Fish feeding intensity intelligent evaluation and dynamic feeding system and method

Abstract

The invention discloses a fish ingestion intensity intelligent evaluation and dynamic feeding system and method, and the method comprises the steps: obtaining a fish shoal ingestion image in real time through video monitoring, extracting a key frame through a Hash difference threshold algorithm, and carrying out the multi-dimensional data enhancement; a three-level fusion model fusing CNN local features, CoordAtt attention features and Transform global features is constructed, and lightweight and high-precision optimization of the model is realized in combination with a transfer learning strategy; according to the feeding intensity grade output by the model in real time, dynamic and accurate feeding control is carried out in combination with the fish biomass, feed waste is effectively reduced, the environmental adaptability of the system is improved, and precision and intelligentization of fish culture are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aquaculture, and in particular to a fish feeding intensity intelligent evaluation and dynamic feeding system and method. BACKGROUND

[0002] In the field of aquaculture, accurate evaluation of fish feeding intensity is an important prerequisite for precise feeding. The existing technologies mainly include manual observation method, simple sensor monitoring method and traditional machine vision technology. Among them, the manual observation method mainly relies on the visual observation of fish feeding behavior by the aquaculture personnel, and determines the feeding amount according to the subjective experience of the fish feeding activity; the simple sensor monitoring method indirectly infers the feeding state of the fish group by detecting water quality parameters (such as dissolved oxygen, water temperature, etc.); the traditional machine vision technology collects fish activity video through fixed camera device, and implements simple classification after manual annotation training samples. There are still the following problems:

[0003] Insufficient evaluation accuracy: the manual observation method relies on subjective experience to judge, and it is difficult to realize the quantification of feeding intensity, and it is difficult to effectively capture the dynamic changes of the fish feeding state in real time; the traditional machine vision technology relies on manual feature extraction, and it is difficult to effectively adapt to complex breeding environments such as light changes and water surface reflections, and the classification accuracy is generally lower than 80%.

[0004] Weak real-time response capability: the existing technologies generally adopt offline analysis mode, and there is a delay of several hours between data collection and feeding decision, which cannot timely respond to the real-time changes of the fish feeding demand, resulting in that the feeding strategy is difficult to accurately match the actual feeding situation.

[0005] Low model lightness level: the mainstream deep learning model has a large parameter size, which is difficult to directly deploy on the edge computing device in the breeding field, and cannot meet the demand of real-time on-site calculation.

[0006] Poor environmental adaptability: the breeding water body is significantly affected by weather and water quality changes, and the feature extraction performance of the existing model significantly decreases under complex environmental conditions such as turbid water and strong light reflection, resulting in a high misjudgment rate of the feeding intensity, and further causing about 30% of the feed waste and water pollution risk.

[0007] Therefore, the present application provides a fish feeding intensity intelligent evaluation and dynamic feeding system and method. SUMMARY

[0008] The present application aims to provide a fish feeding intensity intelligent evaluation and dynamic feeding system and method, which solves the problems of low evaluation accuracy, poor real-time performance, insufficient environmental adaptability and difficulty in light-weight deployment of the existing fish feeding intensity evaluation.

[0009] In a first aspect, the present application provides a fish feeding intensity intelligent evaluation and dynamic feeding system, comprising a data acquisition module, a data preprocessing module, a feature extraction module, a model optimization module, a feeding control module and a dynamic adjustment module, and data transmission between each module:

[0010] The data acquisition module uses a video monitoring device to capture fish feeding behavior video in real time.

[0011] The data preprocessing module extracts key frame images from the video based on the hash difference threshold algorithm, and obtains enhanced images through random cropping, random rotation, brightness adjustment and Gaussian blur processing.

[0012] The feature extraction module constructs a three-level fusion model containing CNN local features, CoordAtt attention features and Transformer global features, and inputs the enhanced images into the model to obtain first feature maps and second feature maps.

[0013] The model optimization module initializes the three-level fusion model to obtain a preliminary model by freezing training with ImageNet pre-training parameters, and optimizes the preliminary model based on the first feature maps and the second feature maps by unfreezing training to obtain an optimized model.

[0014] The feeding control module uses the optimized model to output fish feeding intensity grades in real time, calculates the initial feeding amount combined with the biomass of the fish in the culture pond, and dynamically adjusts the feeding amount by updating the enhanced images to feedback the optimized model based on the real-time collected video.

[0015] The dynamic adjustment module dynamically determines the subsequent feeding amount or the stopping feeding opportunity according to the feeding intensity grades output in real time by the optimized model and the cumulative feeding amount.

[0016] As a preferred technical solution of the first aspect of the present application, the hash difference threshold algorithm is used to extract multiple key frame images, which includes:

[0017] Adjacent video frames are used to generate hash difference values.

[0018] Based on the hash difference values, it is determined whether the current video frame is a redundant frame and the redundancy is performed.

[0019] As a preferred technical solution of the first aspect of the present application, the hash difference values are generated by using adjacent video frames, which includes:

[0020] Hash encoding is performed on the video frame image to generate a hash vector.

[0021] The Hamming distance between adjacent hash vectors is calculated, and the Hamming distance is used as the hash difference value.

[0022] As a preferred technical solution of the first aspect of the application, the construction process of the CoordAtt attention feature branch of the three-level feature fusion model comprises:

[0023] The multi-dimensional enhanced image is respectively encoded in the horizontal and vertical directions.

[0024] The convolution is used to fuse the horizontal and vertical direction spatial feature encodings to generate the CoordAtt attention feature map.

[0025] As a preferred technical solution of the first aspect of the application, the spatial feature encoding process comprises:

[0026] The enhanced image is respectively pooled along the horizontal and vertical directions to generate horizontal and vertical feature descriptors;

[0027] The horizontal and vertical feature descriptors are respectively encoded through convolution and a nonlinear activation function.

[0028] As a preferred technical solution of the first aspect of the application, the optimization process of the frozen training stage comprises:

[0029] The weight parameters of the Transformer global feature branch in the three-level feature fusion model are fixed, and only the CNN local feature branch and the CoordAtt attention feature branch are trained;

[0030] The output features of the frozen Transformer global feature branch are used as a supervision signal to guide the optimization of other branch parameters.

[0031] As a preferred technical solution of the first aspect of the application, the optimization process of the unfreezing training stage comprises:

[0032] The model parameters optimized in the frozen training stage are used as initial parameters, and the weight fixing of the Transformer global feature branch is released;

[0033] The weight parameters of all branches of the three-level feature fusion model are jointly optimized through backpropagation.

[0034] As a preferred technical solution of the first aspect of the application, the processing process of the fish feeding intensity evaluation model for real-time output of the feeding intensity level comprises:

[0035] The first feature map and the second feature map output by the model are respectively globally pooled to obtain a first pooled feature vector and a second pooled feature vector;

[0036] The first pooled feature vector and the second pooled feature vector are spliced and input into a fully connected classification layer to output the feeding intensity level.

[0037] As a preferred technical solution of the first aspect of the application, the process of dynamically determining the feeding amount of the next round or the stopping feeding time point comprises:

[0038] Statistically counting the feeding intensity levels of the continuous multiple frames of images to obtain the level occurrence proportion;

[0039] Based on the level occurrence proportion, the specific value of the feeding amount of the next round or the condition of stopping feeding is objectively determined in combination with the accumulated feeding amount.

[0040] In the second aspect, the application provides a fish feeding intensity intelligent evaluation and dynamic feeding method, which is based on the implementation of the first aspect and comprises the following steps:

[0041] S101: Real-time shooting of fish feeding behavior video by using a video monitoring device;

[0042] S102: Extracting key frame images from the video based on the hash difference threshold algorithm, and obtaining enhanced images through random cropping, random rotation, brightness adjustment and Gaussian blur processing;

[0043] S103: Constructing a three-level fusion model containing CNN local features, CoordAtt attention features and Transformer global features, and inputting the enhanced images into the model to obtain first feature maps and second feature maps;

[0044] S104: Freezing training to initialize the three-level fusion model with ImageNet pre-training parameters to obtain a preliminary model, and then based on the first feature maps and the second feature maps, the preliminary model is optimized by unfreezing training to obtain an optimized model;

[0045] S105: Real-time output of fish feeding intensity levels by using the optimized model, calculation of initial feeding amount in combination with the biomass of fish in the culture pond, and dynamic adjustment of the feeding amount by updating the enhanced images to feed back the optimized model according to the real-time collected video;

[0046] S106: Dynamically determining the subsequent feeding amount or the stopping feeding time point according to the feeding intensity levels output by the optimized model in real time and the accumulated feeding amount.

[0047] In the above technical solution, the application provides the following technical effects and advantages:

[0048] The multi-dimensional enhanced image generated through data enhancement processing is interacted with the three-level feature fusion model, and the feature maps are formed through joint action; then, the model parameter is optimized in stages and nonlinearly by combining the frozen and unfrozen training of two different feature map results, so that a more representative fish feeding intensity evaluation model is obtained; the dynamic adjustment logic of subsequent feedback feeding can rely on the results of the previous model, and the data is continuously updated to re-enter the model, forming a cyclic closed loop of model optimization and feeding strategy adjustment, and further cross-combining the feeding amount decision conditions (grade appearance proportion and cumulative feeding amount) to improve the aquaculture feeding efficiency and economic benefit. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0050] Figure 1 It is a fish feeding intensity intelligent evaluation and dynamic feeding system framework of the present application;

[0051] Figure 2 It is a fish feeding intensity intelligent evaluation and dynamic feeding method flow chart of the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present application more clear, the technical scheme in the embodiments of the present application will be described in more detail below with reference to the drawings in the embodiments of the present application.

[0053] In the drawings, the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The described embodiments are part of the embodiments of the present application, not all of the embodiments. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The embodiments of the present application will be described in detail below with reference to the drawings.

[0054] Embodiment 1

[0055] Please refer to Figure 1 As shown in the figure, the present embodiment provides a fish feeding intensity intelligent evaluation and dynamic feeding system, which comprises a data acquisition module, a data preprocessing module, a feature extraction module, a model optimization module, a feeding control module and a dynamic adjustment module, and each module transmits data between each other;

[0056] a data acquisition module for acquiring fish feeding behavior video and realizing real-time monitoring of the entire pool range;

[0057] Specifically, a Hikvision waterproof camera (DS-2CD3T86FWDV2-I3S) is installed 4 meters above the breeding pool to shoot fish feeding video in real time at a resolution of 3840x2160 and a frame rate of 20 frames / second, with the lens vertically downward covering the entire pool area.

[0058] A data preprocessing module extracts multiple key frame images from the fish feeding behavior video based on the hash difference threshold algorithm, and obtains multi-dimensional enhanced images through random cropping, random rotation, brightness adjustment and Gaussian blur data enhancement processing.

[0059] Specifically, based on the hash difference threshold algorithm, key frames are extracted from the video stream at a frequency of 5 frames / second, and redundant images are removed through hash value comparison, compressing the data size to 20% of the original data size.

[0060] Data enhancement strategy: comprehensive multi-dimensional data enhancement methods such as random cropping (scale 0.8-1.0), rotation (±15°), brightness / contrast adjustment (±20%) and Gaussian blur (3x3 kernel) are used to simulate changes in lighting and viewing angle, and to improve the generalization performance of the model.

[0061] A feature extraction module constructs a three-level feature fusion model, inputs multi-dimensional enhanced images into the three-level feature fusion model, and obtains first and second feature maps representing fish feeding intensity.

[0062] The three-level feature fusion model adopts an improved MobileViT-CoordAtt model architecture, and the three-level feature fusion model includes a CNN local feature branch, a CoordAtt attention feature branch and a Transformer global feature branch. Through the improved MobileViT-CoordAtt model and multi-scale feature fusion strategy, the feature extraction problem under complex background conditions is effectively overcome, and the accuracy of fish feeding intensity evaluation reaches 97.18%, which is 17.18% higher than that of traditional methods, realizing precise quantification of fish feeding behavior.

[0063] Specifically, MobileViT is used as the base network, the local feature extraction capability of CNN and the global modeling capability of Transformer are organically integrated, and parameter lightweight (only 4.09MB) is realized, meeting the real-time inference requirements of edge computing. The model architecture parameter amount is only 4.09MB, the data processing delay is less than 300ms, and it can be directly deployed on edge computing devices to realize second-level response from data monitoring to precise feeding decision, effectively meeting the real-time control requirements of aquaculture sites.

[0064] The CoordAtt attention module is introduced in the global representation branch of the MobileViT network, and the capture ability of the spatial distribution characteristics (such as the degree of aggregation and the motion trajectory) of the fish school is improved through the encoding of the coordinate information in the horizontal and vertical directions. Through the introduction of the CoordAtt attention mechanism, the model proposed by the present application can still maintain stable feature extraction performance in complex environments such as strong light reflection on the water surface, wave disturbance and turbid water quality, and the generalization ability is significantly improved, and the adaptability to the actual breeding environment is stronger.

[0065] A three-level feature fusion strategy is designed, and the input features, local features and global features are fused through Concat operation to enhance the multi-scale expression ability of the model to individual motion details and group behavior patterns.

[0066] The model optimization module is based on transfer learning, and the ImageNet pre-training parameters are used to initialize the three-level feature fusion model in the training stage to obtain a preliminary feeding intensity evaluation model; and based on the first feature map and the second feature map, the preliminary feeding intensity evaluation model is further optimized in a unfreezing training manner to obtain an optimized fish school feeding intensity evaluation model; the transfer learning method is adopted, which can quickly adapt to different fish varieties (such as sea bass and grass carp) under the condition of limited labeled data, and can support further integration of water quality sensors (such as dissolved oxygen and pH) to realize the coordinated regulation of “feeding intensity-water quality parameters” double factors, and greatly improve the adaptability and scientificity of decision-making of the system under different breeding conditions.

[0067] Specifically, the transfer learning training method comprises:

[0068] The ImageNet pre-training model is used for transfer learning fine-tuning.

[0069] On the self-made fish feeding intensity dataset (three categories of strong, medium and weak, a total of 4365 samples), a two-stage optimization strategy of frozen training and unfreezing training is adopted.

[0070] After 200 iterations of optimization, the test set accuracy reaches 97.18%, and the F1-score is more than 97%.

[0071] The feeding control module outputs the fish school feeding intensity level in real time based on the optimized fish school feeding intensity evaluation model, calculates the initial feeding amount combined with the biomass of fish in the breeding pond; the fish school feeding behavior video is collected in real time after feeding with the initial feeding amount, the multi-dimensional enhanced image is updated and fed back to the optimized fish school feeding intensity evaluation model, so as to dynamically adjust the subsequent feeding amount.

[0072] It needs to be explained: equipped with real-time inference engine, receiving pre-processed image data, through the improved MobileViT-CoordAtt model real-time output fish feeding intensity level (strong, medium, weak).

[0073] Combined with the fish biomass of the culture pond, the initial feeding amount (F=Biomass x rate) is calculated, and the feeding strategy is dynamically adjusted according to the real-time evaluation results.

[0074] The dynamic adjustment module dynamically determines the feeding amount of the next round or the timing of stopping feeding according to the cumulative value of the feeding intensity level output by the optimized fish feeding intensity evaluation model and the initial feeding amount in the subsequent feeding process.

[0075] Specifically, a multi-round dynamic feeding strategy is adopted, and the first round feeding amount is 1 / 3 of the maximum feeding amount; 10 frames of images are collected every 2 seconds for real-time evaluation after 5 seconds of the first round feeding; if the proportion of "strong" level is ≥60%, the next round feeding amount continues to maintain 1 / 3 of the maximum feeding amount; if the proportion of "medium" level is ≥60%, the next round feeding amount is adjusted to 1 / 6 of the maximum feeding amount; if the proportion of "weak" level is ≥60% or the cumulative feeding amount reaches the preset maximum value, subsequent feeding is stopped.

[0076] Through the above modular design, real-time and accurate evaluation of fish feeding behavior is realized, and the feeding strategy is dynamically adjusted combined with the feedback mechanism, which greatly improves the efficiency and economic benefit of aquaculture feeding.

[0077] Working principle: after the system is started, the video monitoring device continuously collects fish feeding video, the pre-processing unit extracts key frames and performs data enhancement, and sends them to the central control machine; the improved MobileViT-CoordAtt model analyzes the image in real time, captures the spatial features such as fish aggregation degree and water splash intensity through the CoordAtt module, extracts multi-level feeding related features combined with multi-scale fusion strategy, and evaluates the feeding intensity level; the central control decision system calculates the dynamic feeding strategy according to the evaluation results and biomass, and controls the feeding equipment to feed according to the principle of "regional priority and on-demand adjustment". During the feeding process, the model continuously monitors the feeding state and adjusts the subsequent feeding amount in real time, forming a "monitoring-evaluation-feeding-feedback" closed loop control. Through real-time monitoring and feedback control mechanism, the feeding strategy is dynamically adjusted based on the fish feeding intensity, compared with the traditional feeding method, the feed waste is reduced by 25%-30%, effectively reducing the risk of water pollution and saving the cost of aquaculture.

[0078] The key points of this embodiment include improving the MobileViT-CoordAtt model architecture: embedding the CoordAtt attention module in the MobileViT network and adopting a three-level multi-scale feature fusion strategy model design; video data deduplication and enhancement algorithm based on hash difference threshold: combining hash value comparison and multi-dimensional data enhancement preprocessing method; dynamic feeding strategy based on biomass and feeding intensity: integrating biomass calculation and multi-round real-time evaluation of dynamic adjustment of feeding amount. Other supplements: the system can be extended to integrate water quality sensors (such as dissolved oxygen, pH, and water temperature sensors), further optimize the feeding strategy through multi-source data fusion, realize the collaborative regulation of "feeding intensity-water quality parameters", and improve the adaptability and scientificity of feeding decision-making of the system under different environmental conditions.

[0079] Embodiment 2

[0080] As shown in Figure 2 , the part not detailed in this embodiment is as described in embodiment 1, this embodiment provides a fish feeding intensity intelligent assessment and dynamic feeding method, including the following steps:

[0081] S101: using a video monitoring device to real-time shoot fish school feeding behavior video in a breeding pond;

[0082] S102: extracting multiple key frame images from the fish school feeding behavior video based on a hash difference threshold algorithm, and obtaining multi-dimensional enhanced images through data enhancement processing of random cropping, random rotation, brightness adjustment and Gaussian blur;

[0083] Specifically, the hash difference threshold algorithm extracts multiple key frame images, including:

[0084] Generating a hash difference value using adjacent video frames;

[0085] Determining whether the current video frame is a redundant frame based on the hash difference value and performing deduplication.

[0086] Further, the hash difference value is generated using adjacent video frames, including:

[0087] Hash encoding the video frame image to generate a hash vector;

[0088] Calculating the Hamming distance between adjacent hash vectors, and taking the Hamming distance as the hash difference value.

[0089] S103: constructing a three-level feature fusion model, the three-level feature fusion model includes a CNN local feature branch, a CoordAtt attention feature branch and a Transformer global feature branch; inputting the multi-dimensional enhanced images into the three-level feature fusion model to obtain a first feature map and a second feature map representing the fish school feeding intensity;

[0090] Specifically, the construction process of the CoordAtt attention feature branch of the three-level feature fusion model includes:

[0091] The multi-dimensional enhanced images are respectively subjected to spatial feature encoding in horizontal and vertical directions.

[0092] Convolution is used to fuse the horizontal and vertical direction spatial feature encodings to generate a CoordAtt attention feature map.

[0093] Further, the spatial feature encoding process includes:

[0094] The enhanced images are respectively subjected to pooling operations along horizontal and vertical directions to respectively generate horizontal and vertical feature descriptors.

[0095] The horizontal and vertical feature descriptors are subjected to convolution and a nonlinear activation function to generate corresponding spatial feature encodings.

[0096] S104: Based on transfer learning, the three-level feature fusion model is initialized with ImageNet pre-training parameters through a frozen training stage to obtain a preliminary feeding intensity evaluation model; and based on the first feature map and the second feature map, the preliminary feeding intensity evaluation model is further optimized in an unfreezing training manner to obtain an optimized fish school feeding intensity evaluation model.

[0097] It should be noted that the optimization process of the frozen training stage includes:

[0098] The weight parameters of the Transformer global feature branch in the three-level feature fusion model are fixed, and only the CNN local feature branch and the CoordAtt attention feature branch are trained.

[0099] The output features of the frozen Transformer global feature branch are used as a supervision signal to guide the optimization of other branch parameters.

[0100] Further, the optimization process of the unfreezing training stage includes:

[0101] The model parameters optimized in the frozen training stage are used as initial parameters, and the weight fixing of the Transformer global feature branch is released.

[0102] The weight parameters of all branches of the three-level feature fusion model are jointly optimized through backpropagation.

[0103] S105: Based on the optimized fish school feeding intensity evaluation model, the fish school feeding intensity level is output in real time, and the initial feeding amount is calculated combined with the biomass of fish in the culture pond; after feeding with the initial feeding amount, the fish school feeding behavior video is collected in real time, the multi-dimensional enhanced image is updated and fed back to the optimized fish school feeding intensity evaluation model, so as to dynamically adjust the subsequent feeding amount.

[0104] Specifically, the process of outputting the feeding intensity level in real time according to the optimized fish school feeding intensity evaluation model comprises:

[0105] The first feature map and the second feature map output by the model are respectively globally pooled to obtain a first pooled feature vector and a second pooled feature vector;

[0106] The first pooled feature vector and the second pooled feature vector are spliced and input into a fully connected classification layer to output the feeding intensity level.

[0107] S106: In the subsequent feeding process, the feeding intensity level output in real time according to the optimized fish school feeding intensity evaluation model and the cumulative value of the initial feeding amount are used to dynamically determine the feeding amount of the next round of feeding or the timing of stopping feeding.

[0108] Specifically, the process of dynamically determining the feeding amount of the next round of feeding or the timing of stopping feeding comprises:

[0109] The feeding intensity levels of the continuous multiple frames of images are counted to obtain a level occurrence ratio;

[0110] Based on the level occurrence ratio, the specific value of the feeding amount of the next round of feeding or the condition of stopping feeding is objectively determined in combination with the cumulative feeding amount.

[0111] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A fish feeding intensity intelligent assessment and dynamic feeding system, characterized in that: It includes data acquisition module, data preprocessing module, feature extraction module, model optimization module, feeding control module and dynamic adjustment module, and data transmission is carried out between each module: The data acquisition module uses a video monitoring device to capture the fish feeding behavior video in real time; The data preprocessing module extracts key frame images from the video based on the hash difference threshold algorithm, and obtains enhanced images through random cropping, random rotation, brightness adjustment and Gaussian blur processing; The feature extraction module builds a three-level fusion model including CNN local features, CoordAtt attention features, and Transformer global features, and inputs the enhanced image into the model to obtain the first feature map and the second feature map; The model optimization module initializes the three-level fusion model with ImageNet pre-training parameters through frozen training to obtain a preliminary model, and then optimizes the preliminary model through unfrozen training based on the first feature map and the second feature map to obtain an optimized model; The feeding control module uses an optimization model to output the feeding intensity level of the fish school in real time, calculates the initial feeding amount based on the fish biomass in the aquaculture pond, and dynamically adjusts the feeding amount based on the enhanced image feedback optimization model updated by real-time video acquisition; The dynamic adjustment module dynamically determines the subsequent feeding amount or the timing to stop feeding based on the feeding intensity level and cumulative feeding amount output by the optimization model in real time.

2. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 1, characterized in that: The method of extracting multiple key frame images based on a hash difference threshold algorithm includes: Adjacent video frames are used to generate hash difference values; Based on the hash difference value, it is determined whether the current video frame is a redundant frame and deduplication is performed.

3. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 2, characterized in that: The step of generating a hash difference value using adjacent video frames includes: Perform hash encoding on the video frame image to generate a hash vector; Calculate the Hamming distance between adjacent hash vectors and use the Hamming distance as the hash difference value.

4. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 1, characterized in that: The construction process of the CoordAtt attention feature branch of the three-level feature fusion model includes: Perform spatial feature encoding on the multi-dimensional enhanced image in the horizontal and vertical directions respectively; Convolution is used to fuse horizontal and vertical spatial feature encoding to generate the CoordAtt attention feature map.

5. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 4, characterized in that: The spatial feature encoding process includes: Perform pooling operations on the enhanced image in the horizontal and vertical directions to generate horizontal and vertical feature descriptors respectively; The horizontal and vertical feature descriptors are converted into corresponding spatial feature codes through convolution and nonlinear activation functions.

6. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 1, characterized in that: The optimization process of the frozen training phase includes: The Transformer global feature branch weight parameters in the three-level feature fusion model are fixed, and only the CNN local feature branch and the CoordAtt attention feature branch are trained; The output features of the frozen Transformer global feature branch are used as supervision signals to guide the optimization of other branch parameters.

7. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 6, characterized in that: The optimization process of the unfreeze training phase includes: The model parameters optimized during the freezing training phase are used as the initial parameters to release the weight fixation of the Transformer global feature branch. The weight parameters of all branches of the three-level feature fusion model are jointly optimized through back propagation.

8. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 1, characterized in that: The process of outputting the feeding intensity level in real time according to the optimized fish feeding intensity assessment model includes: Perform global pooling on the first feature map and the second feature map output by the model to obtain a first pooled feature vector and a second pooled feature vector; The first pooled feature vector and the second pooled feature vector are concatenated and input into the fully connected classification layer to output the feeding intensity level.

9. The fish feeding intensity intelligent assessment and dynamic feeding system according to claim 8, characterized in that: The process of dynamically determining the next round of feeding amount or the timing of stopping feeding includes: Count the feeding intensity levels of multiple consecutive frames of images and obtain the occurrence ratio of each level; Based on the proportion of grade occurrence and the accumulated feeding amount, the specific value of the next round of feeding amount or the conditions for stopping feeding can be objectively determined.

10. A method for intelligently assessing and dynamically feeding fish feeding intensity, based on the implementation of the intelligently assessing and dynamically feeding system for fish feeding intensity according to any one of claims 1 to 9, characterized in that: The following steps are involved: S101: Using a video monitoring device to shoot a video of the fish feeding behavior in real time; S102: extracting key frame images from the video based on a hash difference threshold algorithm, and obtaining enhanced images through random cropping, random rotation, brightness adjustment, and Gaussian blur processing; S103: Construct a three-level fusion model including CNN local features, CoordAtt attention features, and Transformer global features, and input the enhanced image into the model to obtain the first feature map and the second feature map; S104: Initializing the three-level fusion model with ImageNet pre-training parameters through frozen training to obtain a preliminary model, and then optimizing the preliminary model through unfrozen training based on the first feature map and the second feature map to obtain an optimized model; S105: Using the optimization model to output the fish feeding intensity level in real time, the initial feeding amount is calculated based on the fish biomass in the aquaculture pond, and the enhanced image feedback optimization model is updated based on the real-time video collection to dynamically adjust the feeding amount; S106: Dynamically determine the subsequent feeding amount or the timing to stop feeding based on the feeding intensity level and the cumulative feeding amount output in real time by the optimization model.

Images