.

[0039] The technical effects achieved by the above scheme include: the system can obtain complete appearance information of blue crabs through multi-perspective image acquisition, avoiding feature omissions caused by a single perspective; image preprocessing improves image quality, laying the foundation for accurate feature extraction and identification; the application of multimodal feature fusion and deep learning technology significantly improves the accuracy of male and female identification; intelligent decision-making and dynamic screening execution agencies realize automated operation, greatly improving identification and screening efficiency, reducing labor costs and labor intensity, and ensuring the reliability of screening results.

[0040] For example, traditional technical solutions have the following technical problems: if the number of cameras is insufficient or the layout is unreasonable, it may not be possible to capture images of key parts of the blue crab, affecting the accuracy of male and female identification; improper angle setting will also cause image deformation or unclear features, which also reduces the identification effect. Based on this, the multi-view image acquisition module has four high-definition cameras, which are respectively set directly above, directly below, to the left and to the right of the blue crab support platform, and the angle between the optical axis of each camera and the horizontal plane of the support platform is 30°-60°, which can cover the complete outer surface morphology of the blue crab.

[0041] This solution is dedicated to solving the problem of how to optimize the number and layout of cameras to ensure the acquisition of high-quality and complete multi-view images of blue crabs, providing strong support for subsequent accurate identification.

[0042] It is worth mentioning that, based on the above, this embodiment further clarifies the camera setup details of the multi-view image acquisition module. The number of high-definition cameras is specifically set to four, located directly above, directly below, to the left, and to the right of the blue crab's support platform. The angle between the optical axis of each camera and the horizontal plane of the support platform is controlled at 30°-60°. This specific layout ensures that the outer surface morphology of the blue crab can be fully covered, obtaining comprehensive and no-blind-angle image information.

[0043] The technical benefits achieved by this solution include: The rational layout of the four cameras captures mud crabs from multiple key angles, ensuring that all features of the carapace, plastron, and lateral carapace are clearly captured, avoiding identification errors caused by missing viewing angles. The specific angle settings ensure that the captured images are more consistent with the actual morphology, reduce image distortion, and improve image quality, thereby enhancing the effectiveness of multi-view image acquisition. This provides more accurate and comprehensive data for subsequent image analysis and sex identification, further enhancing the accuracy and reliability of the system's identification.

[0044] For example, the traditional technical solution has the following technical problems: During the image acquisition process of blue crabs, noise will inevitably be introduced into the image due to environmental factors, equipment performance and other reasons. The traditional fixed window filtering algorithm cannot adapt to the processing requirements of different noise density areas, and may blur the image details while removing noise, affecting subsequent feature extraction and identification. The noise elimination process of the image preprocessing unit adopts an adaptive median filtering algorithm, which dynamically adjusts the filter window size according to the noise density of the local area of ​​the image. The window size calculation formula is:

[0045] ;

[0046] in, is the side length of the filtering window, is the coordinate in the image The absolute difference between the pixel gray value at and the neighborhood mean, is the number of pixels in the current neighborhood, is the preset noise discrimination threshold.

[0047] This solution aims to solve the problem of how to adaptively adjust the filtering parameters according to the actual situation of image noise, so as to effectively eliminate noise while retaining the useful information and details of the image to the greatest extent.

[0048] : Indicates the window side length of the adaptive median filter, which is an odd number, such as 3×3, 5×5, etc., through the rounding function Ensure that the result is an integer to maintain window symmetry.

[0049] : Measures the difference between a single pixel and the neighborhood mean, reflecting the intensity of local noise. It is calculated as the absolute difference between the current pixel's grayscale value and the neighborhood mean. A larger difference indicates a higher likelihood of noise interference for the pixel.

[0050] : The number of pixels in the current neighborhood, which is related to the window size, such as a 3×3 window corresponds to , used for normalized noise intensity calculation.

[0051] : The preset noise discrimination threshold is the key parameter for distinguishing noise pixels from real signals and needs to be calibrated in advance according to the noise characteristics of the blue crab image.

[0052] The formula dynamically resizes the window by following these steps:

[0053] Step 1: Calculate the neighborhood mean. Calculate the pixel grayscale mean within the initial window (e.g. 3×3) with the current pixel as the center. , as a reference value for the local background.

[0054] Step 2: Calculate the absolute difference between the noise intensity sum and the mean of all pixels in the neighborhood The larger the value, the denser the noise in the window.

[0055] Step 3: Normalize the noise intensity by Convert the noise intensity to a relative value: when the value is 0, it means there is no noise in the window, and the window size remains at the minimum value of 3; when the value increases, the window side length is proportionally expanded, increasing by 2 each time, and maintaining an odd number to enhance the denoising ability.

[0056] Step 4: Rounding to determine the window. Ensure that the window side length is an integer by rounding down to avoid image distortion caused by asymmetric windows.

[0057] Disadvantages of traditional fixed window filtering Traditional median filtering uses a fixed window, such as 3×3, which has poor adaptability to changes in noise density: a small window cannot completely remove noise in high-noise areas, and a large window tends to blur edges in clean areas, such as the texture of the blue crab carapace and the outline of the genital pore.

[0058] This formula dynamically adjusts the window based on the noise density to achieve on-demand filtering:

[0059] In the smooth areas of the blue crab image, such as the pure color part of the carapace, there is less noise and the window is kept to the minimum value to protect the details;

[0060] In noise-intensive areas, such as plastron folds or reflective areas, the window automatically expands to enhance the ability to suppress salt and pepper noise and Gaussian noise.

[0061] The adaptively filtered images can not only effectively remove interference such as camera noise and water surface reflections, but also retain key identification features such as the carapace edge and genital pore texture, providing high-quality input for subsequent extraction of regions of interest and significantly improving the reliability of feature extraction.

[0062] It is worth mentioning that: for the image preprocessing unit in the above embodiment, the noise elimination process is described in detail using an adaptive median filtering algorithm. This algorithm dynamically adjusts the filter window size based on the noise density of the local area of ​​the image through a specific formula. In the formula, Represents the side length of the filter window, which is the absolute difference between the grayscale value of each pixel in the image and the neighborhood mean , the number of pixels in the current neighborhood And preset noise discrimination threshold The appropriate window size is calculated to effectively suppress noise.

[0063] The technical benefits achieved by this approach include: an adaptive median filter algorithm that flexibly adjusts the filter window based on the local noise level of the image, increasing the window size in areas with high noise density to enhance denoising, and reducing the window size in areas with less noise to preserve image details. Compared to traditional fixed-window filtering, this algorithm not only removes noise but also better preserves important features of the blue crab image, such as texture and edges. This improves the quality of image preprocessing, provides higher-quality image data for subsequent accurate feature extraction and sex identification, and thus enhances the accuracy and stability of the entire system.

[0064] For example, the traditional technical solutions have the following technical problems: the blue crab clasping crab image contains a large amount of background information and non-critical areas. Directly analyzing the entire image will increase the amount of calculation and affect the accuracy of feature extraction. Accurately locating the carapace contour and segmenting the area outside the plastron genital pore will help focus on key identification features. Traditional methods are difficult to accurately extract these regions of interest under complex backgrounds, resulting in deviations in subsequent feature analysis. Based on this, the region of interest extraction process includes carapace contour positioning and plastron genital pore area segmentation. The carapace contour positioning adopts an edge detection algorithm based on the active shape model. By establishing a priori shape model of the blue crab clasping crab carapace, the carapace edge in the image is iteratively fitted; the plastron genital pore area segmentation adopts a threshold segmentation algorithm based on color space conversion, which converts the image from RGB color space to HSV color space, extracts the V channel component and uses the Otsu algorithm to determine the optimal segmentation threshold.

[0065] This solution aims to solve the problem of how to efficiently and accurately extract regions of interest closely related to sex identification from blue crab images, thereby improving the pertinence and efficiency of image analysis.

[0066] It is worth mentioning that the specific method for extracting and processing the region of interest in the above embodiment is further refined. The carapace outline is located using an edge detection algorithm based on an active shape model. By establishing a priori shape model of the carapace of the mud crab, the carapace edge is iteratively fitted in the image to accurately obtain the carapace outline. The plastron genital pore region is segmented using a threshold segmentation algorithm based on color space conversion. The image is first converted from RGB color space to HSV color space, the V channel component is extracted, and then the Otsu algorithm is used to determine the optimal segmentation threshold to achieve accurate segmentation of the plastron genital pore region.

[0067] For example, traditional technical solutions have the following technical problems: during the operation of the blue crab sex identification and screening system, it is necessary to timely understand the system's operating status and identification results in order to make optimization adjustments; at the same time, due to the different growth stages and species of blue crabs, their male and female characteristics may vary, and traditional systems are difficult to adapt to such changes. The intelligent decision-making control module is integrated with a real-time data monitoring unit and a model update interface. The real-time data monitoring unit can monitor the accuracy and efficiency of identification and screening in real time and generate status reports; the model update interface supports online update training of the male and female identification classifier through external data input to adapt to the identification needs of blue crabs at different growth stages and species.

[0068] This solution aims to solve the problem of how to monitor the system operation status in real time, detect problems in time and optimize them, and how to make the system adaptive so that it can update the model according to the different characteristics of blue crabs and ensure the accuracy and effectiveness of identification and screening.

[0069] It's worth noting that this solution expands the functionality of the intelligent decision-making control module in the previous embodiment, integrating a real-time data monitoring unit and a model update interface. The real-time data monitoring unit monitors key indicators such as identification and screening accuracy and efficiency in real time and generates status reports. The model update interface supports online training of the sex discrimination classifier via external data input, adapting to the identification needs of blue crabs at different growth stages and species.

[0070] The technical effects achieved by the above scheme include: the real-time data monitoring unit can provide real-time feedback on information such as the system's identification accuracy and efficiency. Through status reports, operators can promptly understand the system's operating status, identify potential problems, and take corresponding measures to optimize and adjust, ensuring that the system is always in good operating condition. The model update interface enables the male and female identification classifier to be updated and trained online based on the data of blue crabs at different growth stages and varieties, continuously optimize model parameters, and improve the classifier's ability to recognize different features. This adaptive capability enables the system to better adapt to the complex and changing situations in actual production, continuously maintain a high identification and screening accuracy and stability, and enhance the practicality and adaptability of the system.

[0071] See also Figure 2 Traditional technical solutions have the following technical problems: Most existing methods for identifying male and female blue crabs lack systematicity and comprehensiveness, and cannot fully utilize multi-view image information and multimodal features, resulting in low identification accuracy and poor stability. Based on this, this embodiment provides a multi-view image recognition method, which is applied to the blue crab male and female intelligent identification and screening system as described in any of the above, including the following steps:

[0072] S1: Multi-view image acquisition: collect multi-angle images of the carapace, plastron and lateral carapace of the blue crab through at least three high-definition cameras set at different angles;

[0073] S2: Image preprocessing: grayscale normalization, noise elimination and region of interest extraction of the collected images;

[0074] S3: Multimodal feature extraction: Use a deep learning network to extract the texture features, geometric features, and color features of each view image. The texture feature uses an improved local binary pattern algorithm to calculate the 8-neighborhood local binary pattern value of each pixel. The formula is:

[0075] ;

[0076] in, is the center pixel coordinate, is the gray value of the center pixel, is the gray value of the neighborhood pixel, is a symbolic function, when hour ,otherwise ;

[0077] S4: Feature Fusion and Identification: The multimodal features of each view are spatially aligned and dimensionally spliced, and then input into the male-female identification classifier for training and classification. The classifier uses a multi-layer perceptron model combined with an attention mechanism to adjust the contribution of different viewpoint features to the identification results through attention weights.

[0078] S5: Screening decision and execution: Generate corresponding screening control signals according to the identification and classification results, and drive the dynamic screening actuator to perform sex-classification screening on blue crabs.

[0079] : The coordinates of the center pixel of the current LBP value calculation are used as the reference point of the neighborhood.

[0080] : Grayscale value of the center pixel, ranging from 0 to 255, 8-bit grayscale image, used to measure the relative brightness of neighboring pixels.

[0081] : The grayscale values ​​of the 8 neighboring pixels with the central pixel as the center and a radius of 1, arranged clockwise. to , constituting the comparison unit of local texture.

[0082] : Sign function, which converts the grayscale difference between the neighborhood pixel and the central pixel into a binary value, 0 or 1, to form a binary code of the local texture.

[0083] This formula extracts texture features through the following steps:

[0084] Step 1: Neighborhood comparison Calculate the 8 neighboring pixels separately :If the gray value of the neighborhood pixel is greater than or equal to the central pixel , the sign function outputs 1, otherwise it outputs 0.

[0085] Step 2: Binary encoding weights the outputs of the 8 symbol functions by position. , Corresponding to the lowest bit, Corresponding to the highest bit, weighted summation is performed to obtain an 8-bit binary number 0-255, which is the LBP value of the pixel.

[0086] Step 3: Statistical histogram, perform histogram statistics on the LBP value of the entire image to generate a texture feature vector that reflects the distribution frequency of different texture patterns in the image.

[0087] Limitations of traditional LBP: Traditional LBP uses a fixed neighborhood, such as the 8-neighborhood. However, the texture of the area outside the genital pore of the plastron of the blue crab is delicate, and there is an arc-shaped structure on the edge of the carapace. It is difficult for a fixed neighborhood to capture the direction-sensitive characteristics.

[0088] This embodiment improves the discriminative power of texture features through the following optimizations:

[0089] Neighborhood precision sampling: covers all directions of the center pixel, including up, down, left, right, and four diagonal directions, fully capturing anisotropic features such as the reticular texture of the mud crab carapace and the circular outline of the plastron;

[0090] Grayscale difference symbolization: Converts absolute grayscale values ​​into relative values ​​to eliminate the impact of uneven lighting, such as differences in reflection from cameras at different angles, on texture features and enhance feature robustness.

[0091] Multi-view joint coding: Combining the multi-view images in the above embodiment, the carapace, plastron, and lateral carapace are calculated for each view separately. Through the adaptive weight fusion in the above embodiment, a global texture feature across views is formed, which effectively distinguishes the differences in texture complexity between male and female individuals. For example, the texture around the female genital pore is more regular, and the edge of the male carapace is sharper.

[0092] Texture features are crucial for sex differentiation, serving as a key indicator of male and female morphological differences in mud crabs. The texture of the genital pore area on the female plastron is uniform, with a concentrated distribution of LBP values. The serrated structure on the lateral margins of the male carapace exhibits more dramatic LBP variations. Texture features extracted using this formula complement geometric and color features to form a multimodal feature system, significantly improving the accuracy of the classifier.

[0093] This solution aims to solve the problem of how to efficiently integrate multi-view image data, extract and fuse multimodal features through a complete and systematic method and process, and achieve high-precision image recognition and automated screening of male and female blue crabs to meet actual production needs.

[0094] It's worth noting that this embodiment proposes a multi-view image recognition method for male and female blue crabs, encompassing multiple steps: multi-view image acquisition, image preprocessing, multimodal feature extraction, feature fusion and identification, and screening decision-making and execution. Multi-view image acquisition captures multi-angle images; image preprocessing improves image quality; multimodal feature extraction utilizes a deep learning network to extract texture, geometry, and color features; feature fusion and identification fuses multi-view features for classification and identification; and finally, based on the identification results, screening decision-making and execution implement male and female classification.

[0095] The technical effects achieved by the above scheme include: this method comprehensively collects images of blue crabs through multi-view image acquisition, providing rich data for feature extraction; image preprocessing ensures the quality of input data; multimodal feature extraction extracts features from multiple dimensions and fully exploits image information; feature fusion and identification utilize deep learning and attention mechanisms to effectively integrate multi-view features and improve classification and identification accuracy; and screening decision-making and execution achieve automated operations. The entire method process is systematic and efficient, significantly improving the accuracy and screening efficiency of male and female blue crab image recognition. Compared with traditional methods, it has greater practicality and reliability and can be better applied to the production practices of the blue crab farming industry.

[0096] For example, the traditional technical solution has the following technical problems: Geometric features are important discriminant factors in the identification of male and female blue crabs, but the extraction and quantification of geometric features by previous methods are not accurate and comprehensive enough, and cannot fully play their role in identification. Based on this, in the multimodal feature extraction, the geometric features include the carapace length-to-width ratio, the position coordinates and shape parameters of the plastron genital pore, and the carapace length-to-width ratio calculation formula is:

[0097] ;

[0098] in, is the carapace length-to-width ratio, is the longest straight-line distance of the carapace, is the widest straight-line distance of the carapace; the shape parameters of the plastron genital pore include area, perimeter and circularity; the circularity calculation formula is:

[0099]

[0100] in, is the circularity, is the area of ​​the genital pore, is the circumference of the genital pore area.

[0101] : A dimensionless parameter that reflects the degree of narrowness and length of the carapace. The larger the value, the narrower the carapace, and vice versa, the closer it is to a square.

[0102] : The maximum Euclidean distance between two points on the carapace contour. After fitting the carapace edge using the active shape model, the length of the long side of the circumscribed rectangle of the contour is calculated.

[0103] : The maximum width of the carapace outline perpendicular to the longest distance, that is, the length of the short side of the circumscribed rectangle. 2. Biological significance and identification value There are significant differences in the carapace morphology between male and female individuals of blue crabs: the female carapace is wider and shorter, with a length-to-width ratio of Close to 1.2-1.5 to accommodate the reproductive system; the male carapace is narrower and longer. The formula can reach 1.8-2.2, which is suitable for sports needs. This formula provides intuitive shape characteristics for male and female classification by quantifying the geometric proportions of the carapace. Compared with traditional manual visual inspection, numerical The value avoids subjective errors and can be automatically extracted through machine learning models to improve identification efficiency.

[0104] : A dimensionless parameter with a value range of 0-1. The closer it is to 1, the closer the area is to a circle.

[0105] : The pixel area outside the genital pore of the plastron. After extracting the area using the threshold segmentation algorithm, the number of internal pixels is counted, and the number of pixels with a gray value of 1 is counted.

[0106] : The contour perimeter of the genital pore area is extracted using a pixel-level edge detection algorithm, and the sum of the Manhattan distances of the contour pixels is calculated. The pixel-level edge detection algorithm can use the Sobel operator.

[0107] The shape of the genital pore is the core feature for identifying the sex of blue crabs: the female genital pore is round or oval, and the roundness is Due to differences in reproductive structures, the genital pore area of ​​males is mostly narrow or irregular. .

[0108] This formula converts the qualitative circular / non-circular judgment into quantitative Value, combined with area and perimeter , construct three-dimensional geometric feature vectors ( Compared to a single area or perimeter, circularity eliminates scale effects, such as differences in genital pore area between individuals of different sizes, and only reflects shape characteristics, making it more specific for sex differentiation.

[0109] This scheme aims to solve the problem of how to accurately calculate the geometric characteristics of the genital pores on the carapace and plastron of blue crabs so that they can accurately reflect the differences in morphological structure between male and female individuals and provide more reliable characteristic parameters to improve the accuracy of sex identification.

[0110] It is worth mentioning that: based on the above embodiment, this solution specifically explains the calculation method of geometric features in multimodal feature extraction. The longest straight-line distance through the carapace Distance to the widest straight line The ratio of the calculation; the circularity of the genital pore of the plastron By its area and perimeter , calculated through a specific formula, and also involves parameters such as the area and circumference of the genital pore area. These geometric features serve as important identification basis for subsequent male and female classification.

[0111] The technical benefits achieved by this approach include: By accurately calculating geometric features such as the carapace aspect ratio and the area, circumference, and circularity of the plastron genital pore, it is possible to deeply explore the morphological differences between male and female blue crabs. These quantified geometric features have clear physical meaning and distinguishability. During the feature fusion and identification process, they complement other modal features, providing richer and more discriminative input information for the sex discrimination classifier, significantly improving the accuracy and reliability of sex discrimination and enabling the system to more accurately identify the sex of blue crabs.

[0112] For example, traditional technical solutions have the following technical problems: During the multi-view image feature fusion process, image features from different viewpoints have different degrees of importance for male-female discrimination. The traditional fixed-weight fusion method cannot dynamically adjust the contribution of each viewpoint feature according to the actual situation, which may cause important features to be ignored or minor features to affect the discrimination results. Based on this, the feature fusion process adopts an adaptive weight allocation algorithm to fuse features from different viewpoints. The weight calculation method is:

[0113] ;

[0114] in, For the The weight of the view feature, For the The identification confidence of each view feature, is the Sigmoid activation function, is the number of viewing angles.

[0115] :For the The fusion weight of the perspective features satisfies ;No. The viewing angle features include directly above, directly below, left, and right.

[0116] :Through the male and female identification classifier The prediction confidence of each perspective feature ranges from [0,1]. The higher the value, the greater the potential contribution of the perspective feature to the identification result.

[0117] : Sigmoid function, the confidence Mapping to the (0,1) interval enhances the discrimination of small confidence differences, such as The mapping is 0.65, The mapping is 0.5 to amplify the weight differences of effective features.

[0118] : Number of viewing angles, corresponding to 4 cameras, i.e. 2. Algorithm Logic Analysis This formula implements the adaptive allocation of feature fusion weights through the following steps:

[0119] Step 1: Confidence normalization, using the Sigmoid function to normalize the original confidence Perform nonlinear transformation to suppress the excessive influence of extreme values ​​and avoid single-view features dominating the fusion results; such as confidence close to 1 or 0.

[0120] Step 2: Weight calculation, normalize the Sigmoid output of each view so that the sum of the weights is 1. For example, if the view from the top is , the other three perspectives The sum is 1.2, then , that is, the perspective feature accounts for 40% of the weight during fusion.

[0121] The drawback of traditional fixed-weight fusion: Traditional multi-view fusion uses an average weight, ignoring the differences in discriminative power between features from different viewpoints. For example, the plastron of a mud crab is primarily captured by the camera directly below. If the confidence level of this viewpoint feature is high, a fixed weight will underestimate its importance, leading to discrimination errors.

[0122] Breakthrough point of the adaptive mechanism: This algorithm achieves intelligent focusing of feature fusion through confidence-weight mapping.

[0123] For key perspectives, such as the plastron image taken directly below, if the classifier has high confidence in its features, such as , then give high weight, It can reach 30%-40%, strengthening the influence of core identification characteristics such as genital pore shape and color;

[0124] For secondary perspectives, such as edge textures shot from the side, if the confidence is low, such as , the weight automatically decreases, About 10%-15%, reducing interference from irrelevant features.

[0125] Combined with the attention mechanism multi-layer perceptron (MLP) in the aforementioned embodiment, this weight allocation algorithm enables the classifier to dynamically focus on key features from different viewpoints. This ensures the stability of identification results, especially when individual mud crabs experience posture changes, such as when they roll over or are obscured. By increasing the weight of visible viewpoints, the classifier can achieve a 15%-20% improvement in identification accuracy compared to fixed weights, meeting the precision requirements of high-throughput screening in aquaculture.

[0126] This solution aims to solve the problem of how to adaptively allocate weights based on the actual identification capabilities of each perspective feature, achieve more reasonable and efficient feature fusion, and improve the accuracy of male-female identification.

[0127] It is worth mentioning that: for the feature fusion process in the above embodiment, the method of fusing different view features using the adaptive weight allocation algorithm is described in detail. The weight of each view feature is calculated by the formula , and the The identification confidence of the view feature Related, using Sigmoid activation function The confidence is processed and then the weight of each perspective feature is obtained through normalization calculation to adjust the contribution of different perspective features to the identification results.

[0128] The technical effects achieved by the above scheme include: an adaptive weight allocation algorithm that dynamically assigns weights to different viewpoint features based on the identification confidence of each viewpoint feature. Viewpoint features with strong identification capabilities and containing important identification information are given higher weights, allowing them to play a greater role in the identification process; features with weaker identification capabilities are given lower weights to reduce their interference with the identification results. This method achieves optimized feature fusion, making the feature fusion results more reflective of the true feature differences between male and female blue crabs, effectively improving the performance of the male-female identification classifier, and thereby enhancing the identification accuracy and stability of the entire multi-view image recognition method.

[0129] For example, the traditional technical solution has the following technical problems: the performance of the male-female discrimination classifier directly affects the accuracy of male-female discrimination of blue crabs, and the training quality of the classifier is the key to determining its performance. Traditional training methods may have problems such as unscientific sample selection, unreasonable training objective function, and low parameter optimization efficiency, resulting in poor generalization ability of the classifier and low identification accuracy. Based on this, the training process of the male-female discrimination classifier includes: collecting male and female sample images of blue crabs, performing multi-view acquisition and preprocessing on the sample images, extracting multimodal features and constructing a training data set; using the cross entropy loss function as the training objective function, the formula is:

[0130] ;

[0131] in, is the number of training samples, is the true label of the sample, female is 1, male is 0, Predict probability values ​​for the classifier; optimize the classifier parameters using the stochastic gradient descent algorithm until the loss function converges.

[0132] is the objective function value of the training process. The smaller the value, the smaller the difference between the classifier prediction and the true label.

[0133] : The total number of samples participating in training must include enough male and female samples to avoid class imbalance affecting model training.

[0134] : Binary classification label, female samples are 1 and male samples are 0, reflecting the true attributes of the samples.

[0135] :The classifier is The predicted probability of a sample takes values ​​[0,1], such as This means the model believes the probability that the sample is female is 80%.

[0136] The cross entropy loss function measures the prediction error through the following mechanism:

[0137] Single sample loss, when the sample is female When , the loss term is , if the predicted probability Close to 1, the loss approaches 0; if the predicted probability is low, the loss increases sharply , forcing the model to improve the predicted probability of female samples.

[0138] When the sample is male When , the loss term is , if the predicted probability If the predicted probability is high, such as 0.7, the loss increases. , forcing the model to reduce the probability of misclassifying it as female.

[0139] The losses of all samples are averaged to obtain the overall loss, which is used as the target for model parameter optimization. Using the stochastic gradient descent algorithm, the classifier parameters, such as the weights of the neural network, are updated along the opposite direction of the loss gradient, gradually reducing the loss until convergence.

[0140] Compared with mean square error, the cross entropy of this embodiment has two major advantages in classification problems:

[0141] Gradient stability: When the predicted probability is close to 0 or 1, the gradient of cross entropy does not disappear, while the gradient of MSE approaches 0, causing training to stagnate;

[0142] Probability sensitivity: It directly acts on the classification probability, naturally matches the output logic of the classifier (such as the Softmax layer), and is more suitable for binary classification tasks.

[0143] The optimization of model training combined with multimodal feature input can drive the classifier to focus on learning the difference between male and female samples in carapace length-to-width ratio ( ), plastron roundness ( ), texture LBP value and other features. For example, when a male sample is misjudged as a female, its carapace is long and narrow. , low roundness of plastron The features will be back-propagated through the loss function, forcing the model to adjust the feature weights and enhance the ability to identify these key differences.

[0144] The generalization ability ensures that the classifier trained by the cross-entropy loss function can effectively handle the differences between blue crab varieties, such as populations in different regions and growth stages, such as the changes in characteristics between larvae and adults. Combined with the online update mechanism, it can dynamically adapt to the new sample distribution, avoid overfitting, and ensure that the long-term identification accuracy remains stable at above 95%.

[0145] This solution aims to solve the problem of how to scientifically and effectively train the sex identification classifier to improve its accuracy and generalization ability in sex identification of blue crabs to adapt to the identification needs of different scenarios and samples.

[0146] It is worth mentioning that the training process of the sex discrimination classifier in the above embodiment is described in detail. This includes collecting male and female sample images of blue crabs, performing multi-view acquisition and preprocessing, extracting multimodal features to construct a training dataset, using a cross-entropy loss function as the training objective function, optimizing the classifier parameters using a stochastic gradient descent algorithm, and continuously adjusting the classifier model until the loss function converges and the classifier reaches optimal performance.

[0147] The technical effects achieved by the above scheme include: ensuring the quality and diversity of training data by systematically collecting sample images, performing multi-view acquisition and preprocessing; using the cross-entropy loss function as the training objective function to accurately measure the difference between the classifier's predictions and the true labels, providing a clear direction for parameter optimization; and using the stochastic gradient descent algorithm to efficiently optimize the classifier parameters. Through this scientific training process, the sex discrimination classifier can learn the characteristic patterns of male and female blue crab samples, improving its ability to distinguish and generalize across different samples. This enables the classifier to more accurately distinguish the sex of blue crabs in practical applications, improving the reliability and practicality of the entire image recognition method.

[0148] For example, the traditional technical solution has the following technical problems: in the screening process after the male and female blue crabs are identified, an efficient and reliable actuator is needed to accurately classify and collect the identified blue crabs. The traditional screening method may have problems such as low sorting efficiency, damage to blue crabs, and inability to adapt to individuals of different sizes. The dynamic screening actuator includes a vibrating sorting conveyor belt and a pneumatic sorting device. The vibrating sorting conveyor belt can separate the blue crabs individually and transport them to the sorting area. The pneumatic sorting device blows the corresponding blue crabs into female or male collection containers through high-pressure airflow according to the screening control signal of the intelligent decision-making control module. The airflow injection angle is 45°-60° to the conveyor belt plane, and the injection pressure is dynamically adjusted according to the size of the blue crabs.

[0149] This proposal aims to solve the problem of how to design an automated and adjustable dynamic screening actuator to achieve fast, accurate and non-destructive classification and screening of blue crabs into males and females to meet actual production needs.

[0150] It is worth mentioning that the dynamic screening actuator in the above embodiment is described in detail. It consists of a vibrating sorting conveyor belt and a pneumatic sorting device. The vibrating sorting conveyor belt is responsible for separating individual blue crabs and transporting them to the sorting area to achieve orderly transportation. The pneumatic sorting device, based on the screening control signal from the intelligent decision-making control module, uses high-pressure airflow at a specific angle and pressure to blow the corresponding blue crabs into female or male collection containers to complete the male-female classification and screening operation. The airflow injection pressure can be dynamically adjusted according to the size of the blue crabs.

[0151] The technical effects achieved by the above scheme include: the vibrating sorting conveyor belt can separate and transport blue crabs in an orderly manner, ensuring the continuity and stability of the screening operation; the pneumatic sorting device can achieve accurate sorting of blue crabs by precisely controlling the airflow injection angle and pressure, avoiding the errors and inefficiencies of manual operation. At the same time, the airflow pressure is dynamically adjusted according to the size of the blue crabs, effectively preventing damage to the blue crabs during the sorting process and improving the quality and efficiency of screening. The entire set of dynamic screening actuators realizes automated and precise male and female classification and screening, reduces labor costs, improves production efficiency, and has good practicality and reliability.

[0152] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An intelligent identification and screening system for male and female blue crabs, characterized in that: The system comprises a multi-view image acquisition module, an image preprocessing unit, a feature fusion identification module, an intelligent decision control module, and a dynamic screening actuator; the multi-view image acquisition module comprises at least three high-definition cameras arranged in a surrounding manner at different angles, for acquiring multi-angle images of the carapace, plastron, and lateral carapace of a blue crab; the image preprocessing unit performs grayscale normalization, noise elimination, and region of interest extraction on the acquired images; the feature fusion identification module comprises a multimodal feature extraction network based on deep learning and a male-female identification classifier, capable of fusing texture, geometry, and color features of multi-view images and outputting identification results; the intelligent decision control module generates corresponding screening control signals based on the identification results; The dynamic screening actuator performs sex-classification screening of blue crabs according to the screening control signal; the multi-view image acquisition module includes four high-definition cameras, which are respectively arranged directly above, directly below, to the left, and to the right of the blue crab support platform, and the angle between the optical axis of each camera and the horizontal plane of the support platform is 30°-60°, which can cover the complete outer surface morphology of the blue crab; the noise elimination processing of the image preprocessing unit adopts an adaptive median filtering algorithm, and dynamically adjusts the filter window size according to the noise density of the local area of ​​the image. The window size calculation formula is: ； in, is the side length of the filtering window, is the coordinate in the image The absolute difference between the pixel gray value at and the neighborhood mean, is the number of pixels in the current neighborhood, is a preset noise discrimination threshold; the region of interest extraction processing includes carapace contour positioning and plastron genital pore area segmentation, the carapace contour positioning adopts an edge detection algorithm based on an active shape model, by establishing a priori shape model of the blue crab carapace, the carapace edge in the image is iteratively fitted; the plastron genital pore area segmentation adopts a threshold segmentation algorithm based on color space conversion, the image is converted from RGB color space to HSV color space, the V channel component is extracted and the Otsu algorithm is used to determine the optimal segmentation threshold; the intelligent decision control module is integrated with a real-time data monitoring unit and a model update interface, the real-time data monitoring unit can monitor the accuracy and efficiency of identification and screening in real time, and generate a status report; The model update interface supports online updating and training of the sex identification classifier through external data input to adapt to the identification needs of blue crabs and crab species at different growth stages and species.

2. A multi-view image recognition method, applied to the blue crab male and female intelligent identification and screening system as claimed in claim 1, characterized in that: The following steps are involved: Multi-view image acquisition: High-definition cameras at at least three different angles are used to capture multi-angle images of the carapace, plastron, and lateral carapace of the blue crab. Image preprocessing: grayscale normalization, noise elimination and region of interest extraction of the collected images; Multimodal feature extraction: A deep learning network is used to extract the texture, geometric, and color features of each view image. The texture feature uses an improved local binary pattern algorithm to calculate the 8-neighborhood local binary pattern value of each pixel. The formula is: ； in, is the center pixel coordinate, is the gray value of the center pixel, is the gray value of the neighborhood pixel, is a symbolic function, when hour ,otherwise ; Feature fusion and identification: The multimodal features from each perspective are spatially aligned and dimensionally spliced, and then input into the male-female identification classifier for training and classification. The classifier uses a multi-layer perceptron model combined with an attention mechanism, and adjusts the contribution of different perspective features to the identification results through attention weights. Screening decision and execution: Generate corresponding screening control signals based on the identification and classification results to drive the dynamic screening actuator to perform sex-classification screening on blue crabs.

3. The image recognition method according to claim 2, wherein: In the multimodal feature extraction, the geometric features include the carapace length-to-width ratio, the position coordinates of the plastron genital pore, and the shape parameters. The carapace length-to-width ratio is calculated as follows: ； in, is the carapace length-to-width ratio, is the longest straight-line distance of the carapace, is the widest straight-line distance of the carapace; the shape parameters of the plastron genital pore include area, perimeter and circularity; the circularity calculation formula is: ； in, is the circularity, is the area of ​​the genital pore, is the circumference of the genital pore area.

4. The image recognition method according to claim 2, wherein: During the feature fusion process, an adaptive weight allocation algorithm is used to fuse features from different perspectives. The weight calculation method is: ； in, For the The weight of the view feature, For the The identification confidence of each view feature, is the Sigmoid activation function, is the number of viewing angles.

5. The image recognition method according to claim 2, wherein: The training process of the sex discrimination classifier includes: collecting male and female sample images of blue crabs, performing multi-view acquisition and preprocessing on the sample images, extracting multimodal features and constructing a training data set; using a cross entropy loss function as a training objective function, the formula is: ； in, is the number of training samples, is the true label of the sample, female is 1, male is 0, Predict probability values ​​for the classifier; optimize the classifier parameters using the stochastic gradient descent algorithm until the loss function converges.

6. The image recognition method according to claim 2, wherein: The dynamic screening actuator includes a vibrating sorting conveyor belt and a pneumatic sorting device. The vibrating sorting conveyor belt can separate blue crabs individually and transport them to the sorting area. The pneumatic sorting device blows the corresponding blue crabs into female or male collection containers through high-pressure airflow according to the screening control signal of the intelligent decision-making control module. The airflow injection angle is 45°-60° with the conveyor belt plane, and the injection pressure is dynamically adjusted according to the size of the blue crabs.

Images