CN-122024029-A - Sturgeon hunger state detection method based on generation countermeasure and YOLO model

CN122024029ACN 122024029 ACN122024029 ACN 122024029ACN-122024029-A

Abstract

The invention relates to the technical field of aquaculture intellectualization, and particularly discloses a sturgeon hunger state detection method based on a generation countermeasure and YOLO model, which comprises the following steps that step S1, video image data of residual baits and feeding conditions in water flow are collected in a sturgeon culture pond by utilizing an underwater camera; the method comprises the steps of S2, optimizing and expanding acquired video image data through an improved generation countermeasure network to obtain an expanded data set, S3, pre-training an improved YOLO-v8 model, detecting the optimized and expanded image data in the expanded data set by utilizing the YOLO-v8 model obtained through pre-training, and S4, evaluating the ingestion state and hunger state of sturgeons. The method is suitable for residual bait statistics and fish hunger state analysis in sturgeon industrial culture, provides accurate data support for bait feeding management, optimizes bait utilization efficiency, reduces culture cost and provides efficient technical support for intelligent aquaculture.

Inventors

PANG TAO
Zha Songyou
CHEN XIAOYAN
CHEN DEFANG
LIU CHENMING
XIA HAORAN

Assignees

四川农业大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (7)

1. A sturgeon hunger state detection method based on a generation countermeasure and YOLO model is characterized by comprising the following steps: s1, collecting video image data of residual baits and feeding conditions in water flow in a sturgeon culture pond by using an underwater camera; step S2, optimizing and expanding the acquired video image data through the improved generation countermeasure network to obtain an expanded data set; S3, pre-training an improved YOLO-v8 model, wherein the improved YOLO-v8 model consists of a back bone layer, a Neck layer and a Head layer, judging the performance of the model according to a loss function, continuously updating model training parameters to obtain a pre-trained YOLO-v8 model, and detecting image data after optimizing and expanding in an expanded data set; And S4, inputting an output result of the YOLO-v8 model into a sturgeon hunger degree calculation model, and evaluating the feeding state and the hunger state of the sturgeon by using the model.
2. The sturgeon hunger state detection method based on the model of generating challenge and YOLO according to claim 1, wherein the step S1 comprises: Step S101, a waterproof underwater camera is selected and arranged below a sturgeon culture pond feeder, a lens is aligned with the fed submerged feed, and a monitoring visual field is ensured to cover all underwater feeds; step S102, after sturgeon finishes eating, starting a camera to start recording, and continuously collecting video to obtain video image data covering residual bait; and step 103, performing preliminary processing on the collected video image data, cutting out fragments irrelevant to residual bait, and reserving a key picture.
3. The sturgeon hunger detection method based on the generated challenge and YOLO models according to claim 1, wherein in the step S2, the improved generated challenge network is adopted, the original video image data is optimized through the generated network, and the quality of the video image data is improved, which comprises the following steps: step S201, selecting video image data acquired by a part as a sample set based on the improved generation countermeasure network; The improved generation countermeasure network is a network obtained by replacing standard convolution in the generation network and the discrimination network with depth separable convolution, wherein the depth separable convolution comprises depth convolution and point-by-point convolution; Inputting video image data in the sample set into an improved generation countermeasure network for training; in the process of generating an countermeasure network training for an image enhancement task, a core mechanism of the model is a dynamic game between a generating network and a judging network: The generating network takes the original video image data as input to generate an enhanced image with clear details, true colors and improved contrast; the judging network is responsible for distinguishing whether the input image is real video image data or generating an enhancement result generated by the network; The generating network and the judging network are trained alternately, namely, firstly, the generating network is fixed, the judging network is optimized to accurately identify the authenticity; the video image data in the sample set is subjected to cyclic countermeasure, the generated network continuously adjusts parameters to approximate the distribution of real clear images, meanwhile, the discrimination capability of the discrimination network is continuously improved, and the generated network is forced to generate more realistic enhanced results; finally obtaining a high-performance and stable generation network, namely a trained generation network, through multiple rounds of iteration and constraint of a loss function; Step S202, inputting all acquired video image data into a generation network obtained after training to obtain a new generation picture, namely optimizing the expanded image data; step S203, adding all the optimized and expanded image data into a set to form an optimized and expanded data set.
4. The method for detecting hungry states of sturgeons based on generating a model of countermeasure and YOLO according to claim 3, wherein in step S3, the input image is processed by the modified YOLO-v8 model as follows: step 301, inputting optimized and expanded image data as an input image into a YOLO-v8 model, and completing corresponding initialization preprocessing; Step S302, inputting the initialized and preprocessed image into a backstone layer, and sequentially processing the initialized and preprocessed image by 4 ShuffleBlock modules, wherein the last three modules respectively output characteristic diagrams F2, F3 and F4 with different scales; step S303, inputting the feature graphs F2, F3 and F4 output by the back plane to the Neck plane, and outputting feature graphs P3, P4 and P5 through improved lightweight feature pyramid network processing; Step S304, inputting the feature graphs P3, P4 and P5 output by the Neck layer into a Head layer, wherein the layer comprises three detection heads with three scales, which respectively correspond to the detection of large, medium and small targets, and finally outputting a detection frame, a category and the confidence level thereof.
5. The sturgeon hunger state detection method based on the generation of the challenge and YOLO model according to claim 4, wherein the improved YOLO-v8 model in step S3 comprises: the Backbone layer comprises a ShuffleNetV Backbone network and an SPPF module added at the tail end, wherein the ShuffleNetV Backbone network is formed by stacking 4 pieces of ShuffleBlock and is used for gradually downsampling and extracting features; Neck layers, adopting a lightweight characteristic pyramid network, comprising a channel adaptation layer, an up-sampling fusion block and a down-sampling fusion block: three different scale feature maps F2, F3 and F4 extracted from ShuffleNetV backbone network are subjected to 1X 1 convolution channel adaptation layer to unify the channel number; Then the up-sampling fusion block executes a top-down path, deep features F4 are fused with F3 through up-sampling to generate P4, and then P4 up-sampling is fused with F2 to generate P3; Then, the downsampling fusion block executes a bottom-up path, P3 is fused with P4 through downsampling to generate P4', and then P4' downsampling is fused with F4 to generate P5; All fusion operations adopt a mode of adding element by element and then connecting light convolution, and embed simplified channel attention to strengthen important features, and finally output three-scale feature graphs P3, P4 and P5 and directly send the feature graphs to a Head layer; and a Head layer containing three dimensions of detection heads.
6. The sturgeon hunger state detection method based on the generation of the countermeasure and YOLO model according to claim 5, wherein the pre-training improved YOLO-v8 model comprises selecting part of the expanded image data from the optimized expanded data set, labeling the bounding box and the category, and pre-training the improved YOLO-v8 model by using the labeled image data as an input image to obtain a pre-trained YOLO-v8 model; the method for detecting the optimized and expanded image data in the expanded data set by utilizing the pre-trained YOLO-v8 model comprises the following steps: and respectively taking all image data after the expansion in the expanded data set as input images, inputting a pre-trained YOLO-v8 model, and outputting a detection result by the YOLO-v8 model, wherein the detection result comprises the positions of the boundary boxes in the input images, confidence scores corresponding to the boundary boxes and corresponding categories.
7. The sturgeon hunger state detection method based on the model of generating challenge and YOLO according to claim 1, wherein the step S4 comprises: S401, obtaining a boundary box list according to a detection result output by the YOLO-v8 model, wherein the boundary box list comprises boundary box positions, confidence scores corresponding to the boundary boxes and categories corresponding to the boundary boxes; S402, data preprocessing, namely filtering low confidence detection and calculating the area of each boundary box ; S403, taking each boundary box as a detection target, and performing basic index calculation: Calculating the total target quantity ; Calculate the total area and: ; Calculating average nearest neighbor distance: , is the distance of the ith target to the nearest neighbor target; s404, calculating comprehensive density; calculating residual bait density fraction in the image based on target detection result of YOLO model The method comprises the following steps: Wherein the method comprises the steps of For the reference value of the maximum target number, As the total area of the image is, In order to be a number of the products, Is the area of the substrate to be processed, The constraint conditions of the aggregation degree weight are as follows: s405, calculating a final score; Wherein the method comprises the steps of The method comprises the steps of taking the sturgeon feeding ratio as a theoretical maximum value, wherein the feeding ratio of sturgeon is lower as the final score is higher, the sturgeon feeding ratio is higher, the residual baits in an image are more and denser, and the hunger state is judged by comparing the score change of fixed time after feeding: if the final fraction drop value of the fixed time after feeding is greater than or equal to a preset threshold value, the starvation state is starvation; If the final fractional drop over a fixed time after feeding is less than a preset threshold, the starvation condition is "satiety".

Description

Sturgeon hunger state detection method based on generation countermeasure and YOLO model Technical Field The invention relates to the technical field of aquaculture intellectualization, in particular to a sturgeon hunger state detection method based on a generation countermeasure and YOLO model, which is suitable for bait casting management in an industrial sturgeon cultivation scene. Background In an industrial sturgeon culture system, accurate evaluation of the feeding state of a fish school is an important management link, and the method directly relates to optimization of feed utilization efficiency, improvement of overall culture benefit and reasonable control of production operation cost. Currently, the mainstream industrial sturgeon culture adopts a high-density circulating water culture system, and conventional ingestion evaluation depends on a manual observation method, namely about 20 to 30 minutes after each feeding is finished, and a culture staff visually detects and evaluates the quantity of residual baits at a water outlet of the system. According to experience, if the residual bait quantity is observed to be obviously excessive, the feeding excess is judged, and the feeding quantity of the next time is adjusted accordingly, so that feed waste and water quality deterioration are avoided. However, the traditional method relying on manual experience and visual judgment has the remarkable defects of low working efficiency and high labor cost, and is influenced by subjective judgment, visual errors and environmental factors of individuals, so that the evaluation result is poor in consistency and large in error. Under the large-scale and high-density industrial cultivation scene, the traditional method is increasingly difficult to meet the urgent requirements of fine and real-time management. In the emerging sturgeon culture feeding state evaluation method, a technology based on computer vision gradually becomes a key point of research and application. The technology mainly relies on a camera to collect the residual bait images at the bottom of a culture pond or directly capture the behavior pictures of gathering and ingestion of fish shoals, and judges the real-time ingestion liveness and the feed consumption condition of fish by analyzing the visual information so as to further predict whether the current feeding amount is sufficient. According to the analysis result, the system can automatically decide whether to continuously throw feed or not, and finally, the feed throwing is accurately regulated and controlled according to the feeding demand, so that the purposes of saving cost and optimizing the cultivation effect are achieved. However, the model is often dependent on artificial feature design, has limited adaptability to complex underwater scenes, and has low prediction accuracy generally, and meanwhile, the model is often redundant in structure, large in volume and low in inference speed when dealing with large-scale and high-dimensional data, so that the real-time online monitoring requirement is difficult to meet. In the industrial sturgeon cultivation process, the feed used is usually a sinking pellet feed, which means that the feed can be settled to the bottom of a cultivation pond, so that residual distribution of the feed and feeding activity of fish need to be observed by means of underwater shooting technology. However, images acquired by underwater photography are extremely susceptible to interference from a variety of external factors, including insufficient illumination intensity or maldistribution, changes in turbidity of the water, shielding of various obstacles in the water, and the like. The fundamental problem is that most underwater images often have significant degradation phenomena during acquisition due to the comprehensive influence of special physical and chemical environments under water, such as scattering and absorption effects of light in water, diffusion of suspended particles and chemical action of dissolved substances. This degradation is manifested by blurred image details, significant overall contrast degradation, severe color distortion, and random noise interference, which results in a significant reduction in visual information quality. The degraded underwater image not only brings great challenges to subsequent image enhancement and restoration work and reduces the imaging reliability and practicality, but also limits the effective application of the image in culture monitoring. In addition, due to the variability of the underwater environment, images acquired in different scenes often have complex and various color distribution characteristics, which are mainly influenced by factors such as water depth, water quality components, artificial light source conditions and the like. Therefore, conventional image enhancement techniques are generally difficult to adapt to such non-uniform and dynamically changing color distribution, and have limited processing