A flexible switchable euphausia superba intelligent sorting method and device

Abstract

The application relates to a flexible switchable Euphausia superba intelligent sorting method and device, and belongs to the technical field of aquatic product processing and automatic equipment. The application not only realizes high-precision instance segmentation and flexible reference calculation of Euphausia superba through improvement of a deep learning network and a self-adaptive statistical decision model, but also creatively seamlessly converts a dynamic grading signal generated by front-end vision into a physical driving instruction of a servo motor, so that the rear-end spacing of a longitudinal involute sorting pipe is opened and closed in real time and automatically. The application establishes an intelligent linkage closed loop of "visual perception-dynamic decision-physical execution", can automatically and flexibly adjust a physical sorting channel according to real statistical characteristics of batches of materials, and realizes real flexible nondestructive intelligent grading.

Description

Technical Field

[0001] This invention relates to the field of aquatic product processing and automation equipment technology, and to a flexible and switchable intelligent sorting method and device for Antarctic krill. Background Technology

[0002] Antarctic krill, as a vast biological resource, is rich in high-quality protein and phospholipids, possessing extremely high economic value. In the processing of Antarctic krill, precise size sorting (grading) is a crucial step in enhancing product added value and meeting the needs of different subsequent processing techniques (such as whole-krill freezing, shelling and meat extraction, phospholipid extraction, etc.). According to current production standards in the deep-sea fishing and seafood processing industries, Antarctic krill are typically classified into three grades based on body length: large size (body length > 45mm, mainly used for high-value-added whole krill for consumption and shrimp meat shelling), medium size (body length between 35mm and 45mm, mainly used for extracting high-quality krill oil), and small size (body length < 35mm, mainly used for processing feed-grade krill meal).

[0003] However, existing Antarctic krill sorting technologies still have the following significant problems in practical applications: Mechanical devices lack flexibility, and grading accuracy is limited by fixed structures. Existing drum or tubular sorting equipment often uses fixed screen holes or spacing. For example, Chinese patent document CN114073269A discloses a "shipborne Antarctic krill grading device." Although such devices can achieve preliminary vibration screening, the physical dimensions of their grading screen holes are fixed during manufacturing. When the average body length of raw material batches changes in different fishing seasons or sea areas, this fixed spacing cannot be adaptively adjusted, easily leading to channel blockage or a severe decrease in sorting efficiency.

[0004] Conventional visual algorithms suffer from poor robustness and rigid judgment criteria. With the development of computer vision, some aquatic product testing equipment has introduced image recognition technology. For example, Chinese invention patent CN103056111A discloses a "shrimp quality detection and grading device based on machine vision technology," which proposes a method for grading shrimp based on the extraction of two-dimensional morphological features from images. Compared with Antarctic krill, shrimp are larger and have brighter colors, making them relatively easy to identify using conventional visual detection methods. However, Antarctic krill are characterized by their small size, semi-transparent bodies, light colors, and weak edge contours, making them difficult to distinguish from background impurities, resulting in a high false recognition rate and significant challenges in accurate identification. Furthermore, these conventional visual grading methods often rely on rigid "absolute pixel feature thresholds," which are insufficient to effectively overcome the physical deformation errors caused by the natural curling and lateral bending postures of Antarctic krill.

[0005] The "vision-machine" system is severely fragmented: the visual inspection and mechanical actuators of existing equipment are independent of each other. The statistical features of the group captured by the vision system cannot be directly converted into dynamic adjustment commands for the spacing of the physical sorting channels, and the mechanical device remains in a "fixed parameter sorting" state.

[0006] In summary, developing an intelligent sorting system that can both overcome the errors of semi-transparent target segmentation and curling deformation with high precision, and directly drive the adaptive opening and closing of the back-end mechanical channel through "visual perception" is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies, such as fixed mechanical sorting channels and lack of adaptability of absolute thresholds, and to provide a flexible and switchable intelligent sorting method and device for Antarctic krill. This invention not only achieves high-precision instance segmentation and flexible benchmark calculation of Antarctic krill by improving deep learning networks and adaptive statistical decision models, but also innovatively and seamlessly converts the dynamic grading signals generated by front-end vision into physical drive commands for servo motors, thereby controlling the opening and closing of the rear end spacing of the longitudinally involute sorting tube in real time and automatically. This invention establishes an intelligent linkage closed loop of "visual perception—dynamic decision-making—physical execution," which can automatically and flexibly adjust the physical sorting channel according to the actual statistical characteristics of material batches, achieving truly flexible and non-destructive intelligent grading.

[0008] The technical solution of the present invention is as follows: A flexible and switchable intelligent sorting method for Antarctic krill, comprising the following steps: Step 1: Acquisition of Antarctic krill image data and construction of augmented dataset; To support the precise classification of large, medium, and small levels, and to provide data support for the subsequent transformation of the "one-dimensional physical volume length" in the national grading standard into the "two-dimensional pixel area" resistant to curling interference, a high-fidelity dataset containing complete morphological features and physical benchmarks was first constructed, specifically including: Step 1.1, Morphological Physical Calibration and Image Acquisition: Antarctic krill samples of large, medium, and small sizes were randomly selected beforehand, and each sample was assigned a unique number. First, the krill samples were flattened without damaging the krill body, and their standard physical body length L was manually measured. The standard physical body length L is the distance from the base of the compound eye to the end of the telson. Then, krill samples with the same number were placed in the image acquisition area, and their visual images were acquired in a naturally bent or naturally positioned state. The corresponding two-dimensional pixel area S was extracted based on the target region of the krill in the visual image. Thus, a one-to-one correspondence was established between the standard physical body length L of the same sample and the pixel area S in the visual image under natural conditions, forming a "body length - pixel area" paired dataset, providing calibration data for subsequent conversion from the physical body length grading standard to the pixel area grading standard.

[0009] Step 1.2, Refined Connected Component Annotation: Pixel-level annotation of the Antarctic krill outline is performed using polygon annotation to obtain an Antarctic krill mask. Enhanced annotation is applied to the semi-transparent body connection points of the Antarctic krill to ensure the mask completely covers the entire "head-body-tail" profile, providing accurate ground truth values ​​for subsequent pixel-area-based morphological calculations and grading. Step 1.3, Sample Balancing and Augmentation: Data augmentation and quantity balancing were performed on Antarctic krill samples of large, medium, and small sizes. A combination of techniques, including flipping, rotating, sharpening, adjusting brightness and contrast, injecting salt-and-pepper noise, and enhancing the red channel (where salt-and-pepper noise was used to simulate impurities and bubble interference in the production line), was employed to expand the dataset size and prevent the model from overfitting to a specific size or background during training.

[0010] Step 2: Construct an Antarctic krill sorting model based on an improved YOLOv12 instance segmentation network and adaptive statistical decision-making. An improved YOLOv12 instance segmentation network is used to segment Antarctic krill images acquired and labeled in step one, extract pixel areas, and determine size levels. Specifically, this step uses the YOLOv12 instance segmentation network as the base network and improves its feature extraction, feature denoising, feature fusion, and post-processing decision stages to adapt it to detection scenarios where Antarctic krill are translucent, easily curled, have a low proportion of small targets, and have significant background interference from the flow line. The YOLOv12 instance segmentation network includes a backbone network, a feature fusion network, and a segmentation prediction head. The backbone network is used to extract multi-scale features from the input image and output shallow, medium, and deep feature maps; the feature fusion network is used to upsample, downsample, and fuse channels for features at different scales; and the segmentation prediction head is used to output the category information, bounding box information, and instance mask information of the Antarctic krill target. The present invention inputs the acquired Antarctic krill images into the YOLOv12 instance segmentation network, and after improved multi-scale feature processing, obtains the instance mask of each Antarctic krill target, and calculates the corresponding pixel area based on the instance mask.

[0011] Specifically, it includes the following sub-steps.

[0012] Step 2.1: Insert small features into the shallow feature output of the backbone network to maintain the pathway without loss. In the original YOLOv12 instance segmentation network, the backbone network performs layer-by-layer convolution and downsampling on the input image to obtain feature maps of different scales. The shallow feature maps contain more edge, texture, and local contour information, playing a crucial role in identifying the head, tail, and small individuals of Antarctic krill. However, Antarctic krill are small, and small samples account for a low proportion of the overall image. If directly subjected to multiple downsampling and channel compression, the tiny features such as the head and tail are easily weakened, leading to missed detections or incomplete mask boundaries in subsequent segmentation prediction heads. To address these issues, this invention inserts a lossless micro-feature preservation path after the shallow feature output of the backbone network P3 of the original YOLOv12 instance segmentation network and before entering the feature fusion network. This lossless micro-feature preservation path includes a global average pooling unit, an adaptive one-dimensional convolution unit, a sigmoid activation unit, and a feature recalibration unit. The shallow feature map output from the backbone network P3 layer is denoted as... Global average pooling unit pairs Compression is performed in the spatial dimension to obtain channel description vectors. Then, an adaptive one-dimensional convolutional unit automatically determines the kernel size based on the number of channels in the channel description vectors and models the local relationships between adjacent channels to obtain channel interaction features. Next, a sigmoid activation unit normalizes the channel interaction features into channel weights. Finally, a feature recalibration unit modifies the channel weights with the original shallow feature map. Element-wise multiplication yields the enhanced shallow feature map. .

[0013] The calculation process is expressed as follows: In the formula, This is a shallow feature map output by the P3 layer of the backbone network; This is a global average pooling operation; This is a one-dimensional convolution operation with a kernel size of k. Use the Sigmoid activation function; This indicates element-wise multiplication; This is a shallow feature map after lossless enhancement of pathways by minimizing features. After this processing, the shallow feature responses representing the head, tail, and edges of small individuals of Antarctic krill are enhanced, reducing information loss of small samples in subsequent feature fusion processes.

[0014] Step 2.2: Insert a deep hybrid domain noise reduction filter at the deep feature output of the backbone network. In the sorting scenario of Antarctic krill production lines, the background of images often contains bubbles, debris, reflective areas, and other impurities. These interfering targets may resemble small-sized Antarctic krill in shape and brightness, making them easy for instance segmentation networks to misidentify as Antarctic krill targets and further participate in subsequent area statistics, causing the mean of the current batch of samples and the dynamic grading threshold to shift.

[0015] To address the aforementioned issues, this invention inserts a deep hybrid domain noise reduction filter after the deep feature output of the P5 backbone network and before the SPPF module in the original YOLOv12 instance segmentation network. This deep hybrid domain noise reduction filter includes a channel attention submodule and a spatial attention submodule. The channel attention submodule suppresses invalid channels that strongly respond to background highlights and debris textures; the spatial attention submodule reduces the feature responses of bubbles, debris, and background regions in their spatial locations.

[0016] The specific processing procedure is as follows: First, the deep feature map output by the P5 layer of the backbone network is denoted as... The channel attention submodules respectively... Global average pooling and global max pooling are performed to obtain two sets of channel description vectors. These two sets of channel description vectors are then input into a shared multilayer perceptron to generate channel attention weights. And the channel attention weight is compared with Element-wise multiplication yields the channel-denoised feature map. Then, the spatial attention submodule performs average pooling and max pooling on the channel-dimension feature map, respectively, and concatenates the two results before inputting them into the input module. Convolutional layers generate spatial attention suppression weights. Finally, the spatial suppression attention weights are multiplied element-wise with the channel-suppressed feature map to obtain the denoised deep feature map. .

[0017] The calculation process is expressed as follows: In the formula, This is the deep feature map output by the P5 layer of the backbone network; This is the channel attention weight generation function; The feature map is after channel attention noise reduction processing; Let this be the spatial attention suppression weight generation function; This indicates element-wise multiplication; This is the deep feature map after processing by the deep hybrid domain noise suppression filter.

[0018] After this processing, the feature responses of bubbles, debris, and reflective areas in the background of the production line are suppressed, and the deep features entering the subsequent feature fusion network are more concentrated in the real Antarctic krill target area.

[0019] Step 2.3: Fuse multi-scale features and obtain a fused feature map. After completing steps 2.1 and 2.2, the backbone network outputs the enhanced shallow feature maps. Mid-layer feature map and noise-suppressed deep feature maps The three types of feature maps mentioned above are input into the feature fusion network of the YOLOv12 instance segmentation network. The feature fusion network fuses feature information at different scales through upsampling, downsampling, feature concatenation, and C2f feature aggregation modules. Specifically, shallow features provide information on the edges, contours, and local textures of Antarctic krill; mid-level features provide information on the shrimp's body structure and local morphology; and deep features provide target semantic information and background suppression information. After processing by the feature fusion network, a fused feature map is obtained for segmentation prediction. .

[0020] The output of this step As input for the subsequent cross-space topology repair module.

[0021] Step 2.4: Insert a cross-spatial topology repair module after the feature fusion network. Antarctic krill bodies are semi-transparent, which can easily lead to areas with low grayscale, discontinuous textures, or weak responses in the head-tail connection region in images. If the fused feature map is directly input into the segmentation prediction head, the segmentation results may show mask breaks in the head, body, and tail of the same Antarctic krill, causing a real individual to be segmented into multiple disconnected regions, resulting in underestimated pixel area and misjudgment of size.

[0022] To address the aforementioned issues, this invention inserts a cross-spatial topology repair module after the feature aggregation module of the C2f feature fusion network in the original YOLOv12 instance segmentation network and before the segmentation prediction head. The cross-spatial topology repair module includes a location encoding unit, a feature mapping unit, a spatial correlation calculation unit, and a residual aggregation unit.

[0023] The specific processing procedure is as follows: First, the fused feature map output in step 2.3 is denoted as... The global position coding matrix is ​​obtained through position coding units. and the global position encoding matrix Adding the fused feature map to the fused feature map yields an enhanced feature map containing spatial location information. : Then, the feature mapping unit uses Convolution enhances the feature map respectively Perform a linear mapping to obtain the query matrix. Key matrix Sum matrix The spatial correlation calculation unit calculates based on the query matrix. AND key matrix The similarity is used to calculate the spatial association weights between different spatial locations, and then normalized using the Softmax function. Subsequently, the residual aggregation unit uses the normalized spatial association weights to adjust the value matrix. Perform weighted aggregation and merge with the original feature map Perform residual connections to obtain the fused feature map after topology repair. : In the formula, The fused feature map output by the feature fusion network; This is the global location encoding matrix; An enhanced feature map after incorporating spatial location information; , , These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; This is the normalization function; This is a weighted aggregation operation; This is the output feature map after cross-spatial topology repair.

[0024] After this treatment, the long-distance spatial correlation between the head, translucent body and tail of Antarctic krill is enhanced, and the characteristic continuity of the transparent body region is improved, thereby reducing instance masking breakage.

[0025] Step 2.5: Output Antarctic krill instance masks using the segmentation prediction head and calculate pixel areas. The topology repair feature map obtained in step 2.4 Input the segmentation prediction head of the YOLOv12 instance segmentation network. The segmentation prediction head outputs the target category, bounding box, and instance mask for each candidate target. Antarctic krill targets are selected based on the target category, and the instance mask of each Antarctic krill target is binarized. The regions with a value of 1 in the binary mask are set as the target mask regions.

[0026] For the For each Antarctic krill target, the number of pixels within the target mask area is counted to obtain the pixel area of ​​the Antarctic krill target. : In the formula, For the first The pixel area of ​​an Antarctic krill target; For the first An Antarctic krill target at pixel coordinates The mask value at a given location is 1 when the pixel belongs to the Antarctic krill target mask area, and 0 otherwise. and These represent the width and height of the input topology repair feature map, respectively.

[0027] Therefore, the system transforms the segmentation results of Antarctic krill instances in the front-end image into two-dimensional pixel area data that can be used for size classification.

[0028] Step 2.6: Construct a length-surface morphological mapping model from physical volume length to pixel area. Antarctic krill sizes are typically classified based on their physical body length. However, krill produced on industrial production lines are often in a naturally curved or sideways position. Classifying krill directly based on the length of straight lines or skeletons in images is easily affected by changes in posture. To reduce the interference of curled posture on the classification results, this invention converts the one-dimensional physical body length classification standard into a two-dimensional pixel area classification standard.

[0029] Specifically, the system reads the "body length - pixel area" paired dataset obtained from the offline calibration in step one, and then uses the standard physical body length of the sample. Pixel area in natural state The correspondence between them is used to establish a long-face morphological mapping model: In the formula, Physical body length The corresponding standard pixel area; This is the standard physical body length of Antarctic krill; The pixel physical equivalent under the current image acquisition conditions; and The regression coefficient between volume length and pixel area; is the fitting constant.

[0030] Substitute the preset large-scale boundary length and small-scale boundary length into the length-surface morphological mapping model to obtain the corresponding large-scale reference area and small-scale reference area.

[0031] When the dividing line length for large sizes is 45 mm and the dividing line length for small sizes is 35 mm in the grading standard, the system will respectively... and Substituting into the above mapping model, we obtain the corresponding large-scale reference area. and small-size benchmark area : in, The reference pixel area corresponding to the large-size boundary. This represents the baseline pixel area corresponding to the smaller size boundary. Through the above conversion, the system transforms the physical volume length grading standard into a pixel area grading standard that can be directly calculated and compared in machine vision.

[0032] Step 2.7: Construct a statistical dynamic mapping model and output the specification grading results. After obtaining the pixel area of ​​each Antarctic krill target in the current image or the current time window, a statistical dynamic mapping model is further established to adapt to the changes in the distribution of raw material specifications and image acquisition conditions in different batches.

[0033] First, a sampling time window is set, and the actual Antarctic krill target areas identified and retained by the instance segmentation network within this sampling time window are statistically analyzed to calculate the average area of ​​the current batch. : In the formula, This represents the average area of ​​the Antarctic krill target within the current sampling time window. The target number of Antarctic krill within this time window; For the first The pixel area of ​​an Antarctic krill target.

[0034] Then, read the mean area of ​​the calibrated sample from step one. Based on the relationship between the current batch area mean and the calibration sample area mean, a batch correction factor is calculated. : In the formula, This is a batch correction factor; To determine the mean area of ​​the calibration sample; This is a dynamic correction factor, with a value range of [value range missing]. .

[0035] Subsequently, the large-size and small-size reference areas obtained in step 2.6 are corrected according to the batch correction factor to obtain the large-size dynamic threshold corresponding to the current batch. and small-scale dynamic threshold : In the formula, The current batch's large-size dynamic grading threshold; This is the dynamic grading threshold for small specifications in the current batch.

[0036] Finally, for the first sampling time within the current sampling time window For each Antarctic krill target that enters the grading process, the system calls the pixel area obtained from the instance mask statistics in step 2.5. and will Compared with the current batch of large-scale dynamic grading threshold and small-size dynamic grading threshold Compare and output specification classification results : In the formula, For the first Size determination results for one Antarctic krill target; This refers to large-sized Antarctic krill; This indicates medium-sized Antarctic krill; This refers to small-sized Antarctic krill.

[0037] Based on the size determination results, the system visually marks Antarctic krill targets of different sizes and transmits the grading results to the subsequent mechanical sorting mechanism. Specifically, the system can mark large-sized targets in red, medium-sized targets in yellow, and small-sized targets in green, thereby achieving a unified output of Antarctic krill visual recognition results, pixel area calculation results, and size grading results.

[0038] Step 3: Construct an adaptive physical sorting execution mechanism with a "vision-mechanical" closed loop; To address the fundamental flaw of existing mechanical sorting devices with fixed sieve holes or spacing, which cannot respond to fluctuations in raw material batches, this invention transforms the specification grading results (flexible control factors) output by the Antarctic krill sorting model in step two into physical opening instructions for the underlying mechanical structure, and finally outputs the physical sorting results, achieving truly flexible sorting. Specifically, as follows: (1) Cross-domain signal conversion and servo drive: The main control system extracts the dynamic grading threshold (i.e., the large-size dynamic threshold that dynamically fluctuates with the current raw material batch average) output in step two. and small-scale dynamic threshold ), and through the preset mechanical mapping relationship, it is converted into the physical opening target value required by the back-end sorting channel. It should be noted that the physical opening is not a fixed value set by humans in advance, but is inherited from the optimal boundary value determined by the real-time statistical decision of the vision network. Subsequently, the system adopts an industrial control module and uses a mature servo position control algorithm to convert the physical opening target value into motor drive pulses or displacement commands to control the movement of the back-end drive mechanism. (2) Flexible mapping of physical channel and sorting output: After receiving the command in step (1), the drive mechanism drives the transmission mechanism of the Antarctic krill intelligent sorting device to move, so that the back end of the sorting tube is displaced laterally, thereby continuously changing the spacing of the longitudinally gradually opening sorting channel to match the physical opening target value output by the vision network. Subsequently, the material passes through the sorting channel that has completed adaptive adjustment to complete the final physical screening and result output. (3) Innovative effect: This mechanism completely breaks the limitation of the traditional sorting machine relying on manual shutdown and blind adjustment. The system captures the characteristics of the raw materials in real time through the front-end vision network, automatically calculates the optimal dividing point, and converts it into a signal to drive the opening and closing of the back-end sorting channel in real time. It truly realizes the integrated intelligent perception of "eyes, brain, and hands", enabling purely mechanical devices to have flexible execution capabilities that adapt to biological morphological differences.

[0039] The beneficial effects of this invention are: (1) Constructing a "vision-machine" closed-loop linkage architecture to achieve flexible adaptive adjustment of the sorting channel: This invention breaks through the defect of fixed spacing in traditional mechanical sorting and innovatively converts the dynamic grading threshold calculated in real time by the vision network directly into the physical drive signal of the servo motor. The system automatically and steplessly adjusts the opening of the rear end of the involute sorting tube, so that the mechanical channel "changes with the shrimp", completely eliminating the tediousness of manual stoppage and blind adjustment, and realizing truly fully automatic non-destructive sorting.

[0040] (2) By integrating allometric growth patterns and dynamic statistical decision-making, this invention overcomes grading errors caused by the translucency and curling deformation of shrimp bodies. The invention repairs the fractures in the translucent body at the visual level and filters out industrial impurities. Simultaneously, it abandons rigid one-dimensional length measurement and fixed thresholds, utilizing "length-surface" morphological mapping to completely eliminate measurement errors caused by the natural curling of the shrimp body. The system combines the dynamic average of the current batch of raw materials with real-time floating grading criteria, effectively resisting interference from fluctuations in raw material specifications across different fishing seasons and improving grading accuracy. Attached Figure Description

[0041] Figure 1 This is a flowchart of the block module of the present invention.

[0042] Figure 2 The results are the classification results before and after model improvement; where (a) is the segmentation result before model improvement, and (b) is the result after model improvement and classification.

[0043] Figure 3 The improved network structure diagram of the YOLOv12 instance segmentation network.

[0044] Figure 4 This is a nonlinear mapping relationship between the body length of Antarctic krill and the pixel area, and a grading benchmark curve diagram, as shown in this embodiment of the invention.

[0045] Figure 5 Overall structural diagram.

[0046] Figure 6 A schematic diagram of the rear-end adjustment mechanism and a partial structure of the sorting module.

[0047] Figure 7 A partial structural diagram of the drive mechanism.

[0048] Figure 8 Side view of the overall structure.

[0049] Figure 9 Rear-end adjustable distance mechanism.

[0050] Explanation of reference numerals in the attached drawings: 1. Frame; 2. Sorting module; 3. Receiving trough assembly; 4. Side baffle; 5. Feed hopper; 6. Rear guide mechanism; 21. Sorting tube; 22. Front fixing plate; 23. Rear connecting plate; 25. Sorting channel; 31. First receiving trough; 32. Second receiving trough; 33. Third receiving trough; 34. Fourth receiving trough; 41. Left side baffle; 42. Right side baffle; 51. Feed hopper body; 61. Rear guide beam; 71. Slider body; 72. Sorting tube mounting hole; 73. Drive pin; 81. Inclined chute drive plate; 82. Inclined chute; 91. Drive screw; 92. Coupling; 93. Motor mounting base; 94. Drive motor. Detailed Implementation

[0051] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0052] Step 1: Constructing the image dataset of Antarctic krill includes the following: Step 1.1, Morphological Physical Calibration and Image Acquisition: Antarctic krill samples of large, medium, and small sizes were randomly selected beforehand, and each sample was assigned a unique number. First, the krill samples were flattened without damaging the krill body, and their standard physical body length L was manually measured. The standard physical body length L is the distance from the base of the compound eye to the end of the telson. Then, krill samples with the same number were placed in the image acquisition area, and their visual images were acquired in a naturally bent or naturally positioned state. The corresponding two-dimensional pixel area S was extracted based on the target region of the krill in the visual image. Thus, a one-to-one correspondence was established between the standard physical body length L of the same sample and the pixel area S in the visual image under natural conditions, forming a "body length - pixel area" paired dataset, providing calibration data for subsequent conversion from the physical body length grading standard to the pixel area grading standard.

[0053] Step 1.2, Refined Connected Component Annotation: The outline of the Antarctic krill is annotated at the pixel level using polygonal annotation to obtain an Antarctic krill mask. In this embodiment, LabelMe software is used for polygonal pixel-level annotation. Considering the characteristic that the translucent body of the Antarctic krill is easily visually broken, the annotators, based on morphological experience, artificially closed and connected the transparent areas between the "krill head" and "krill tail" to ensure that the ground truth mask (set as krill category) covers the complete "head-body-tail" individual.

[0054] Step 1.3, Sample Balancing and Enhancement: Before training the improved YOLOv12 instance segmentation network model, sample balancing and enhancement were performed on large, medium, and small Antarctic krill samples. First, basic data enhancement was performed simultaneously on the original labeled images and their corresponding masks, including horizontal flipping, random rotation, image sharpening, brightness adjustment, contrast adjustment, salt-and-pepper noise injection, and red channel enhancement. Horizontal flipping and random rotation were used to simulate different placement postures of Antarctic krill during transport; brightness and contrast adjustment were used to simulate fluctuations in ambient light; salt-and-pepper noise was used to simulate water bubbles, debris, and particle interference in the conveyor belt image; and red channel enhancement was used to enhance the color difference between the Antarctic krill target and the background.

[0055] Subsequently, a combination of Mixup and Mosaic enhancement strategies was employed during the training phase of the improved YOLOv12 instance segmentation network model. Mixup was used to proportionally blend different sample images and their labels to increase the continuity of the sample distribution; Mosaic was used to randomly crop, scale, and stitch multiple sample images into the same training image to increase the frequency of small-sized Antarctic krill in a single training image and enrich the target combination forms under different background conditions. Through the combination of these basic and combined enhancements, the number of training samples was expanded, reducing the risk of overfitting the improved YOLOv12 instance segmentation network model to a single size, pose, or background condition, and improving the adaptability of the subsequent improved YOLOv12 instance segmentation network model to Antarctic krill of different sizes.

[0056] Step 2: Specific Implementation of Improving the YOLOv12 Instance Segmentation Network and Hierarchical Decision Process In this embodiment, the YOLOv12 instance segmentation network is trained and applied using the Antarctic krill image dataset constructed in step one. The YOLOv12 instance segmentation network takes Antarctic krill images as input and outputs the Antarctic krill target category, bounding box, and instance mask. To adapt to the characteristics of Antarctic krill—semi-transparent, small in size, with curved postures, and subject to significant background interference from the flow line—this embodiment makes the following improvements to the backbone network, feature fusion network, and post-processing hierarchical module of the original YOLOv12 instance segmentation network.

[0057] Step 2.1: Specific Implementation of Non-destructive Preservation of Pathways for Minor Features In this embodiment, the backbone network of the original YOLOv12 instance segmentation network outputs feature maps at three scales: P3, P4, and P5. Among them, the P3 feature map retains more edge, texture, and local contour information, and is mainly used to express the detailed features of the head, tail, and small individuals of Antarctic krill.

[0058] In this embodiment, a lossless preservation path for minor features is added after the shallow feature output of the backbone network P3 and before the feature fusion network. That is, the P3 feature map in the original network is directly fed into the feature fusion network; in this embodiment, the P3 feature map first undergoes lossless preservation of minor features before being fed into the feature fusion network. Specifically, the P3 feature map is input into a global average pooling unit to obtain global description information for each channel; then, this channel description information is input into an adaptive one-dimensional convolutional unit to establish local connections between adjacent channels; subsequently, a sigmoid activation unit generates channel weights; finally, a feature recalibration unit applies these channel weights to the original P3 feature map to obtain an enhanced P3 feature map. This enhanced P3 feature map is then used as a shallow input to the subsequent feature fusion network.

[0059] With the above settings, the network enhances the shallow responses of shrimp heads, tails, and small individuals before entering the feature fusion stage, reducing the risk of these responses being weakened in the subsequent feature fusion process.

[0060] Step 2.2, Specific Implementation of the Deep Hybrid Domain Noise Suppression Filter In this embodiment, the P5 feature map represents the deep semantic features output by the backbone network. Since images of Antarctic krill processing lines are prone to contain bubbles, debris, reflections, and background impurities, these disturbances may create strong responses in the deep features, affecting subsequent instance segmentation results. In this embodiment, a deep hybrid domain noise reduction filter is connected after the P5 deep feature output of the backbone network and before the SPPF module. That is, the P5 feature map in the original network directly enters the SPPF module; in this embodiment, the P5 feature map is first processed by the deep hybrid domain noise reduction filter before entering the SPPF module and subsequent feature fusion networks.

[0061] In practice, the P5 feature map is first input into the channel attention submodule. This submodule performs global average pooling and global max pooling on the P5 feature map to obtain two sets of channel description vectors, and generates channel attention weights through a shared multilayer perceptron. These channel attention weights are then compared with... Element-wise multiplication yields a channel-denoised feature map, used to weaken channels that strongly respond to interference such as background highlights and debris textures. This channel-denoised feature map is then input into the spatial attention submodule. The spatial attention submodule generates spatial attention suppression weights in the spatial dimension to reduce the response at corresponding locations of bubbles, debris, and reflective areas. Finally, the spatial suppression attention weights are multiplied element-wise with the channel-denoised feature map to obtain a denoised deep feature map.

[0062] The denoised P5 feature map is used as deep input to the subsequent feature fusion network, enabling the network to focus more on the real Antarctic krill target area at the semantic level.

[0063] Step 2.3, Specific Implementation of Multi-Scale Feature Fusion After completing steps 2.1 and 2.2, this embodiment inputs the enhanced P3 feature map, the P4 feature map output by the backbone network, and the denoised P5 feature map into the feature fusion network of the YOLOv12 instance segmentation network. In the feature fusion network, the system performs upsampling, downsampling, feature concatenation, and C2f feature aggregation on the three types of feature maps according to the multi-scale fusion path of the YOLOv12 instance segmentation network. Specifically, the enhanced P3 feature map is used to supplement edge and local detail information of Antarctic krill, the P4 feature map is used to provide shrimp body structure and local morphological information, and the denoised P5 feature map is used to provide target semantic information and background suppression information.

[0064] After the above fusion process, the feature fusion network outputs a fused feature map. This fused feature map is not directly input into the segmentation prediction head, but is instead fed into the cross-spatial topology repair module for further processing.

[0065] Step 2.4, Specific Implementation of the Cross-Space Topology Repair Module In this embodiment, to address the issues of weak feature response and easy mask breakage in the translucent body region of Antarctic krill, a cross-spatial topology repair module is integrated after the C2f feature aggregation module and before the segmentation prediction head in the feature fusion network. That is, in the original YOLOv12 instance segmentation network, the fused feature map output by the C2f feature aggregation module directly enters the segmentation prediction head; in this embodiment, the fused feature map first passes through the cross-spatial topology repair module before entering the segmentation prediction head.

[0066] The cross-spatial topology repair module includes a location encoding unit, a feature mapping unit, a spatial correlation calculation unit, and a residual aggregation unit. In specific implementation, the location encoding unit first adds global location encoding information to the fused feature map, enabling the fused feature map to carry pixel spatial location information. Subsequently, the feature mapping unit uses three 1×1 convolutions to map the fused features, obtaining query features, key features, and value features. The spatial correlation calculation unit calculates the association weights between different spatial locations based on the similarity between the query features and key features, and normalizes them using the Softmax function. The residual aggregation unit then uses these association weights to perform weighted aggregation of the value features and performs a residual connection with the original fused features to obtain the topology-repaired fused feature map.

[0067] Through the above processing, the spatial correlation between the head, translucent body, and tail of the Antarctic krill is enhanced, and weak response regions can obtain feature supplementation from adjacent high response regions. The topology-repaired fused feature map is then input into the segmentation prediction head to output an Antarctic krill instance mask with better connectivity.

[0068] Step 2.5: Specific Implementation of Instance Mask Output and Area Extraction In this embodiment, the segmentation prediction head receives the topology repair fusion feature map output in step 2.4 and performs category prediction, bounding box prediction, and instance mask prediction for each candidate target. The system first filters out Antarctic krill targets based on the category prediction results, removing non-target categories; then, it retains valid targets based on a confidence threshold; subsequently, it binarizes the instance mask of the retained targets, setting pixels belonging to the Antarctic krill target region to 1 and background pixels to 0. The system counts the number of pixels with a value of 1 in the binary mask to obtain the pixel area of ​​the Antarctic krill target.

[0069] Through this processing, the image results output by the segmentation prediction head are converted into area data that can be directly used by subsequent grading modules. Each valid Antarctic krill target corresponds to an instance mask and a pixel area value.

[0070] Step 2.6, Specific Implementation of the Long-Face Morphological Mapping Model In this embodiment, the system calls the "body length-pixel area" paired dataset obtained in step one. In this dataset, each Antarctic krill sample corresponds to a manually measured standard physical body length and a pixel area extracted from a natural-state image. Before formal grading, the system establishes a mapping relationship between body length and pixel area based on the above paired data. Then, the preset large-size and small-size boundary body lengths are substituted into this mapping relationship to obtain the corresponding large-size and small-size reference areas. The above reference areas no longer depend on the straight line length or skeleton length of the shrimp body in the image, but are used as area boundaries in subsequent grading judgments.

[0071] Through this processing, the system converts the physical body length grading standard into a pixel area grading standard, thereby reducing the impact of Antarctic krill's natural bending, lateral bending, or inconsistent posture on size determination.

[0072] Step 2.7, Specific Implementation of Statistical Dynamic Mapping and Specification Determination In this embodiment, the system uses a set time window as a unit to statistically analyze the area of ​​valid Antarctic krill targets identified within that time window, obtaining the area distribution information for the current batch. Based on the number of valid targets and the pixel area of ​​each target within that time window, the system calculates the average area of ​​the current batch.

[0073] Subsequently, the system reads the mean area of ​​the calibrated samples from step one, and obtains a batch correction factor based on the difference between the mean area of ​​the current batch and the mean area of ​​the calibrated samples. The system then uses this batch correction factor to correct the large-size and small-size reference areas obtained in step 2.6, forming the large-size dynamic threshold and small-size dynamic threshold for the current batch.

[0074] When a new Antarctic krill target enters the classification and judgment stage, the system compares the pixel area of ​​the target with the current batch dynamic threshold: if the target area is greater than or equal to the large size dynamic threshold, it is judged as a large size Antarctic krill; if the target area is less than the small size dynamic threshold, it is judged as a small size Antarctic krill; if the target area is between the small size dynamic threshold and the large size dynamic threshold, it is judged as a medium size Antarctic krill.

[0075] After the determination is completed, the system sends the specification results to the subsequent mechanical sorting actuator. Simultaneously, the system visually marks targets of different sizes on the monitoring interface: large targets are marked in red, medium targets in yellow, and small targets in green. This enables continuous processing of Antarctic krill image input, instance segmentation, area extraction, dynamic grading, and mechanical sorting signal output.

[0076] Step 3: The specific process of implementing "vision-mechanical" closed-loop adaptive physical sorting In this step, the system receives the Antarctic krill size grading results and flexible control factors output from step two, and converts them into execution control signals for the sorting device, thereby enabling adaptive adjustment of the sorting channel opening and continuous output of Antarctic krill of different sizes.

[0077] like Figure 9 As shown, the sorting device includes a frame 1, a sorting module 2 mounted on the frame 1, a receiving trough assembly 3 mounted below the sorting module 2, a feeding hopper 5 mounted at the front end of the sorting module 2, a rear guide mechanism 6 mounted at the rear end of the sorting module 2, a chute drive mechanism 8, and a drive mechanism 9. The drive mechanism 9 is located at the rear end of the slant drive mechanism 8, and includes a motor mounting base 93 mounted on the frame 1, on which a controlled drive motor 94 is mounted. When the drive motor 94 is working, its output shaft drives the drive screw 91 to rotate with high precision via the coupling 92.

[0078] like Figure 5 , Figure 6 , Figure 7 and Figure 8As shown, the sorting device of the present invention is supported by a frame 1. A sorting module 2 is mounted on the frame 1, which consists of multiple sorting tubes 21 arranged along the length direction. A front-end fixing plate 22 is provided at the front end of the sorting module 2, and the front end of each sorting tube 21 is rotatably connected to or hinged to the front-end fixing plate 22, so that the front end position of the sorting tube 21 is relatively fixed. A rear-end connecting plate 23 and a rear-end guiding mechanism 6 are provided at the rear end of the sorting module 2. The rear-end guiding mechanism 6 includes a transversely arranged rear-end guiding beam 61, which is fixed to the rear end of the rear-end connecting plate 23 and parallel to the rear-end connecting plate 23. The rear end of each sorting tube 21 passes through the rear-end connecting plate 23 and the rear-end guiding beam 61 in sequence and is supported in the sorting tube mounting hole 72 on the corresponding slider body 71, and can slide in the sorting tube mounting hole 72.

[0079] Each slider body 71 is slidably mounted on the rear guide beam 61, and each slider body 71 has a downwardly protruding drive pin 73 at its bottom. The inclined groove drive mechanism 8 includes an inclined groove drive plate 81, which is fixed to a screw nut on a drive screw 91. When the drive screw 91 rotates, the screw nut moves linearly along the drive screw 91, thereby driving the inclined groove drive plate 81 to move linearly. The inclined groove drive plate 81 has multiple inclined grooves 82 distributed at an angle, and each drive pin 73 extends into the corresponding inclined groove 82. When the inclined groove drive plate 81 moves, the drive pin 73 undergoes lateral displacement under the constraint and guidance of the inclined groove 82, thereby forcibly driving the slider body 71 to slide laterally along the rear guide beam 61. The rear guide beam 61 contains a straight slot to provide clearance space for the sorting tube 21 during movement. Because the front end of the sorting tube 21 is constrained by the front fixing plate 22, its rear end expands or contracts relative to the slider body 71, thereby dynamically changing the spacing of the gradually opening sorting channel 25 formed between adjacent sorting tubes 21.

[0080] After the system completes the channel opening adjustment, Antarctic krill enters through the feed hopper 5 located above the front end of the sorting module 2. The feed hopper 5 includes a feed hopper body 51 for containing materials. The Antarctic krill advances along the length of the conveying path formed by the sorting pipes 21. During this process, side baffles 4 are set on both sides of the sorting module 2, including a left baffle 41 and a right baffle 42, to prevent the Antarctic krill from falling from the sides. As the Antarctic krill of different sizes advances, they gradually fall into the receiving trough assembly 3 located directly below at the corresponding size position as the width of the sorting channel 25 gradually opens. The receiving trough assembly 3 is sequentially divided into a first receiving trough 31, a second receiving trough 32, a third receiving trough 33, and a fourth receiving trough 34 along the conveying direction, thereby realizing continuous separation and precise output of large, medium, and small Antarctic krill samples.

[0081] It should be understood that any content not described in detail in this specification belongs to the prior art. Those skilled in the art should understand that they can modify or substitute the technical solutions of this invention; however, any modifications, substitutions, or improvements made within the spirit and principles of this invention should be within the protection scope of this invention.

Claims

1. A flexible and switchable intelligent sorting method for Antarctic krill, characterized in that, The steps are as follows: Step 1: Acquisition of Antarctic krill image data and construction of augmented dataset; Step one specifically includes: Step 1.1, Morphological Physical Calibration and Image Acquisition: Antarctic krill samples of large, medium, and small sizes are randomly selected beforehand, and each sample is assigned a unique number. First, the krill samples are flattened without damaging the krill body, and their standard physical body length L is measured manually. The standard physical body length L is the distance from the base of the compound eye to the end of the telson. Then, krill samples with the same number are placed in the image acquisition area, and their visual images are acquired in a naturally bent or naturally positioned state. The corresponding two-dimensional pixel area S is extracted based on the target area of ​​the krill in the visual image. Thus, a one-to-one correspondence is established between the standard physical body length L of the same sample and the pixel area S in the visual image under natural conditions, forming a "body length - pixel area" paired dataset, providing calibration data for subsequent conversion from the physical body length grading standard to the pixel area grading standard. Step 1.2, Refined Annotation of Connected Components: The outline of the Antarctic krill is annotated at the pixel level using the polygon annotation method to obtain the Antarctic krill mask; Enhanced annotation is performed on the semi-transparent body connection points of the Antarctic krill to ensure that the mask completely covers the entire "head-body-tail" view; Step 1.3, Sample Balancing and Augmentation: Data augmentation and quantity balancing were performed on Antarctic krill samples of large, medium and small sizes; a combination of flipping, rotating, sharpening, brightness and contrast adjustment, salt and pepper noise injection, and red channel enhancement techniques were used to expand the dataset size. Step 2: Construct an Antarctic krill sorting model based on an improved YOLOv12 instance segmentation network and adaptive statistical decision-making; An improved YOLOv12 instance segmentation network is used to segment Antarctic krill images acquired and labeled in step one, extract pixel areas, and determine size levels. The YOLOv12 instance segmentation network includes a backbone network, a feature fusion network, and a segmentation prediction head. The backbone network is used to extract multi-scale features from the input image and output shallow, medium, and deep feature maps. The feature fusion network is used to upsample, downsample, and fuse channels for features at different scales. The segmentation prediction head is used to output the category information, bounding box information, and instance mask information of the Antarctic krill target. Step 3: Construct an adaptive physical sorting execution mechanism with a "vision-mechanical" closed loop; The specification and grading results output by the Antarctic krill sorting model in step two are converted into physical opening instructions for the underlying mechanical structure, and finally the physical sorting results are output to achieve true flexible sorting. Step three is as follows: (1) Cross-domain signal conversion and servo drive: The main control system extracts the dynamic grading threshold output in step two and converts it into the physical opening target value required by the back-end sorting channel through mechanical mapping relationship; then, the industrial control module is used to convert the physical opening target value into motor drive pulse or displacement command using servo position control algorithm to control the movement of the back-end drive mechanism. (2) Flexible mapping of physical channels and sorting output: After receiving the instruction in step (1), the drive mechanism drives the transmission mechanism of the Antarctic krill intelligent sorting device to move, causing the rear end of the sorting tube to move laterally, thereby continuously changing the spacing of the longitudinally gradually opening sorting channel so that it matches the physical opening target value output by the vision network; subsequently, the material passes through the sorting channel that has completed adaptive adjustment to complete the final physical screening and result output.

2. The flexible and switchable intelligent sorting method for Antarctic krill according to claim 1, characterized in that, Step two specifically includes: Step 2.1: Insert small features into the shallow feature output of the backbone network to maintain the pathway without loss. After the shallow feature output of the backbone P3 layer of the original YOLOv12 instance segmentation network and before entering the feature fusion network, a small feature lossless preservation path is inserted. This small feature lossless preservation path includes a global average pooling unit, an adaptive one-dimensional convolutional unit, a sigmoid activation unit, and a feature recalibration unit. The shallow feature map output by the backbone P3 layer is denoted as... Global average pooling unit pairs Compression is performed in the spatial dimension to obtain channel description vectors. Then, an adaptive one-dimensional convolutional unit automatically determines the kernel size based on the number of channels in the channel description vectors and models the local relationships between adjacent channels to obtain channel interaction features. Next, a sigmoid activation unit normalizes the channel interaction features into channel weights. Finally, a feature recalibration unit modifies the channel weights with the original shallow feature map. Element-wise multiplication yields the enhanced shallow feature map. ; The calculation process is expressed as follows: In the formula, This is a shallow feature map output by the P3 layer of the backbone network; This is a global average pooling operation; This is a one-dimensional convolution operation with a kernel size of k. Use the Sigmoid activation function; This indicates element-wise multiplication; This is a shallow feature map after pathway enhancement without loss of minor features; Step 2.2: Insert a deep hybrid domain noise reduction filter at the deep feature output of the backbone network. A deep hybrid domain noise reduction filter is inserted after the deep feature output of the P5 backbone network of the original YOLOv12 instance segmentation network and before the SPPF module. The deep hybrid domain noise reduction filter includes a channel attention submodule and a spatial attention submodule. The channel attention submodule is used to suppress invalid channels that respond strongly to background highlights and debris textures. The spatial attention submodule is used to reduce the feature responses of bubbles, debris, and background regions in spatial location. The specific processing procedure is as follows: First, the deep feature map output by the P5 layer of the backbone network is denoted as... The channel attention submodule respectively... Global average pooling and global max pooling are performed to obtain two sets of channel description vectors. These two sets of channel description vectors are then input into a shared multilayer perceptron to generate channel attention weights. And the channel attention weight is compared with Element-wise multiplication yields the channel-denoised feature map; then, the spatial attention submodule performs average pooling and max pooling on the channel-dimensional feature map, and concatenates the two before inputting them. Convolutional layers generate spatial attention suppression weights. Finally, the spatial suppression attention weights are multiplied element-wise with the channel-suppressed feature map to obtain the denoised deep feature map. ; The calculation process is expressed as follows: In the formula, This is the deep feature map output by the P5 layer of the backbone network; This is the channel attention weight generation function; The feature map is after channel attention noise reduction processing; Let this be the spatial attention suppression weight generation function; This indicates element-wise multiplication; This is the deep feature map after processing by the deep hybrid domain noise reduction filter; Step 2.3: Fuse multi-scale features and obtain a fused feature map. After completing steps 2.1 and 2.2, the backbone network outputs the enhanced shallow feature maps. Mid-layer feature map and noise-suppressed deep feature maps The feature fusion network of the YOLOv12 instance segmentation network inputs three types of feature maps together. The feature fusion network fuses feature information at different scales through upsampling, downsampling, feature concatenation, and C2f feature aggregation modules. Shallow features provide edge, contour, and local texture information of Antarctic krill; mid-level features provide shrimp body structure and local morphological information; and deep features provide target semantic information and background suppression information. After processing by the feature fusion network, a fused feature map is obtained for segmentation prediction. ; The output of this step As input for subsequent cross-space topology repair modules; Step 2.4: Insert a cross-spatial topology repair module after the feature fusion network. A cross-spatial topology repair module is inserted after the feature fusion network C2f feature aggregation module of the original YOLOv12 instance segmentation network and before the segmentation prediction head; the cross-spatial topology repair module includes a location encoding unit, a feature mapping unit, a spatial correlation calculation unit, and a residual aggregation unit; The specific processing procedure is as follows: First, the fused feature map output in step 2.3 is denoted as... The global position coding matrix is ​​obtained through position coding units. and the global position encoding matrix Adding the fused feature map to the fused feature map yields an enhanced feature map containing spatial location information. : Then, the feature mapping unit uses Convolution enhances the feature map respectively Perform a linear mapping to obtain the query matrix. Key matrix Sum matrix The spatial correlation calculation unit calculates based on the query matrix. AND key matrix The similarity is used to calculate the spatial association weights between different spatial locations, and then normalized using the Softmax function. Subsequently, the residual aggregation unit uses the normalized spatial association weights to adjust the value matrix. Perform weighted aggregation and merge with the original feature map Perform residual connections to obtain the fused feature map after topology repair. : In the formula, The fused feature map output by the feature fusion network; This is the global location encoding matrix; An enhanced feature map after incorporating spatial location information; , , These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; This is the normalization function; This is a weighted aggregation operation; This is the output feature map after cross-spatial topology repair; Step 2.5: Output Antarctic krill instance masks using the segmentation prediction head and calculate pixel areas. The topology repair feature map obtained in step 2.4 Input the segmentation prediction head of the YOLOv12 instance segmentation network; the segmentation prediction head outputs the target category, bounding box and instance mask for each candidate target; Antarctic krill targets are selected according to the target category, and the instance mask of each Antarctic krill target is binarized, and the region with a value of 1 in the binary mask is set as the target mask region; For the For each Antarctic krill target, the number of pixels within the target mask area is counted to obtain the pixel area of ​​the Antarctic krill target. : In the formula, For the first The pixel area of ​​an Antarctic krill target; For the first An Antarctic krill target at pixel coordinates The mask value at a given location is 1 when the pixel belongs to the Antarctic krill target mask area, and 0 otherwise. and These represent the width and height of the input topology restoration feature map, respectively. Step 2.6: Construct a length-surface morphological mapping model from physical volume length to pixel area. Read the "body length - pixel area" paired dataset obtained from the offline calibration in step one, and then, based on the standard physical body length of the samples... Pixel area in natural state The correspondence between them is used to establish a long-face morphological mapping model: In the formula, Physical body length The corresponding standard pixel area; The standard physical body length of Antarctic krill; The pixel physical equivalent under the current image acquisition conditions; and The regression coefficient between volume length and pixel area; These are the fitting constants; Substitute the preset large-scale boundary length and small-scale boundary length into the length-surface morphological mapping model to obtain the corresponding large-scale reference area and small-scale reference area. When the dividing line length for large sizes is 45 mm and the dividing line length for small sizes is 35 mm in the grading standard, the system will respectively... and Substituting into the above mapping model, we obtain the corresponding large-scale reference area. and small-size benchmark area : in, The reference pixel area corresponding to the large-size boundary. The reference pixel area is the boundary of the small size. Through the above conversion, the system converts the physical volume length classification standard into a pixel area classification standard that can be directly calculated and compared in machine vision. Step 2.7: Construct a statistical dynamic mapping model and output the specification grading results. After obtaining the pixel area of ​​each Antarctic krill target in the current image or the current time window, a statistical dynamic mapping model is further established to adapt to the changes in the distribution of raw material specifications and image acquisition conditions in different batches. First, a sampling time window is set, and the actual Antarctic krill target areas identified and retained by the instance segmentation network within this sampling time window are statistically analyzed to calculate the average area of ​​the current batch. : In the formula, This represents the average area of ​​the Antarctic krill target within the current sampling time window. The target number of Antarctic krill within this time window; For the first The pixel area of ​​an Antarctic krill target; Then, read the mean area of ​​the calibrated sample from step one. Based on the relationship between the current batch area mean and the calibration sample area mean, a batch correction factor is calculated. : In the formula, This is a batch correction factor. To determine the mean area of ​​the calibration sample; This is a dynamic correction factor, with a value range of [value range missing]. ; Subsequently, the large-size and small-size reference areas obtained in step 2.6 are corrected according to the batch correction factor to obtain the large-size dynamic threshold corresponding to the current batch. and small-scale dynamic threshold : In the formula, This is the dynamic grading threshold for the large specifications in the current batch; The dynamic grading threshold for small specifications in the current batch; Finally, for the first sampling time within the current sampling time window For each Antarctic krill target that enters the grading process, the system calls the pixel area obtained from the instance mask statistics in step 2.

5. and will Compared with the current batch of large-scale dynamic grading threshold and small-size dynamic grading threshold Compare and output specification classification results : In the formula, For the first Size determination results for one Antarctic krill target; This refers to large-sized Antarctic krill; This indicates medium-sized Antarctic krill; This refers to small-sized Antarctic krill; Based on the specification determination results, the system visually marks Antarctic krill targets of different specifications and transmits the grading results to the subsequent mechanical sorting execution mechanism.

3. A flexible and switchable intelligent sorting device for Antarctic krill, characterized in that, For implementing a flexible and switchable intelligent sorting method for Antarctic krill as described in any one of claims 1-2, the intelligent sorting device for Antarctic krill includes a frame (1), a sorting module (2) disposed on the frame (1), a receiving trough assembly (3) disposed below the sorting module (2), a feeding hopper (5) disposed at the front end of the sorting module (2), a rear guide mechanism (6) disposed at the rear end of the sorting module (2), an inclined trough drive mechanism (8), and a drive mechanism (9); The drive mechanism (9) is located at the rear end of the inclined slot drive mechanism (8), including a motor mounting base (93) mounted on the frame (1), on which a controlled drive motor (94) is mounted; when the drive motor (94) is working, its output shaft drives the drive screw (91) to rotate with high precision via the coupling (92); The sorting device is supported by a frame (1); a sorting module (2) is provided on the frame (1), which is composed of multiple sorting tubes (21) arranged along the length direction; a front fixing plate (22) is provided at the front end of the sorting module (2), and the front end of each sorting tube (21) is rotatably connected to the front fixing plate (22), so that the front end position of the sorting tube (21) is relatively fixed; a rear connecting plate (23) and a rear guiding mechanism (6) are provided at the rear end of the sorting module (2), and the rear guiding mechanism (6) includes a rear guiding beam (61) arranged laterally, which is fixed to the rear end of the rear connecting plate (23) and parallel to the rear connecting plate (23); the rear end of each sorting tube (21) passes through the rear connecting plate (23) and the rear guiding beam (61) in sequence and is supported in the sorting tube mounting hole (72) on the corresponding slider body (71), and can slide in the sorting tube mounting hole (72); Each slider body (71) is slidably mounted on the rear guide beam (61), and each slider body (71) has a driving pin (73) protruding downward from its bottom; the inclined groove drive mechanism (8) includes an inclined groove drive plate (81), which is fixed to the screw nut on the drive screw (91). When the drive screw (91) rotates, the screw nut moves linearly along the drive screw (91), thereby driving the inclined groove drive plate (81) to move linearly; the inclined groove drive plate (81) is provided with multiple inclined grooves (82) distributed at an incline, and each driving pin (73) extends into the corresponding inclined groove (82). 2) Inside; when the inclined groove drive plate (81) moves, the drive pin (73) generates a lateral displacement under the constraint and guidance of the inclined groove (82), thereby forcibly driving the slider body (71) to slide laterally along the rear guide beam (61); the rear guide beam (61) contains a straight groove, which provides clearance space for the sorting tube (21) to move; since the front end of the sorting tube (21) is constrained by the front end fixing plate (22), its rear end is relatively expanded or contracted under the drive of the slider body (71), thereby dynamically changing the spacing of the gradually opening sorting channel (25) formed between adjacent sorting tubes (21); After the system completes the channel opening adjustment, Antarctic krill enters through the feed hopper (5) located above the front end of the sorting module (2); the feed hopper (5) includes a feed hopper body (51) for containing materials; the Antarctic krill moves along the length direction on the conveying path formed by the sorting pipe (21). During this process, side baffles (4) are set on both sides of the sorting module (2), including a left baffle (41) and a right baffle (42), to prevent the Antarctic krill from falling from the side; as the Antarctic krill of different sizes moves forward, as the width of the sorting channel (25) gradually opens, they fall into the receiving trough assembly (3) set directly below at the corresponding size position; the receiving trough assembly (3) is divided into a first receiving trough (31), a second receiving trough (32), a third receiving trough (33) and a fourth receiving trough (34) in sequence along the conveying direction, thereby realizing the continuous separation and precise output of large, medium and small Antarctic krill samples.

Images