Fish feeding method and device based on multi-task active learning framework

Abstract

The invention discloses a fish feeding method and device based on a multi-task active learning framework, and the method comprises the steps: fusing the active learning framework on the basis of a multi-task learning model, and respectively calculating the uncertainty scores of a feeding intensity classification task and a residual feed counting task in combination with information entropy and approximate Bayesian reasoning, and calculating a total uncertainty score in a weighting mode to select a to-be-marked sample in an active learning process. Through interactive training of the annotation expert and the model, not only can the performance similar to that of completely supervised learning be achieved with fewer training samples, but also errors in manual annotation can be often corrected by the model in the interaction process, so that the effects of reducing the data annotation cost and improving the model recognition accuracy are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fish farming, and in particular relates to a fish feeding method and device based on a multi-task active learning framework. Background Art

[0002] As an innovative farming model, marine ranching aims to meet the growing market demand for aquatic products and reduce over-reliance on wild fishery resources by implementing large-scale and intensive farming strategies in the marine environment. Its efficient operation relies on cutting-edge intelligent feeding technology, which integrates a variety of advanced means such as sensor monitoring, computer vision recognition and artificial intelligence algorithms to achieve real-time monitoring of the farming environment and precise management of fish status.

[0003] Although AI-based aquaculture feeding methods have significant advantages in terms of scientificity and accuracy, and have demonstrated excellent performance in areas such as large-scale aquaculture needs, their application is also accompanied by a series of technical challenges. First, the cost of model training is high. Not only does it require the classification and labeling of a large amount of image data for feeding intensity, but it also requires accurate labeling of the location of leftover bait. This process involves a lot of data preprocessing and manual intervention, which greatly increases the training cost of the model. Secondly, because the classification characteristics of feeding intensity in different feeding scenarios may not be obvious enough, and the high density of leftover bait makes positioning difficult, labeling errors are prone to occur during the classification process, such as over-labeling, under-labeling, or wrong labeling of leftover bait locations. These problems will directly affect the accuracy and generalization ability of the training model. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a fish feeding method and device based on a multi-task active learning framework.

[0005] To achieve the above object, the present invention adopts the following technical solution: A fish feeding method based on a multi-task active learning framework, comprising: Step S1, constructing a fish feeding behavior dataset and a multi-task active learning framework; wherein the multi-task active learning framework includes: labeling experts, unlabeled sample pool , mark sample pool , multi-task learning models and multi-task reasoning networks; Step S2: Randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The multi-task learning model includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and residual bait counting network are connected to the feature sharing network respectively. Step S3: Use the multi-task reasoning network to perform sample reasoning on the initial model in the unlabeled sample pool, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression for each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top uncertainty score with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained based on the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; Step S4: input the real-time video of the fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

[0006] As a preference, unlabeled sample pools Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training; labeled sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

[0007] Preferably, in step S3, the query strategy function for unlabeled sample selection is used to determine the samples to be labeled in the iterative process; wherein the query strategy function includes: a classification query strategy and a regression query strategy, the classification query strategy uses category information entropy to calculate the classification uncertainty of the feeding intensity of each image; the regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the leftover bait count of each sample image; the total uncertainty score of each image is determined by weighted summation, and the samples to be labeled in the iterative process are determined according to the total uncertainty score.

[0008] Preferably, the multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by the current round of training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by the current round of training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence.

[0009] The present invention also provides a fish feeding device based on a multi-task active learning framework, comprising: The first processing module is used to construct a fish feeding behavior data set and a multi-task active learning framework; wherein the multi-task active learning framework includes: a labeling expert, an unlabeled sample pool , mark sample pool , multi-task learning models and multi-task reasoning networks; The second processing module is used to randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The multi-task learning model includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and residual bait counting network are connected to the feature sharing network respectively. The third processing module is used to perform sample reasoning on the initial model in the unlabeled sample pool through the multi-task reasoning network, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression of each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top model with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained based on the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; The fourth processing module is used to input the real-time video of fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

[0010] As a preference, unlabeled sample pools Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training; labeled sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

[0011] Preferably, the third processing device determines the samples to be marked in the iteration process through the query strategy function of unlabeled sample selection; wherein the query strategy function includes: classification query strategy and regression query strategy, the classification query strategy uses category information entropy to calculate the classification uncertainty of feeding intensity of each image; the regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the leftover bait count of each sample image; the total uncertainty score of each image is determined by weighted summation, and the samples to be marked in the iteration process are determined according to the total uncertainty score.

[0012] Preferably, the multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by the current round of training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by the current round of training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence.

[0013] Based on the multi-task learning model, the present invention integrates the active learning framework, and combines information entropy and approximate Bayesian reasoning to select the uncertainty scores of the feeding intensity classification task and the residual bait counting task in the samples to be labeled. It can not only achieve performance similar to that of fully supervised learning with fewer training samples, but also enable the model to correct errors in manual labeling during the interaction process, so as to reduce the cost of data labeling and improve the recognition accuracy of the model, providing a feasible technical solution for solving the problem of precise feeding of fish. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0015] Figure 1 This is a flow chart of a fish feeding method based on a multi-task active learning framework according to an embodiment of the present invention; Figure 2 Schematic diagram of a multi-task active learning framework according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the feeding intensity classification network; Figure 4 It is the structural diagram of the residual bait counting network; Figure 5 Schematic diagram of the structure of the approximate Bayesian inference network. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Embodiment 1: like Figure 1 , 2 As shown, an embodiment of the present invention provides a fish feeding method based on a multi-task active learning framework, comprising: Step S1, constructing a fish feeding behavior dataset and a multi-task active learning framework; wherein the multi-task active learning framework includes: labeling experts, unlabeled sample pool , mark sample pool , multi-task learning model and multi-task reasoning network; divide the fish feeding behavior dataset into training set and test set; label the feeding intensity classification label and the residual bait location label for the test set, and do not label the feeding intensity classification label and the residual bait location label for the training set; store the training set in the unlabeled sample pool middle; Step S2: Randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. The annotation content includes the feeding intensity category and the amount of residual bait. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. Step S3: Use the multi-task reasoning network to perform sample reasoning on the initial model in the unlabeled sample pool, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression for each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top candidate with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained on the basis of the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; the threshold of the evaluation index is: classification accuracy>90%, mean absolute error<20, root mean square error<25; Step S4: input the real-time video of the fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

[0019] As an implementation of an embodiment of the present invention, in step S1, constructing a fish school feeding behavior data set includes: Step 11: Use the constructed experimental environment to obtain a video of the feeding behavior of the fish school; wherein the experimental environment includes: a recirculating aquaculture system, a video acquisition system, and a data transmission system; Step 12, obtaining a fish school feeding image based on the fish school feeding behavior video; wherein, the image is screened by extracting one frame every 5 seconds to obtain the fish school feeding image; Step 13: Crop the fish feeding image to remove the irrelevant background in the fish feeding image. The resolution of the cropped image is 1024. 1024, then the cropped fish feeding image is scaled, and the final scaled image resolution is 512 512; Step 14: The cropped and scaled fish feeding images are used as a fish feeding behavior dataset; wherein the fish feeding behavior dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1:2.

[0020] Furthermore, the multi-task active learning framework includes: annotation experts ( ), unlabeled sample pool , mark sample pool and multi-task learning models Through the connection of unlabeled data streams, labeled data streams and human intervention, the next round of training can be carried out only after each sample selection and expert labeling by the active learning algorithm.

[0021] Annotation Expert ( ): Use labelme software to annotate the test set. The image label content contains two parts: feeding intensity classification label and residual bait location label.

[0022] The intensity of the feeding images is divided into two categories of labels - "Active" and "Not Active", as shown in Table 1. The annotated classification labels are saved in txt format.

[0023] Labeling classification standard: First, two experienced experts will conduct a classification evaluation of the feeding intensity of the image. If the evaluation results are the same, the label of the image will be determined. If they are different, a third expert will be invited to conduct a labeling evaluation until the classification evaluation results of each labeling expert are consistent. Table 1 For marking the position of the remaining bait, control points are used to mark the position of the remaining bait in the image, and the position coordinates of the remaining bait are saved in JSON format.

[0024] Unlabeled sample pool Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training.

[0025] Mark sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

[0026] like Figures 3 to 5 As shown, the multi-task learning model It includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and the residual bait counting network are connected to the feature sharing network respectively. Among them, the feature sharing network designs the corresponding network structure according to different application scenarios; for example, lightweight networks such as MobileNet and ShuffleNet can be selected to reduce model parameters. The feeding intensity classification network includes: the first convolution layer, the second convolution layer, the first maximum pooling layer, the third convolution layer, the fourth convolution layer, the second maximum pooling layer, the first average pooling layer (AvgPool) and the fully connected layer (FC) connected in sequence; among them, the convolution kernel size of the first convolution layer, the second convolution layer, the third convolution layer, and the fourth convolution layer is 3*3, and the step size is 1. Through the first maximum pooling layer and the second maximum pooling layer, the feature map output by the feature sharing network can be downsampled twice, and the feeding intensity category (Active or NotActive) of each map can be output through the average pooling operation. The residual bait counting network includes: 5 fifth convolution layers connected in sequence, and the convolution kernel size of the fifth convolution layer is 3 3, the expansion rate d is 2; wherein, the dilated convolution can effectively count the residual baits in the dense area, and the final output density map is summed to obtain the number of residual baits in the entire image.

[0027] As an implementation of an embodiment of the present invention, in step S2, the loss function of training the multi-task active model includes: The binary cross entropy loss is used as the regression loss for feeding intensity classification, i.e., in, is the number of samples in the training set, is the predicted classification category, is the probability of belonging to this category.

[0028] The regression output of the residual bait counting network is the number of residual baits in each image, and the loss function used is the mean absolute error loss, that is, in, is the number of test images, is the actual number of residual bait in the image, is the amount of residual bait predicted by the model.

[0029] The loss function for training a multi-task active learning model is as follows: in, and Respectively represent the variance of the classification loss and regression loss following their respective distributions, is the regularization term.

[0030] As an implementation method of the present invention, in step S3, the sample to be labeled in the iteration process is determined by the query strategy function selected by the unlabeled sample; wherein the query strategy function includes: classification query strategy and regression query strategy , the classification query strategy uses the category information entropy to calculate the uncertainty of the feeding intensity classification of each image ; The regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the residual bait count for each sample image ; Determine the total uncertainty score of each image by weighted summation, and determine the samples to be marked in the iteration process according to the total uncertainty score.

[0031] Furthermore, the multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by the current round of training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by the current round of training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence; wherein the number of neurons in the first linear layer and the second linear layer are 1024 and 512 respectively, and the neuron deletion ratios of the first Dropout layer and the second Dropout layer are 0.3 and 0.15 respectively.

[0032] Furthermore, the classification uncertainty of the feeding intensity of each image is calculated using the category information entropy: , the information entropy of each category output by the multi-task model is: represents the classification uncertainty, and They respectively represent the conditional probabilities that the image belongs to the feeding intensity of "Not Active" and "Active". Since there are only two categories, the conditional probabilities of the two are 1. The closer the conditional probabilities of the two are, the worse the classification performance of the model for the image, and the greater the amount of information expressed, that is, the higher the uncertainty score.

[0033] Furthermore, the approximate Bayesian inference network is used to calculate the regression uncertainty of the residual bait count of each sample image. Specifically, in the deep Bayesian neural network, the parameters of the model have a prior probability distribution, which can be used to quantify the prediction uncertainty in the regression task. The sampling method, that is, Monte Carlo Dropout regularization, is used to predict the uncertainty of regression. In the actual experimental process, the regression uncertainty is determined by calculating the variance of each output result in the inference stage. The formula is as follows: in, represents the regression uncertainty, Represents the regression result of each inference output, is the average value of the output results in the inference phase. To use the sampling-based Monte Carlo sampling times, The final regression uncertainty is obtained by taking the average of Monte Carlo samples.

[0034] By weighting the above classification uncertainty and regression uncertainty, the final total uncertainty score for each image is obtained ,Right now: in, is a weight factor used to balance the uncertainty of the two tasks.

[0035] The training process of the multi-task active learning framework implemented in this paper is as follows: randomly select Image samples are used as initial samples and are manually annotated by experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The initial model is used to perform sample inference in the unlabeled sample pool, and the uncertainty score of the feeding intensity classification and the regression uncertainty score of the residual bait count of each image sample in the unlabeled sample pool are calculated respectively. The total uncertainty score is calculated in a weighted manner, and the top uncertainty score with the highest uncertainty score is selected. The image samples are taken as samples to be labeled, and continue to be labeled by experts. After the labeling is completed, they are put into the labeled sample pool. The model is trained based on the initial model. The above operations are repeated until the threshold of the evaluation index is reached or the maximum number of iterations is reached. The algorithm ends and the final training model is obtained. Here, the number of initial sample selections is , the number of samples selected in subsequent iterations ; The detailed process is shown in Algorithm 1.

[0036] As an implementation method of the present invention, in step S4, different feeding strategies are formulated according to the feeding intensity of the fish school and the density level of the residual bait. There are four types of feeding strategies: continuous feeding, feeding stagnation, feeding suspension and feeding end. The following is a detailed introduction to these four feeding strategies: 1. Continuous feeding: The fish are very active in feeding, most of them actively come out of the water to look for bait, and the amount of bait in the pond is small, which cannot meet the feeding needs of the fish; 2. Feeding stagnation: The fish are actively feeding, but the bait in the pond can meet the feeding needs of the fish. Most fish only eat the bait nearby and will not actively look for bait. At this time, stop feeding in the short term; 3. Feeding suspension: The fish school is not very active in feeding. The fish school basically does not actively eat the nearby bait, and feeding is stopped for a period of time; 4. End of feeding: The fish are inactive in feeding, most of them will not eat the bait, and there are still a lot of leftover bait in the pond. At this time, feeding is stopped and no more feeding is done.

[0037] Embodiment 2: The embodiment of the present invention also provides a fish feeding device based on a multi-task active learning framework, comprising: The first processing module is used to construct a fish feeding behavior data set and a multi-task active learning framework; wherein the multi-task active learning framework includes: a labeling expert, an unlabeled sample pool , mark sample pool , multi-task learning models and multi-task reasoning networks; The second processing module is used to randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The multi-task learning model includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and residual bait counting network are connected to the feature sharing network respectively. The third processing module is used to perform sample reasoning on the initial model in the unlabeled sample pool through the multi-task reasoning network, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression of each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top model with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained based on the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; The fourth processing module is used to input the real-time video of fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

[0038] As an implementation method of the present invention, the unlabeled sample pool Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training; labeled sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

[0039] As an implementation mode of an embodiment of the present invention, the third processing device determines the samples to be marked in the iteration process through the query strategy function of unlabeled sample selection; wherein the query strategy function includes: classification query strategy and regression query strategy, the classification query strategy uses category information entropy to calculate the classification uncertainty of feeding intensity of each image; the regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the leftover bait count of each sample image; the total uncertainty score of each image is determined by weighted summation, and the samples to be marked in the iteration process are determined according to the total uncertainty score.

[0040] As an implementation method of an embodiment of the present invention, the multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by the current round of training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by the current round of training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence.

[0041] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A fish feeding method based on a multi-task active learning framework, characterized in that: include: Step S1, constructing a fish feeding behavior dataset and a multi-task active learning framework; in; The multi-task active learning framework includes: labeling experts, unlabeled sample pool , mark sample pool , multi-task learning models and multi-task reasoning networks; Step S2: Randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The multi-task learning model includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and residual bait counting network are connected to the feature sharing network respectively. Step S3: Use the multi-task reasoning network to perform sample reasoning on the initial model in the unlabeled sample pool, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression for each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top uncertainty score with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained based on the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; Step S4: input the real-time video of the fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

2. The fish feeding method based on a multi-task active learning framework as claimed in claim 1, characterized in that: Unlabeled sample pool Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training; labeled sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

3. The fish feeding method based on a multi-task active learning framework as claimed in claim 2, characterized in that: In step S3, the query strategy function for unlabeled sample selection is used to determine the samples to be labeled in the iteration process; wherein, the query strategy function includes: classification query strategy and regression query strategy, the classification query strategy uses category information entropy to calculate the classification uncertainty of feeding intensity for each image; the regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the residual bait count for each sample image; the total uncertainty score of each image is determined by weighted summation, and the samples to be labeled in the iteration process are determined according to the total uncertainty score.

4. The fish feeding method based on a multi-task active learning framework as claimed in claim 3, characterized in that: The multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by current round training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by current round training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence.

5. A fish feeding device based on a multi-task active learning framework, characterized in that: include: The first processing module is used to construct a fish feeding behavior dataset and a multi-task active learning framework; in; The multi-task active learning framework includes: labeling experts, unlabeled sample pool , mark sample pool , multi-task learning models and multi-task reasoning networks; The second processing module is used to randomly select from the unlabeled sample pool Image samples are used as initial samples and annotated by annotation experts. After annotation, they are put into the labeled sample pool. The multi-task learning model is initially trained based on the annotated initial samples to obtain the initial model. The multi-task learning model includes: feature sharing network, feeding intensity classification network, and residual bait counting network. The feeding intensity classification network and residual bait counting network are connected to the feature sharing network respectively. The third processing module is used to perform sample reasoning on the initial model in the unlabeled sample pool through the multi-task reasoning network, calculate the uncertainty score of the feeding intensity classification and the uncertainty score of the residual bait count regression of each image sample in the unlabeled sample pool, and calculate the total uncertainty score in a weighted manner, and select the top model with the highest total uncertainty score. Image samples are taken as samples to be labeled, and continue to be labeled by experts. After labeling, they are put into the labeled sample pool. The model is iteratively trained based on the initial model until the threshold of the evaluation index is reached or the maximum number of iterations is reached, and the final multi-task learning model is obtained; The fourth processing module is used to input the real-time video of fish feeding behavior into the final multi-task learning model to predict the fish feeding intensity and the amount of residual bait, and output the corresponding feeding strategy.

6. The fish feeding device based on a multi-task active learning framework as claimed in claim 5, characterized in that: Unlabeled sample pool Contains: all initial fish feeding behavior image samples and the remaining fish feeding behavior image samples after each iteration of the multi-task learning model training; labeled sample pool It is used to store sample datasets of fish feeding behavior images that have been annotated by experts and annotated samples that are continuously updated through iterations during the training process of the multi-task learning model.

7. The fish feeding device based on a multi-task active learning framework as claimed in claim 6, characterized in that: The third processing device determines the samples to be marked in the iteration process through the query strategy function of unlabeled sample selection; wherein the query strategy function includes: classification query strategy and regression query strategy, the classification query strategy uses category information entropy to calculate the classification uncertainty of feeding intensity of each image; the regression query strategy uses an approximate Bayesian inference network to calculate the regression uncertainty of the residual bait count of each sample image; the total uncertainty score of each image is determined by weighted summation, and the samples to be marked in the iteration process are determined according to the total uncertainty score.

8. The fish feeding device based on a multi-task active learning framework as claimed in claim 7, characterized in that: The multi-task reasoning network includes: an approximate Bayesian network and a multi-task learning model obtained by current round training; the approximate Bayesian network is placed after the residual bait counting network in the multi-task learning model obtained by current round training; the approximate Bayesian reasoning network includes a second average pooling layer, a first Dropout layer, a first linear layer, a second Dropout layer, and a second linear layer connected in sequence.

Images