Multi-layer caged waterfowl feeding method, system and device based on behavior feedback

Abstract

The invention discloses a multi-layer cage-rearing waterfowl feeding method, system and device based on behavior feedback, and belongs to the technical field of waterfowl rearing, and the method comprises the steps that a high-definition camera on a movable trough collects waterfowl image data and a two-dimensional code label of a rearing cage in real time; the core processing module accurately positions a cage number by decoding the two-dimensional code, and segments a waterfowl body area image and a mouth area image by adopting a multi-channel fusion method; calculating the ingestion desire index and the body condition score of the waterfowl based on the segmented mouth area and the waterfowl body area, and feeding the waterfowl based on the ingestion desire index and the body condition score; a high-definition camera installed at a trough operation track terminal collects trough images; and the core processing module is used for carrying out pollution image analysis and pollution degree judgment on the movable trough, and automatically generating a skip car cleaning prompt based on the pollution degree judgment. According to the method, reasonable feeding of waterfowls is achieved, feed waste and pollution are reduced, and breeding benefits and the automation level are improved.

Description

Technical Field

[0001] This application relates to the technical field of waterfowl breeding, and particularly to a multi-layer cage feeding method, system and device for waterfowl based on behavior feedback. Background Art

[0002] At present, waterfowl breeding is mainly divided into two modes: cage breeding and non-cage breeding. Compared with the non-cage breeding mode, the waterfowl cage breeding mode helps to better control the feed intake and feeding time of waterfowl, ensure that waterfowl obtain appropriate nutrition, and effectively manage feed costs. In addition, the cage breeding mode is convenient for cleaning the house, can improve the breeding environment, prevent water resource pollution, strengthen disease prevention and control, and improve animal welfare.

[0003] With the rapid development of the waterfowl breeding industry, the traditional waterfowl feeding methods face many technical challenges. At present, large-scale farms generally adopt two methods: manual feeding and pipeline feeding, but both of these methods have obvious defects: manual feeding has a large labor intensity, the feed is easily contaminated and unevenly distributed; pipeline feeding is prone to blockage and feed residue problems. More critically, most of the existing waterfowl feeding equipment directly applies the feeding equipment for meat / egg chickens, and this design does not fully consider the special physiological characteristics of waterfowl, such as higher activity, larger body size and unique flat-beak feeding method. Therefore, in practical applications, problems such as feed accumulation, uneven feeding, and feed trough pollution often occur, seriously affecting the feeding efficiency and the growth performance of waterfowl. Especially in a high-density breeding environment, due to problems such as poor feed transportation, high blockage rate, and uneven distribution of traditional equipment, the breeding effect is greatly reduced. In addition, most of the current domestic waterfowl feeding equipment generally lacks intelligent control, and many devices still adopt fixed feeding strategies, unable to dynamically adjust according to environmental parameters and the health status of waterfowl, and it is difficult to achieve precise individual feeding. This not only exacerbates the problems of feed waste and pollution, but also restricts the improvement of breeding efficiency. Therefore, developing an intelligent feeding equipment that adapts to the special needs of waterfowl, improves feeding efficiency, reduces waste and improves the health status of waterfowl has become an urgent need for the industry development. And in the cage rack breeding mode, the breeding density is large and the scale effect is significant. It is necessary to regularly inspect every day to check the health status and growth of waterfowl. In particular, it is necessary to feed or supplement feed for waterfowl many times, and this work belongs to repetitive labor.

[0004] To solve the above problems, some researchers have currently designed intelligent feeding methods and developed corresponding devices, but there are still some problems. The prior art has designed a stepped chicken cage feeding device. The stepped design enables the feed truck to feed multiple feed troughs simultaneously, which can effectively reduce the floor area of the equipment in the farm. However, it lacks intelligent food storage and cannot automatically regulate the feeding amount of feed, which may lead to the accumulation of surplus feed. In addition, the prior art has also designed a rail-type cage frame poultry automatic feeding system and its control method, which can achieve simple food storage detection and then dynamically regulate the feeding amount of feed. However, it lacks functions such as intelligent monitoring and alarm of the feeding desire, health status, and contaminated feed of waterfowl. Summary of the Invention

[0005] In view of the above deficiencies in the prior art, the present application provides a multi-layer cage-raised waterfowl feeding method, system, and device based on behavior feedback to dynamically regulate the feeding amount of feed according to the feeding desire and health status of waterfowl, as well as intelligent monitoring and alarm of contaminated feed, reduce feed waste and pollution, and significantly improve the breeding efficiency and automation level.

[0006] To achieve the above invention purpose, the technical solution adopted by the present application is as follows: In the first aspect, the present application provides a multi-layer cage-raised waterfowl feeding method based on behavior feedback, including: S1: A high-definition camera on the mobile feed trough continuously collects image data of waterfowl in front of the feed trough and the two-dimensional code label on the breeding cage. S2: The core processing module decodes the collected two-dimensional code label to obtain the cage number of the collected waterfowl, and uses a multi-channel fusion method to segment the waterfowl image data into a waterfowl body area image and a beak area image. S3: Based on the segmented beak area image, calculate the feeding desire index of the waterfowl. S4: Based on the segmented waterfowl body area image, extract the key image features of the waterfowl body structure, construct a body condition scoring model, and combine the body condition score obtained from the feeding desire index and the body condition scoring model to control the feeding of the waterfowl. S5: A high-definition camera installed at the end of the feed trough running track collects feed trough images. The core processing module performs pollution image analysis and pollution degree judgment on the mobile feed trough, and based on the pollution degree judgment, automatically generates a feed truck cleaning prompt.

[0007] Further, the multi-channel fusion method used in S2 to segment the waterfowl body and beak from the waterfowl image data specifically includes: A1: Convert the collected image data into the HSV color space and the Lab color space, extract the S channel in the HSV color space to identify the image of the waterfowl body area; extract the a channel in the Lab color space to extract the image of the waterfowl beak area. A2: Binarize the image of the waterfowl body area in the S channel and the image of the waterfowl beak area in the a channel, and use morphological filtering to remove noise from the binarized image. A3: Use the CV model to accurately segment the image of the waterfowl body area, and use gray-scale morphology processing and Gaussian filtering to optimize the edges of the segmented image; use the RSF model, local area fitting, and narrow-band optimization algorithm to optimize the extraction of the waterfowl beak boundary in the beak area image.

[0008] Further, the S3 specifically includes: S301: Based on the segmented beak area image, analyze the opening and closing actions of the waterfowl beak and the change in the distance to the feeding trough in consecutive frames. S302: Use the action frequency, proximity, and beak speed change as indicators to construct a waterfowl feeding desire index:

[0009] Where, is the feeding desire index, is the beak opening and closing frequency per unit time, is the distance from the beak centroid to the leading edge of the feeding trough, is the beak action speed, in pixels / frame, , and are empirical coefficients.

[0010] Further, the S4 specifically includes: S401: Based on the segmented waterfowl body area image, extract the binary mask image of the waterfowl body area image. S402: Based on the segmented waterfowl body area image, extract the key image features of the waterfowl body structure. S403: Normalize the key image features to obtain the scoring values of the standardized features of each key image feature. S404: Fuse the scoring values of each standardized feature to obtain a body condition scoring model:

[0011] Where, is the waterfowl body condition score, is the body area is the scoring value of is the body length-width ratio The scoring value, is the edge curvature fluctuation of the scoring value, is the tail height ratio of the scoring value, is the symmetry coefficient of the scoring value, is the weight coefficient; S405: Feed control is performed on waterfowl based on the body condition score obtained by combining the feeding desire index and the body condition scoring model: If and , extend the residence time of the feeding truck and increase feeding; if and , do not perform feeding, and start the RDHL-DETR model to detect dead waterfowl; if and , shorten the residence time of the feeding truck, where is the lowest threshold, is the highest threshold, is the lowest body weight threshold, is the highest body weight threshold.

[0012] Furthermore, the RDHL-DETR model specifically includes: Replace the standard convolution module in the original DETR model with a RepConv module; replace the multi-head self-attention mechanism in the traditional AIFI module with a deformable multi-head attention mechanism; introduce a local feature fusion module.

[0013] Furthermore, training the RDHL-DETR model specifically includes: B1: Collect image data of dead waterfowl in different scenarios, and use annotation tools to annotate the image data; B2: Divide the annotated image data into a training set, a validation set, and a test set; B3: Train the RDHL-DETR model using the training set; B4: Evaluate the trained RDHL-DETR model using the validation set, and save the best network training weights according to the validation results; B5: Evaluate the generalization ability of the RDHL-DETR model using the test set to ensure the stability of the RDHL-DETR model.

[0014] Furthermore, the S5 specifically includes: S501: The high-definition camera installed at the terminal of the feeding trough track collects the feeding trough image; S502: Perform brightness equalization and geometric distortion correction processing on the collected feeding trough image; S503: Extract the main visual area image containing the feed trough from the processed feed trough image according to the preset feed trough structure template; S504: Use the fusion algorithm of color space, edge structure and texture complexity to detect the contaminated area of the main visual area image of the feed trough, and obtain the pollution score index; S505: Perform pollution determination based on the pollution score index and the set pollution threshold: If the pollution score index is greater than the pollution threshold, it is determined that the feed trough is seriously polluted, prompt the need for manual or automatic cleaning operations, and generate a pollution alarm record including the feed trough number, detection time, pollution score value and image screenshot.

[0015] Further, in the S504, the fusion algorithm of color space, edge structure and texture complexity is used to detect the contaminated area of the main visual area image of the short feed trough, and obtain the pollution score index, which specifically includes: C1: Convert the main visual area image into the Lab color space; C2: Enhance the areas with abnormal saturation in the image in the Lab color space; C3: Use the local gradient information of the image to extract the edges and patch contours of the feed trough, and construct a pollution mask image using color features; C4: Based on the pollution mask image, construct a pollution scoring function to quantify the pollution degree, and the pollution scoring function is:

[0016] Where is the pollution score index, is the detected area of the contaminated area, is the total area of the feed trough, is the average saturation concentration of the contaminated area, is the local information entropy of the contaminated area, and and are empirical selection coefficients set according to the proportion required by manual labor.

[0017] In the second aspect, the present application provides a multi-layer cage aquaculture waterfowl feeding system based on behavior feedback, including: A data acquisition device, including a camera and a wireless transmission module. The camera collects waterfowl image data, two-dimensional code tags on the breeding cage and feed trough images, and then transmits them to the edge image processing module through the wireless transmission module; The edge image processing module deploys a multi-channel color space image segmentation model and an image processing model, and transmits the processed image to the intelligent recognition module; The intelligent recognition module deploys a feeding desire assessment module, a waterfowl condition scoring module, an RDHL-DETR module, and a feeder pollution recognition module. The images of the waterfowl body area and the waterfowl beak area segmented by the edge image processing module are transmitted to the feeding desire assessment module and the waterfowl condition scoring module for calculation, and the calculation results are transmitted to the intelligent feeding control module as feeding instructions; the RDHL-DETR module identifies whether the waterfowl is dead according to the output of the feeding desire assessment module; the feeder image processed by the edge image processing module is used as the input of the feeder pollution recognition module, and the feeder pollution recognition module transmits the recognition result to the waterfowl intelligent feeding platform; An intelligent feeding control module, and the intelligent feeding control module controls the residence time of the mobile short feeder according to the feeding instruction; A waterfowl intelligent feeding platform, and the waterfowl intelligent feeding platform collects and analyzes data related to waterfowl feeding behaviors, waterfowl death judgment data, and feeder pollution recognition data, and generates relevant reports for display.

[0018] In a third aspect, the present application provides a multi-layer cage waterfowl feeding device based on behavior feedback, including: A plurality of breeding cages, an egg conveyor belt, a QR code label, a mobile short feeder, a waterfowl inspection high-definition camera, a feeder pollution detection high-definition camera, and a core processing module; a plurality of the waterfowl inspection high-definition cameras are mounted on the mobile short feeder, the feeder pollution detection high-definition camera is installed at the feeder inspection terminal, and the QR code label is located at the front end of the plurality of breeding cages.

[0019] The beneficial effects of the present application are: The multi-layer cage waterfowl feeding method, system, and device provided by the present application dynamically regulate the feeding amount of feed through the feeding desire and condition status of waterfowl, realize the reasonable feeding of waterfowl, reduce feed waste, and reduce feed pollution through feeder pollution recognition and feeder cleaning, significantly improving the breeding efficiency and automation level. Description of the Drawings

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application, and those of ordinary skill in the art can also obtain other embodiments according to these drawings.

[0021] Figure 1 It is a schematic flowchart of a multi-layer cage waterfowl feeding method provided by an embodiment of the present application.

[0022] Figure 2Schematic diagram of a multi - layer cage - raised waterfowl feeding system based on behavior feedback provided by an embodiment of the present application.

[0023] Figure 3 Schematic diagram of a structure of a multi - layer cage - raised waterfowl feeding device based on behavior feedback provided by an embodiment of the present application.

[0024] Among them: 1 - breeding cage, 2 - egg conveyor belt, 3 - QR code label, 4 - mobile short feeding trough, 5 - high - definition waterfowl inspection camera, 6 - high - definition camera for detecting feeding trough pollution, 7 - core processing module. Detailed implementation manners

[0025] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art based on the present application belong to the scope of protection of the present application.

[0026] Embodiment 1: An embodiment of the present application provides a multi - layer cage - raised waterfowl feeding method based on behavior feedback. This method can be referred to Figure 1 , Figure 1 As shown in the flow chart of a multi - layer cage - raised waterfowl feeding method based on behavior feedback provided by an embodiment of the present application, it includes: S1: The high - definition camera on the mobile feeding trough collects the image data of the waterfowl in front of the feeding trough and the QR code label on the breeding cage in real time.

[0027] In an embodiment of the present application, the camera is fixedly installed at the front end of the feeding vehicle. As the feeding vehicle moves horizontally along the track between cage positions, it collects the image data of the waterfowl in front of each feeding trough in real time, and synchronizes with the QR code encoding of each cage position to realize the binding of image collection and individual positioning, eliminating the deployment of multiple cameras, reducing costs while improving the traceability of waterfowl behavior tracking and body condition data.

[0028] S2: The core processing module decodes the collected QR code label to obtain the cage number of the collected waterfowl, and uses a multi - channel fusion method to segment the image data of the waterfowl into the body area image and the beak area image of the waterfowl.

[0029] Further, the multi - channel fusion method for segmenting the body and beak of the waterfowl from the image data of the waterfowl in S2 specifically includes: A1: Color channel processing. The acquired image data is converted into the HSV color space and the Lab color space. The S channel is extracted in the HSV color space for identifying the body area image of waterfowl; the a channel is extracted in the Lab color space specifically for extracting the beak area. In this way, different regions can be effectively distinguished and preparations can be made for subsequent fine segmentation; A2: Coarse segmentation. After the color channels are extracted, the system performs binary processing on the images of the S and a channels respectively. The Otsu method is used to threshold the image, with the threshold set to 0.6, and noise is removed through morphological filtering (dilation and erosion operations) to retain the main features in the image. Then, the largest connected region in the image is extracted as the initial contour of the active contour, preparing for the next step of fine segmentation; A3: Fine segmentation. On the basis of coarse segmentation, the system accurately segments the body area of waterfowl through the CV model, and optimizes the edges by combining gray-scale morphological processing and Gaussian filtering to ensure more accurate segmentation results. At the same time, for the beak area of waterfowl, the system uses the RSF model and combines local region fitting and narrow-band optimization algorithms to improve the extraction accuracy of the beak boundary, thus providing clear image data support for accurately judging the feeding actions and health status of waterfowl.

[0030] It can be understood that introducing the HSV-S and Lab-a color channels for body and beak recognition respectively effectively improves the edge segmentation accuracy and system recognition robustness in complex backgrounds.

[0031] S3: Calculate the feeding desire index of waterfowl based on the segmented beak area image.

[0032] Specifically: S301: After the beak area is segmented, track the changes in the opening and closing actions of the beak and the distance to the feeding trough in consecutive frames; S302: Set the action frequency, proximity, and beak speed change as indicators to comprehensively evaluate the feeding desire. The feeding desire index of waterfowl is:

[0033] Where, is the feeding desire index, is the opening and closing frequency of the beak per unit time, is the distance from the centroid of the beak to the leading edge of the feeding trough, is the beak action speed, in pixels per frame, 、 and are empirical coefficients, adjusted according to the training set.

[0034] It is understandable that the feeding behavior is quantified based on parameters such as the frequency of mouth opening and closing, displacement distance, and the degree of proximity to the feeding trough, so as to achieve dynamic and accurate recognition of the individual's feeding desire and control of the feeding response.

[0035] S4: Based on the segmented waterfowl body region image, extract the key image features of the waterfowl body structure, construct a body condition scoring model, and combine the feeding desire index and the body condition score obtained from the body condition scoring model to control the feeding of the waterfowl.

[0036] Specifically: S401: After the body image acquisition is completed and the image segmentation is performed, extract the binary mask image of the waterfowl body region as the input for body condition scoring. Among them, the body region should have a complete boundary to exclude other waterfowl or occlusion interference; S402: Based on the segmented waterfowl body region image, extract the key image features of the waterfowl body structure: body area : Refers to the number of pixels within the body contour region, representing the total volume; body length-width ratio : Calculate the ratio of its width to height from the minimum bounding rectangle of the segmentation contour; edge curvature fluctuation : Measure the contour smoothness through the change of the first derivative of the edge curvature; tail height ratio : Calculate the ratio of the height of the tail region to the height of the entire body; symmetry coefficient : Evaluate the left-right symmetry based on the area difference on both sides of the image central axis; S403: Feature normalization Each of the above features is processed by standardization and converted into a scoring value in the range of [0,1], denoted as: , , , ,

[0037] Among them, the normalization method uses Gaussian function fitting; S404: Fuse the scoring values of the standardized features into a comprehensive score to obtain the body condition scoring model:

[0038] Among them, is the body condition score of the waterfowl, and the value range is , is the scoring value of the body area , is the scoring value of the body length-width ratio , is the scoring value of the edge curvature fluctuation , is the scoring value of the tail height ratio The scoring value, is the symmetry coefficient The scoring value, is the weight coefficient, obtained based on expert experience; S405: Combine the feeding desire index and the body condition score obtained from the body condition scoring model to control the feeding of waterfowl: If and , it indicates that the feeding desire is strong and the body condition of the waterfowl is normal. Prolong the residence time of the feeding vehicle and increase the feeding time; If and , do not feed, prompt that there are health risks such as abnormal body shape or malnutrition in the waterfowl, and start the RDHL-DETR model to detect dead waterfowl; If and , it indicates that the weight of the waterfowl may exceed the standard. Shorten the residence time of the feeding vehicle to adjust the feeding duration of the waterfowl. Among them, is the lowest threshold, is the highest threshold, is the lowest weight threshold, is the highest weight threshold.

[0039] It can be understood that the lowest weight threshold and the highest weight threshold can be set by the farm management personnel. Calculate the body condition score of the waterfowl through the structural features (such as area, aspect ratio, edge curvature, etc.) in the image, and cooperate with the desire index to realize the automatic visual evaluation of the individual nutritional status and health level.

[0040] Furthermore, for the waterfowl death recognition mechanism triggered under the condition of low feeding desire, an improved end-to-end depth detection model RDHL-DETR is used for automatic judgment. Specifically: In one embodiment of the present application, (1) in the feature extraction stage, RDHL-DETR replaces the standard convolution module (HGBlock) in the original DETR framework with a RepConv module. This improvement enhances the computational efficiency of feature extraction and the generalization ability of the model when processing waterfowl images with different structures. The RepConv module can effectively extract more detailed information while reducing the computational cost, providing more efficient and accurate feature data for the subsequent feature fusion stage; (2) to further improve the model's processing ability for high-resolution images, RDHL-DETR introduces a deformable multi-head attention mechanism (DMHA) into the traditional AIFI module, replacing the original traditional multi-head self-attention (MHSA) mechanism. DMHA can dynamically perceive the target deformation and position offset in the image, especially showing stronger robustness when detecting the postures of dead waterfowl such as "abdomen up" or "head down". This improvement significantly enhances the recognition accuracy of the model in occluded or high-dynamic scenarios; (3) when dealing with complex backgrounds or partial occlusions, traditional cross-scale feature fusion methods (such as CCFF) may result in the loss of small target information. To overcome this problem, RDHL-DETR introduces a local feature fusion module (LFFM). This module adaptively adjusts the model's perception ability of the region of interest by selectively expanding the receptive field, effectively extracting the detailed features of dead waterfowl, such as "fluffy feathers" and "straight legs", significantly improving the recognition accuracy of the model in complex environments.

[0041] It can be seen that an improved DETR (Detection Transformer) model framework - RDHL-DETR is proposed to improve the detection accuracy and speed of dead waterfowl in complex environments. The model has made multiple technical improvements based on DETR and can effectively handle the influence of environmental factors such as occlusion, uneven illumination, and complex backgrounds on the detection accuracy. By integrating RepConv, deformable multi-head attention mechanism (DMHA), and local feature fusion module (LFFM), RDHL-DETR solves the problem of low recognition accuracy of traditional models under occlusion, complex illumination, and variable postures, significantly improving the performance and robustness of the detection model, especially the recognition ability of small targets (such as dead waterfowl) in dynamic environments.

[0042] Furthermore, the RDHL-DETR model is trained, specifically including: First, collect image data containing waterfowl deaths in different scenarios and use the Labelme annotation tool to accurately annotate the images. Then, divide the dataset into a training set, a validation set, and a test set according to the ratio of 7:2:1. To ensure the training effect, the size of all images is uniformly adjusted to 640x640 pixels. During the training process, use RDHL-DETR as the model architecture and conduct 200 rounds of training. After each round of training, use the validation set to evaluate the model and save the best network training weights according to the validation results. Finally, evaluate the generalization ability of the model through the test set to ensure the stable performance of the model in actual applications and avoid overfitting. This training process helps to improve the accuracy and robustness of the model in the stain detection task. At the same time, through the evaluation of the validation set and the test set, ensure that the model has good generalization ability and can work stably in different waterfowl death scenarios.

[0043] S5: The high-definition camera installed at the terminal of the feeding trough operation track collects the feeding trough images. The core processing module analyzes the pollution images of the mobile feeding trough and judges the pollution degree, and based on the pollution degree judgment, automatically generates a cleaning prompt for the feeding truck.

[0044] Specifically: S501: The camera is fixedly installed at the terminal detection position of the feeding truck operation track. After the feeding truck completes the entire feeding process, each feeding trough is aligned with the terminal camera in turn, and accurate image acquisition is achieved through motor or guide rail positioning. S502: The collected images are first processed by brightness equalization and geometric distortion correction to remove visual interference caused by light fluctuations or angle errors. S503: According to the preset feeding trough structure template, extract the main visual region (ROI) containing the short feeding trough in the image for subsequent pollution analysis. The region extraction can be completed by combining structure projection and gray-scale matching techniques to ensure the accuracy and consistency of positioning. S504: After obtaining the feeding trough ROI region image, the system uses a fusion algorithm based on color space, edge structure, and texture complexity to detect the pollution region. Specifically: First, enhance the saturation abnormal regions in the image in the Lab color space, and extract dark, brown, or blackish regions as pollution candidate pixels by setting thresholds. Secondly, use the local gradient information of the image to extract the edges and patch contours of the feeding trough, and combine color features to construct a pollution mask image. Finally, quantify the pollution degree through the following pollution scoring function:

[0045] Among them, is the pollution scoring index, is the area of the detected pollution region, is the total area of the feeding trough, is the average saturation concentration of the polluted area (color concentration index), is the local information entropy of the polluted area, reflecting texture complexity and distribution dispersion), and and are empirical selection coefficients set according to the proportion of manual labor required, used to balance the impact of various indicators on the final score. This score comprehensively considers the pollution range, color characteristics, and distribution structure, and can effectively distinguish different pollution levels such as slight material residue and severe fecal contamination.

[0046] S505: Pollution warning determination and prompt. When the calculated pollution score is greater than the set threshold (such as 0.65), it is considered that the feeder is severely polluted and manual or automatic cleaning operations are required. At this time, the intelligent waterfowl feeding system will automatically generate a pollution alarm record, including the feeder number, detection time, pollution score value, and image screenshot, and synchronize the alarm information through the management terminal pop-up window, APP push, or background management system to remind the breeding personnel to deal with it in time. At the same time, the intelligent waterfowl feeding system will record the historical pollution score change curve of each feeder, which can be used for subsequent feeder structure improvement, position adjustment, or cleaning frequency optimization decisions.

[0047] Embodiment 2: The present application provides a multi-layer cage-type waterfowl feeding system based on behavior feedback. This system is as Figure 2 shown, Figure 2 is a schematic diagram of a multi-layer cage-type waterfowl feeding system based on behavior feedback provided by an embodiment of the present application, including: A data acquisition device, including a camera and a wireless transmission module. The camera acquires waterfowl image data, two-dimensional code labels on the breeding cages, and feeder images, and then transmits them to the edge image processing module through the wireless transmission module; The edge image processing module deploys a multi-channel color space image segmentation model and an image processing model, and transmits the processed images to the intelligent recognition module; The intelligent recognition module deploys a feeding desire evaluation module, a waterfowl physical condition scoring module, an RDHL-DETR module, and a feeder pollution recognition module. The waterfowl body area image and waterfowl beak area image segmented by the edge image processing module are transmitted to the feeding desire evaluation module and the waterfowl physical condition scoring module for calculation, and the calculation results are transmitted to the intelligent feeding control module as feeding instructions; the RDHL-DETR module identifies whether the waterfowl is dead according to the output of the feeding desire evaluation module; the feeder image processed by the edge image processing module is used as the input of the feeder pollution recognition module, and the feeder pollution recognition module transmits the recognition result to the intelligent waterfowl feeding platform; An intelligent feeding control module, which controls the residence time of the mobile short feed trough according to the feeding instruction; A waterfowl intelligent feeding platform, which collects and analyzes data related to waterfowl feeding behaviors, waterfowl death judgment data, and feed trough pollution identification data, and generates relevant reports for display.

[0048] In an embodiment of the present application, the waterfowl intelligent feeding platform will record the historical pollution score change curves of each feed trough, which can be used for subsequent decisions on feed trough structure improvement, position adjustment, or cleaning frequency optimization; and the waterfowl intelligent feeding platform relies on core technologies such as camera vision recognition, image analysis, and individual behavior modeling to record data such as the feeding time, feeding desire index, cage position, feed truck residence time, actual feeding amount, feed remaining, feeding behavior images, and body condition scores of each waterfowl, forming a waterfowl individual feeding behavior report; forming a group waterfowl feeding performance report, which is used to display the overall feeding trend, average feeding amount, residual feed rate, and body condition score distribution of waterfowl in each area, assisting in optimizing the feeding strategy and feed truck operation path, and improving the group feeding efficiency and resource utilization rate; when the feeding desire index or body condition score is continuously abnormal, the system calls the death recognition model to automatically judge and alarm, generating a death report containing fields such as location information, recognition screenshots, and confidence levels, assisting in precise cleaning and health management; after the feed truck completes a round of feeding, the terminal camera sequentially conducts pollution identification and scoring on all short feed troughs. If the pollution degree is higher than the set threshold, the system automatically generates a pollution warning report, prompting the administrator to arrange cleaning and maintenance in a timely manner to avoid feed waste and disease transmission; supports users to feedback problems such as abnormal feeding, missed death detection, feed trough pollution, or equipment failures through a small program, and the waterfowl intelligent feeding platform automatically summarizes and generates a processing report to improve the feedback closed-loop efficiency and service quality; provides operation management functions: provides functions such as setting feeding desire and body condition score thresholds, monitoring the operation status of equipment, configuring regional feeding parameters, managing feed trough pollution scores, and managing individual files, supporting refined breeding control of different regions, groups, and stages; supports chart display of multi-dimensional data, including feeding behaviors, body condition changes, death warnings, pollution scores, equipment operation status, and user feedback processing, forming a visual data dashboard to achieve "grasping the overall operation situation of the whole field with one picture".

[0049] Embodiment 3: The present application provides a multi-layer cage-type waterfowl feeding device based on behavior feedback. This device is as Figure 3 shown, Figure 3 is a schematic diagram of a multi-layer cage-type waterfowl feeding device based on behavior feedback provided by an embodiment of the present application, including: A number of breeding cages 1, egg conveyor belts 2, QR code tags 3, mobile short feed troughs 4, high-definition cameras 5 for waterfowl inspection, high-definition cameras 6 for detecting feed trough pollution, and a core processing module 7; a plurality of the high-definition cameras 5 for waterfowl inspection are mounted on the mobile short feed trough 4, the high-definition camera 6 for detecting feed trough pollution is installed at the feed trough inspection terminal, and the QR code tag 3 is located at the front end of the number of breeding cages 1.

[0050] A method, system and device for feeding multi-layer cage waterfowl based on behavior feedback provided by this application integrate a camera at the front end of a mobile short feed trough and combine QR code positioning to realize the synchronous binding of image acquisition and individual information, eliminate the deployment of multiple cameras, improve the traceability of waterfowl behavior tracking and body condition data while reducing costs; introduce HSV-S and Lab-a color channels for body and beak recognition respectively, effectively improving the edge segmentation accuracy and system recognition robustness in complex backgrounds; jointly control the feed feeding amount based on the waterfowl body condition index and feeding desire; at the same time, automatically enable the death recognition model when the feeding desire is continuously abnormal, and use the improved DETR structure combined with the attention mechanism to achieve high-accuracy recognition of dead waterfowl, significantly reducing missed detections and false alarms. In addition, after the inspection of the feed truck is completed, the degree of feed trough pollution is detected by the terminal camera, and if the pollution score exceeds the threshold, a cleaning reminder is triggered to improve the efficiency of environmental hygiene management.

[0051] It should be noted that those of ordinary skill in the art will realize that the embodiments described here are to help readers understand the principles of this application, and it should be understood that the protection scope of this application is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of this application based on the technical revelations disclosed in this application, and these deformations and combinations are still within the protection scope of this application.

Claims

1. A multi-layer cage feeding method for waterfowl based on behavior feedback, characterized in that, Including: S1: The high-definition camera on the mobile feeding trough collects the image data of waterfowl in front of the feeding trough and the QR code label on the breeding cage in real time; S2: The core processing module decodes the collected QR code label to obtain the cage number of the waterfowl being collected, and uses a multi-channel fusion method to segment the image data of the waterfowl into the body area image and the beak area image of the waterfowl; S3: Based on the segmented beak area image, calculate the feeding desire index of the waterfowl; S4: Based on the segmented body area image of the waterfowl, extract the key image features of the waterfowl body structure, construct a body condition scoring model, and combine the feeding desire index and the body condition score obtained from the body condition scoring model to control the feeding of the waterfowl; S5: The high-definition camera installed at the terminal of the feeding trough running track collects the feeding trough image, and the core processing module analyzes the pollution image of the mobile feeding trough and judges the pollution degree, and based on the pollution degree judgment, an automatic cleaning prompt for the feeding truck will be generated.

2. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 1, wherein, In the S2, the method of using multi-channel fusion to segment the body and beak of the waterfowl from the image data of the waterfowl specifically includes: A1: Convert the collected image data into the HSV color space and the Lab color space, extract the S channel in the HSV color space to identify the body area image of the waterfowl; extract the a channel in the Lab color space to extract the beak area image of the waterfowl; A2: Binary process the body area image of the waterfowl in the S channel and the beak area image of the waterfowl in the a channel, and use morphological filtering to remove noise from the binary-processed image; A3: Use the CV model to accurately segment the body area image of the waterfowl, and use gray-scale morphology processing and Gaussian filtering to optimize the edges of the segmented image; use the RSF model, local area fitting and narrowband optimization algorithm to optimize the extraction of the waterfowl beak boundary in the beak area image.

3. The method for feeding multi-layer cage - raised waterfowl based on behavior feedback according to claim 2, characterized in that, The S3 specifically includes: S301: Based on the segmented beak area image, analyze the opening and closing actions of the waterfowl beak and the distance change close to the feeding trough in consecutive frames; S302: Use the action frequency, closeness, and beak speed change as indicators to construct the feeding desire index of the waterfowl: Among them, is the feeding desire index, is the opening and closing frequency of the mouth per unit time, is the distance from the centroid of the mouth to the leading edge of the feeding trough, is the mouth movement speed, with the unit of pixel / frame, 、 and are empirical coefficients.

4. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 3, wherein, The S4 specifically includes: S401: Based on the segmented body area image of the waterfowl, extract the binary mask image of the body area image of the waterfowl; S402: Based on the segmented body area image of the waterfowl, extract the key image features of the waterfowl body structure; S403: Normalize the key image features to obtain the scoring values of the standardized features of each key image feature; S404: Fuse the scoring values of each standardized feature to obtain the body condition scoring model: Among them, is the body condition score of waterfowl, is the score value of the body area , is the score value of the body length-width ratio , is the score value of the edge curvature fluctuation , is the score value of the tail height ratio , is the score value of the symmetry coefficient , is the weight coefficient; S405: Based on the body condition score obtained by combining the feeding desire index and the body condition scoring model, control the feeding of waterfowl: If and , extend the residence time of the feed truck and increase feeding; If and , do not feed and start the RDHL-DETR model to detect dead waterfowl; If and , shorten the residence time of the feed truck, where is the lowest threshold, is the highest threshold, is the lowest weight threshold, is the highest weight threshold.

5. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 4, wherein The RDHL-DETR model specifically includes: Replace the standard convolution module in the original DETR model with a RepConv module; replace the multi-head self-attention mechanism in the traditional AIFI module with a deformable multi-head attention mechanism; introduce a local feature fusion module.

6. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 5, wherein, Training the RDHL-DETR model specifically includes: B1: Collect the image data of waterfowl deaths in different scenarios, and use annotation tools to annotate the image data; B2: Divide the labeled image data into a training set, a validation set, and a test set; B3: Use the training set to train the RDHL-DETR model; B4: Use the validation set to evaluate the trained RDHL-DETR model, and save the best network training weights according to the validation results; B5: Use the test set to evaluate the generalization ability of the RDHL-DETR model to ensure the stability of the RDHL-DETR model.

7. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 1, wherein The specific steps of S5 include: S501: The high-definition camera installed at the terminal of the trough running track collects trough images; S502: Perform brightness equalization and geometric distortion correction on the collected trough images; S503: According to the preset trough structure template, extract the main visual area image containing the trough from the processed trough images; S504: Use the fusion algorithm of color space, edge structure, and texture complexity to detect the contaminated area in the main visual area image of the trough, and obtain the pollution score index; S505: Based on the pollution score index and the set pollution threshold, perform pollution determination: if the pollution score index is greater than the pollution threshold, it is determined that the trough is seriously polluted, prompting the need for manual or automatic cleaning operations, and generating a pollution alarm record containing the trough number, detection time, pollution score value, and image screenshot.

8. The method for feeding multi-layer cage-raised waterfowl based on behavior feedback according to claim 7, wherein In S504, the fusion algorithm of color space, edge structure, and texture complexity is used to detect the contaminated area in the main visual area image of the short trough, and obtain the pollution score index, which specifically includes: C1: Convert the main visual area image to the Lab color space; C2: Enhance the abnormal saturation area in the image in the Lab color space; C3: Use the local gradient information of the image to extract the trough edge and patch contour, and construct a pollution mask image using color features; C4: Based on the pollution mask image, construct a pollution scoring function to quantify the pollution degree, and the pollution scoring function is: Among them, is the pollution scoring index, is the area of the detected pollution area, is the total area of the feeding trough, is the average saturation concentration of the pollution area, is the local information entropy of the pollution area, , and are empirical selection coefficients set according to the proportion of manual labor required.

9. A multi-layer cage aquaculture waterfowl feeding system based on behavior feedback, characterized in that, including: The data acquisition device includes a camera and a wireless transmission module. After the camera collects waterfowl image data, the QR code label on the breeding cage, and the trough image, it transmits them to the edge image processing module through the wireless transmission module; The edge image processing module deploys a multi-channel color space image segmentation model and an image processing model, and transmits the processed image to the intelligent recognition module; The intelligent recognition module deploys a feeding desire evaluation module, a waterfowl physical condition scoring module, an RDHL-DETR module, and a trough pollution recognition module. The waterfowl body area image and the waterfowl beak area image segmented by the edge image processing module are transmitted to the feeding desire evaluation module and the waterfowl physical condition scoring module for calculation, and the calculation results are transmitted to the intelligent feeding control module as feeding instructions; the RDHL-DETR module identifies whether the waterfowl is dead according to the output of the feeding desire evaluation module; the trough image processed by the edge image processing module is used as the input of the trough pollution recognition module, and the trough pollution recognition module transmits the recognition result to the waterfowl intelligent feeding platform; The intelligent feeding control module, and the intelligent feeding control module controls the residence time of the mobile short trough according to the feeding instruction; The waterfowl intelligent feeding platform collects and analyzes data related to waterfowl feeding behaviors, waterfowl death judgment data, and feeder pollution identification data, and generates relevant reports for display.

10. A multi-layer cage aquaculture waterfowl feeding device based on behavior feedback, characterized in that, It includes: Several breeding cages (1), egg conveyor belts (2), QR code tags (3), mobile short feeders (4), high-definition waterfowl inspection cameras (5), high-definition feeder pollution detection cameras (6), and a core processing module (7); a plurality of the high-definition waterfowl inspection cameras (5) are mounted on the mobile short feeder (4), the high-definition feeder pollution detection camera (6) is installed at the feeder inspection terminal, and the QR code tag (3) is located at the front end of the several breeding cages (1).

Images