A shrimp and crab fry feeding decision method and system based on bait response characteristics

Abstract

The application discloses a shrimp and crab fry feeding decision method and system based on bait response characteristics, and relates to the technical field of machine learning. The shrimp and crab fry feeding decision method based on bait response characteristics acquires spatial bait response image sequences corresponding to each spatial response region, performs continuous behavior response analysis, constructs a spatiotemporal bait response feature sequence, inputs the spatiotemporal bait response feature sequence into a pre-trained spatiotemporal process evolution perception network, generates bait response process hidden state features through joint coding of spatial interaction and time dependence, performs matching interaction through a group of learnable bait response evolution primitives, generates a bait response evolution vector, and outputs a feeding state probability distribution through mapping processing. The application generates a feeding decision through the feeding state probability distribution, thereby accurately representing the group bait response behavior trajectory of the entire process from feeding to satiation, and further improving the decision accuracy.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a method and system for making feeding decisions for shrimp and crab seedlings based on feed response characteristics. Background Technology

[0002] In the industrialized cultivation of shrimp and crab seedlings, especially in the high-density seedling cultivation stage, the feeding process has a direct impact on the survival and growth of the seedlings. Under the stimulation of feed, the seedlings in the culture pond will exhibit a series of behavioral responses such as feeding, gathering, and competing for food. These behaviors are visually manifested as changes in the group's movement state, involving multiple dimensions such as movement intensity, spatial distribution, and directional trends. Industrial cameras installed above the seedling pond can continuously acquire overhead video streams to record the behavioral changes of the seedling group before and after feeding.

[0003] Several technical solutions have emerged for automated decision-making regarding the feeding of shrimp and crab larvae. These solutions typically determine feeding needs by analyzing the movement information or distribution status of the population in the rearing pond, and then establish a correlation between the extracted behavioral indicators and feeding strategies. Common practices include performing target detection and motion analysis on the acquired images to obtain statistics on the population's activity level, and then adjusting the feeding amount according to preset rules or empirical thresholds. Another solution uses deep learning methods to directly learn feature representations related to the feeding status from the original images and map them into feeding decision instructions.

[0004] The limitations of existing technologies include at least the following problems: In the scenario of high-density seedling cultivation in industrialized settings, the behavioral response of seedling populations in the breeding ponds to feed stimuli is captured by cameras as a continuous image sequence. Existing models typically extract frames within a fixed time window and use statistical features as input, compressing the feeding response process, which has a dynamic evolutionary pattern, into a static feature expression. This results in the population gradually declining from the feeding peak to the satiated remnant and the response states corresponding to non-feeding disturbances caused by water flow or molting during the feeding interval, which have different evolutionary sources but similar instantaneous characteristics. This leads to aliasing of representations in the feature space, making it difficult for existing technologies to stably match the actual needs of high-density seedling cultivation for the timing of feed delivery. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a feeding decision-making method and system for shrimp and crab seedlings based on feed response characteristics, which solves the problem that existing technologies have difficulty distinguishing between satiety remnants and non-feeding disturbances, leading to inaccurate feeding decisions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a shrimp and crab fry feeding decision-making method based on feed response features, comprising the following steps: dividing the image acquisition range corresponding to the shrimp and crab fry rearing pond into several spatial response regions; continuously acquiring feed response image sequences during the shrimp and crab fry feeding process, and obtaining spatial feed response image sequences corresponding to each spatial response region; performing continuous behavioral response analysis on the spatial feed response image sequences, extracting feed response features, and constructing a spatiotemporal feed response feature sequence; inputting the spatiotemporal feed response feature sequence into a pre-trained spatiotemporal process evolution perception network; performing joint encoding of spatial interaction and temporal dependence on the spatiotemporal feed response feature sequence to generate latent state features of the feed response process; matching and interacting the latent state features of the feed response process with a set of learnable feed response evolution primitives to generate a feed response evolution vector; mapping the feed response evolution vector to output a feeding state probability distribution; and generating a feeding decision based on the feeding state probability distribution.

[0007] Further, the steps for dividing the image acquisition range into several spatial response regions and obtaining the spatial bait response image sequence corresponding to each spatial response region are as follows: obtain the bait placement position within the image acquisition range, and set several spatial response regions within the image acquisition range based on the bait placement position as the region division benchmark; record the spatial response region positional relationship within the image acquisition range for each spatial response region; and perform region cropping on the bait response image sequence according to the region boundary in the spatial response region positional relationship to obtain the spatial bait response image sequence corresponding to each spatial response region.

[0008] Further, the specific steps for extracting bait response features are as follows: The spatial bait response image sequence is slidably cropped according to a set window length to obtain several sliding window image groups; inter-frame motion estimation and seedling response region segmentation statistics are performed on each sliding window image group to obtain the population motion vector field and population response density distribution; dynamic evolution analysis is performed on the population motion vector field and population response density distribution to extract a set of change features, including motion state change features, aggregation state change features, diffusion state change features, and direction change features; the change feature sets of the same spatial response region within the same sliding window are arranged to form bait response features.

[0009] Furthermore, the specific steps for constructing the spatiotemporal bait response feature sequence are as follows: Based on the temporal position of each sliding window in the spatial bait response image sequence, the bait response features of the same spatial response region within different sliding windows are temporally sorted; for sliding windows with the same temporal position, the bait response features are spatially sorted to form spatial response feature groups corresponding to the sliding windows; based on the spatial relationship of the spatial response regions, the regional adjacency identifiers of each spatial response region are determined and written into the bait response features respectively; the spatial response feature groups corresponding to different sliding windows are continuously arranged to form the spatiotemporal bait response feature sequence.

[0010] Furthermore, the steps for jointly encoding the spatial interaction and temporal dependence of the spatiotemporal bait response feature sequence are as follows: Feature mapping is performed on the bait response features in the spatiotemporal bait response feature sequence to obtain spatial response feature vectors; spatial association weights are calculated based on region adjacency identifiers, and the spatial response feature vectors are weighted and aggregated to obtain spatial interaction feature vectors; the spatial interaction feature vectors are used for temporal evolution modeling through a gating mechanism to generate historical evolution features corresponding to each sliding window; the historical evolution features are arranged and fused to generate latent state features of the bait response process corresponding to each sliding window.

[0011] Further, the specific steps for generating the bait response evolution vector are as follows: Calculate the matching response values ​​between the latent state features of the bait response process and each bait response evolution primitive, and normalize them to obtain the evolution response weights; according to the evolution response weights, weightedly combine a set of learnable bait response evolution primitives to obtain primitive aggregation features; calculate the weight change between the evolution response weights and the evolution response weights corresponding to adjacent sliding windows, and generate evolutionary transfer features based on the weight change; fuse the latent state features of the bait response process, primitive aggregation features, and evolutionary transfer features to obtain the bait response evolution vector.

[0012] Furthermore, the pre-training steps of the spatiotemporal process evolution perception network are as follows: acquire historical bait response image sequences labeled with feeding states, and generate historical spatiotemporal bait response feature sequences; output the corresponding historical feeding states based on the historical bait response evolution vector sequences; determine the classification loss based on the difference between the historical feeding states and the feeding state labels; analyze the evolutionary continuity relationship between the bait response evolution vectors corresponding to adjacent sliding windows, and determine the evolutionary continuity loss; analyze the state change relationship between the feeding states corresponding to adjacent sliding windows to determine the state, and convert the state transition constraint loss; construct an initial joint optimization objective based on the classification loss, evolutionary continuity loss, and state transition constraint loss, and perform evolutionary primitive expansion processing to form a comprehensive joint optimization objective, and perform joint iterative optimization until the preset training termination condition is met, thus obtaining the pre-trained spatiotemporal process evolution perception network.

[0013] Furthermore, the specific steps of the evolutionary primitive expansion process are as follows: Before training begins, a predetermined number of initial bait response evolutionary primitives are set, and each initial bait response evolutionary primitive is set as a trainable parameter; during each forward propagation, the evolutionary response weights are calculated respectively; the feature similarity between each initial bait response evolutionary primitive is extracted, and the primitive separation constraint is calculated; the primitive response concentration constraint is calculated based on the distribution of the evolutionary response weights; the primitive evolution constraint is calculated based on the relationship between the changes in the evolutionary response weights and the corresponding feeding state labels; the primitive separation constraint, the primitive response concentration constraint, and the primitive evolution constraint are added to the joint optimization objective to form a comprehensive joint optimization objective.

[0014] Furthermore, the feeding state probability distribution includes the probability of a feeding response enhancement state, the probability of a feeding response maintenance state, and the probability of a feeding response weakening state. The specific steps for generating feeding decisions are as follows: Time-weighted summation is performed on the various feeding state probabilities in the feeding state probability distributions corresponding to the most recent multiple sliding windows to obtain a cumulative set of feeding states; the comprehensive trend of the weakening feeding response state is calculated; the cumulative set of feeding states, the comprehensive trend of the weakening feeding response state, and the feeding state probability distribution are compared with preset feeding conditions, and the corresponding feeding decisions are generated based on the comparison results.

[0015] A feeding decision-making system for shrimp and crab larvae based on feed response features includes: a data acquisition unit for acquiring spatial feed response image sequences corresponding to each spatial response region; a feed response feature construction unit for constructing spatiotemporal feed response feature sequences; a spatiotemporal process evolution perception unit, including a spatiotemporal coding module for generating latent state features of the feed response process; a process evolution module for generating feed response evolution vectors; a state classification module for outputting the probability distribution of feeding states; and a feeding decision-making unit for generating feeding decisions.

[0016] The present invention has the following beneficial effects: (1) The shrimp and crab seedling feeding decision method based on bait response features constructs a spatiotemporal bait response feature sequence and performs joint encoding of spatial interaction and temporal dependence. This allows the dynamic evolution of the seedling population’s behavioral response to bait stimulation. After dividing the spatial response region, the change features of each region are extracted independently. Spatial relationships are preserved through adjacency identification. Then, the model is built along the time progression. The hidden state features of each sliding window not only integrate the current performance of each region, but also carry the evolution memory of multiple previous moments. This fully represents the population’s bait response behavior trajectory from feeding to satiation, thereby improving the accuracy of decision-making.

[0017] (2) The shrimp and crab seedling feeding decision method based on bait response features sets learnable bait response evolution primitives and matches them with latent state features. This enables the automatic discovery and differentiation of the response patterns of the population at different stages under bait stimulation. In training, primitive separation constraints keep primitives distinct, response concentration constraints prompt each window to match a few primitives, and evolution constraints guide the weight changes of adjacent windows to be consistent with the feeding state label. This allows each primitive to correspond to the change patterns of different response stages after training convergence. The bait response evolution vector generated during inference simultaneously contains the current state, matching pattern, and migration trend, thus enabling accurate assessment of the feeding state.

[0018] (3) The feeding decision method for shrimp and crab seedlings based on feed response features introduces evolutionary continuity loss and state transition constraint loss during the training process. This allows the process representation learned by the network to maintain a smooth transition within the same feeding stage and to follow the natural evolution law of group feeding behavior between different stages. Furthermore, the feed response evolution vectors of adjacent sliding windows are constrained to be close to each other when the state remains unchanged, avoiding representation jumps caused by water flow fluctuations or individual differences. When the state changes, the vectors are guided to migrate in a reasonable direction to prevent representation paths that violate the actual law, such as jumping directly from increased feeding to decreased feeding. This improves the stability of feeding state discrimination.

[0019] (4) The shrimp and crab seedling feeding decision system based on feed response features acquires spatial feed response image sequences corresponding to each region through the data acquisition unit, the feed response feature construction unit performs continuous behavior analysis on the images of each region and constructs spatiotemporal feed response feature sequences, the spatiotemporal coding module in the spatiotemporal process evolution perception unit performs joint encoding of spatial interaction and time dependence on the feature sequences to generate latent state features, the process evolution module converts the latent state features into evolution vectors through evolution primitive matching, the state classification module maps the evolution vectors into feeding state probability distributions, and finally the feeding decision unit generates feeding decisions based on the probability distributions, thereby improving the reliability of feeding decisions.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a shrimp and crab seedling feeding decision-making method based on feed response characteristics according to the present invention.

[0022] Figure 2 This is a block diagram of a shrimp and crab seedling feeding decision system based on feed response characteristics according to the present invention. Detailed Implementation

[0023] Please see Figure 1The present invention provides a technical solution: a shrimp and crab seedling feeding decision method based on feed response characteristics, comprising the following steps: dividing the image acquisition range corresponding to the shrimp and crab seedling breeding pond into several spatial response regions, continuously acquiring feed response image sequences of shrimp and crab seedlings in the image acquisition range during the feeding process, showing changes in behavior response caused by feed stimulation, and obtaining spatial feed response image sequences corresponding to each spatial response region from the feed response image sequences; Continuous behavioral response analysis was performed on the spatial bait response image sequences corresponding to each spatial response region. The bait response features of each spatial response region within each sliding window were extracted, and a spatiotemporal bait response feature sequence was constructed. The spatiotemporal bait response feature sequence is input into a pre-trained spatiotemporal process evolution perception network, which includes a spatiotemporal encoder, a process evolution attention layer, and a state classifier. In the spatiotemporal encoder, the spatiotemporal bait response feature sequence is jointly encoded with spatial interaction and temporal dependence to generate the latent state features of the bait response process corresponding to each sliding window. In the process evolution attention layer, the latent state features of the bait response process are matched and interacted with a set of learnable bait response evolution primitives. The bait response evolution vectors corresponding to each sliding window are generated through linear mapping. Each bait response evolution primitive represents a change pattern in the response trajectory of the population under bait stimulation. In the state classifier, the bait response evolution vector is mapped and processed to output the feeding state probability distribution corresponding to each sliding window; Feeding decisions are generated based on the probability distribution of feeding states corresponding to each sliding window.

[0024] The specific steps for generating the probability distribution of feeding states are as follows: By performing state mapping on the bait response evolution vector, we obtain state response values ​​representing the states of enhanced feeding response, maintained feeding response, and weakened feeding response, respectively: Slide window The bait response evolution vector is denoted as Its dimensions are , To provide a globally unified time step index, a trainable state mapping matrix and a state bias vector are set in the state classifier to perform a linear mapping on the bait response evolution vector, thereby obtaining the probability values ​​corresponding to the feeding response enhancement state, feeding response maintenance state, and feeding response weakening state. An enhanced feeding response indicates that the population is in a phase of increased or peak feeding intensity; a sustained feeding response indicates that the population is in an active and stable peak feeding period with feeding intensity maintained at a high level; and a weakened feeding response indicates that the population's feeding intensity has transitioned from peak to declining phase. Normalizing the probability values ​​corresponding to the states of enhanced feeding response, maintained feeding response, and weakened feeding response yields the probability distribution of the feeding states corresponding to the sliding window, i.e.: The three state response values ​​are normalized using an exponential normalization function to obtain the probability of enhanced feeding response, the probability of maintained feeding response, and the probability of weakened feeding response. The values ​​of the three are all in the range of [0, 1] and the sum is 1. The feeding state probability distribution completely describes the evolution stage of the group's feeding response within the current sliding window.

[0025] Feed response image sequence refers to a top-down video stream continuously captured by an industrial camera from above the seedling pond at a fixed frame rate, with each frame completely covering the image acquisition range of the pond; The spatial response zone is a series of continuously distributed sub-regions centered on the bait placement point, used to capture the differentiated responses of groups to bait stimuli at different spatial locations. The spatiotemporal bait response feature sequence is a feature sequence formed by organizing the bait response features of each spatial response region within each sliding window according to the spatial arrangement relationship and the temporal sequence. This sequence retains both spatial adjacency information and temporal evolution information. The spatiotemporal process evolution perception network consists of three parts: a spatiotemporal encoder, a process evolution attention layer, and a state classifier. The spatiotemporal encoder is used to jointly encode the spatial interaction and temporal dependence of the input sequence. The process evolution attention layer is used to transform the latent state features into process evolution representations through evolutionary primitive matching. The state classifier is used to output the feeding state probability distribution based on the process evolution representations. The bait response evolution primitive is a set of learnable parameter vectors in the process evolution attention layer. Each primitive is automatically differentiated during training to represent a potential response change pattern of the population under bait stimulation, without establishing a pre-established one-to-one correspondence with the state of enhanced, maintained or weakened feeding. The bait response evolution vector is a fixed-dimensional vector output by the process evolution attention layer, which condenses the evolutionary process information of the group's behavioral response to bait stimuli within the current window.

[0026] Specifically, the steps for dividing the image acquisition area into several spatial response regions and obtaining the spatial bait response image sequence corresponding to each spatial response region from the bait response image sequence are as follows: The bait placement location of the feeding device within the image acquisition range is obtained. Using the bait placement location as the region division benchmark, several continuously distributed spatial response regions are set within the image acquisition range, namely: The reference coordinates of the bait placement location within the image acquisition range are determined by the image coordinates of the installation location of the baiting device or the bait entry point. Using the reference coordinates as the center, the image acquisition range is divided into multiple concentric rings (e.g., 5) along the radial direction, and each ring is then divided into multiple fan-shaped regions (e.g., 8) along the circumferential direction, resulting in a total of several spatial response regions. Each spatial response region has a unique region identifier with a value range of 1 to N. It also records the region boundaries, region locations, and adjacency relationships between each spatial response region and adjacent spatial response regions within the image acquisition range, forming the spatial response region positional relationships, i.e.: For each spatial response region, record the closed region whose boundary is determined by the radial start and end radii and the circumferential start and end angles, and the region position is the image coordinates of the centroid of that region; The adjacency relationship is defined as follows: if two spatial response regions are adjacent in the radial direction, that is, located in the same circumferential position of adjacent rings, or adjacent in the circumferential direction, that is, located in adjacent circumferential positions of the same ring, then the two regions are determined to be adjacent spatial response regions. Construct an adjacency matrix to record the adjacency relationships between regions. A matrix element of 1 indicates that two different regions in the corresponding row and column are adjacent, while a matrix element of 0 indicates that they are not adjacent or are the same region. The spatial response region position relationship includes the region boundary, region position, and adjacency matrix of each region. Based on the regional boundaries in the spatial response region positional relationship, each frame of the bait response image sequence is cropped to obtain the spatial bait response image corresponding to each spatial response region in each frame. Then, according to the acquisition time sequence of the bait response image sequence, the spatial bait response images belonging to the same spatial response region are arranged to form the spatial bait response image sequence corresponding to each spatial response region. For each frame of the bait response image sequence, based on the boundary of each spatial response region, all pixels in that region are extracted using a masking method to form the spatial bait response image corresponding to each spatial response region in that frame. The image position in each spatial bait response image retains its global image coordinates in the original bait response image. After performing the same operation on all frames in the bait response image sequence, the spatial bait response images belonging to the same spatial response region are arranged in ascending order of frame number to form a spatial bait response image sequence corresponding to each spatial response region. The sequence length is the same as the total number of frames in the original bait response image sequence.

[0027] The specific steps for extracting bait response features are as follows: According to the set window length and window movement step size, the spatial bait response image sequence corresponding to each spatial response region is slidably cropped to obtain sliding window image groups corresponding to each spatial response region at different time positions. Each sliding window image group includes multiple frames of spatial bait response images that are consecutive in time, that is: Window length The value ranges from 30 frames to 120 frames, with 60 frames being the preferred setting, and the window movement step size. The value range is from 1 / 6 to 1 / 2 of the window length, and it is preferably set to 15 frames; For each spatial response region, starting from the first frame of the spatial bait response image sequence, a continuous sequence is extracted. The frame space bait response images form the first sliding window image group; Then move forward along the timeline. Frame, extract the next group of consecutive frames Frames, forming the second sliding window image group; And so on, to obtain the corresponding spatial response regions at different time locations. The first sliding window image group, the first spatial response region of each... Each sliding window is strictly aligned in time, collectively defining a globally unified time step. If the remaining number of frames is less than the window length If the insufficient portion at the end is not formed, a new sliding window image group will be discarded. Inter-frame motion estimation and seedling response region segmentation statistics are performed on adjacent spatial bait response images within each sliding window image group to obtain the group motion vector field between adjacent spatial bait response images. The group motion vector field includes the motion direction and motion amplitude corresponding to each image position, as well as the seedling response region corresponding to each frame of spatial bait response image. Based on the response pixel distribution of each image position within the seedling response region, the group response density distribution is obtained, i.e.: Brightness normalization and water surface disturbance suppression processing were performed on each frame of spatial bait response image to eliminate high-frequency brightness changes caused by water surface ripples; A device area mask is generated based on the fixed area of ​​the feeding device, and a bait particle mask is generated based on the bright or high-contrast particle area within a preset time range after the bait enters the water. During the background subtraction process, the pixels covered by the above two masks are excluded. The bright or high-contrast particle area refers to the connected region whose gray value is higher than the average gray value of the background plus 3 times the standard deviation, or the connected region whose local contrast is greater than the preset contrast threshold. After the above processing, effective pixels that undergo continuous changes relative to the background reference image before feeding (maintaining the change in images for no less than a preset number of consecutive frames) and whose change amplitude exceeds a preset response threshold are marked as response pixels; the response threshold is determined based on the background grayscale standard deviation of the corresponding pixel position, and is preferably set to 2.5 to 3.5 times the background grayscale standard deviation; During the feeding process, the background model is updated according to the image positions that are not marked as response pixels according to the preset update cycle, and the set of all response pixels constitutes the seedling response area; A binary region mask is generated based on the region boundary of each spatial response region to limit the effective pixel range of the region. For the spatial bait response images of two adjacent frames within the sliding window, the dense optical flow estimation algorithm is used to calculate the motion vectors at each effective image position. Only the motion vectors that are both inside the region mask in the two adjacent frames are retained. Pixels outside the region boundary and those less than the preset boundary width are not included in the statistics. The first Frame and the The group motion vector field between frames is defined as the first... The group motion vector field of the frame, where the motion vector contains two attributes: motion direction and motion amplitude; The population response density distribution is defined as the response pixel density value at each image location within the seedling response region. This density value is calculated using a region centered at that location with a radius of [missing information]. The ratio of the number of responding pixels within a local window to the total number of pixels in the window is calculated, and the radius is... The preferred value is 1 / 40 to 1 / 20 of the length of the shorter side of the image; Dynamic evolution analysis is performed on the group motion vector field and the group response density distribution to extract the change feature set within the corresponding sliding window, including motion state change features, aggregation state change features, diffusion state change features, and direction change features, i.e.: For the group motion vector field of each frame within the sliding window, binning statistics are performed according to motion direction and motion amplitude to construct motion direction histogram and motion amplitude histogram; Motion state change characteristics: Calculate the statistical distance between the motion amplitude histograms of each adjacent frame within the window to obtain a statistical distance sequence. Calculate the mean and standard deviation of this sequence, which together serve as the motion state change characteristics. The statistical distance can be the Bach distance. Direction change feature: Calculate the statistical distance between the motion direction histograms of each adjacent frame within the window to obtain a statistical distance sequence, and calculate the mean and standard deviation of the sequence, which together serve as the direction change feature; Clustering state change characteristics: For the population response density distribution of each frame within the sliding window, pixels with response pixel density exceeding the density threshold are marked as high-density response pixels. The density threshold is determined based on the statistical value of response pixel density in the corresponding spatial response area in the historical feeding samples under the same breeding pond and the same seedling size conditions, and is preferably set to 1.2 to 1.8 times the historical average. The total number of all high-density response pixels in each frame is calculated as the area. And calculate the average response density of these high-density response pixels. Then calculate their first-order difference sequences respectively. and ;calculate The mean and standard deviation, and The mean and standard deviation of the data are concatenated into a fixed-dimensional vector, which serves as a feature of cluster state changes. Diffusion state change characteristics: For the seedling response area of ​​each frame within the sliding window, calculate the minimum and maximum distances between all response pixel positions and the bait placement positions, and form the minimum distance sequence and the maximum distance sequence respectively; The mean and standard deviation of the difference sequence between adjacent frames of the minimum distance sequence are used as the inner edge distance variation feature to characterize the edge movement trend of the group closer to the feeding center. The mean and standard deviation of the difference sequence between adjacent frames of the maximum distance sequence are used as features of the outer edge distance variation to characterize the edge diffusion trend of the population on the side away from the feeding center. The inner edge distance variation feature and the outer edge distance variation feature are concatenated into a fixed-dimensional vector, which is used as the diffusion state variation feature. Arrange the set of changing features of the same spatial response region within the same sliding window to form the bait response features of the corresponding spatial response region within the sliding window.

[0028] Specifically, the steps for constructing the spatiotemporal bait response feature sequence are as follows: Based on the temporal position of each sliding window in the spatial bait response image sequence, the bait response features of the same spatial response region within different sliding windows are sorted temporally, i.e.: For each spatial response region, its bait response feature vectors within different sliding windows are processed according to a globally unified time step. Arranged in ascending order, this forms a temporal characteristic sequence of the spatial response region; For sliding windows with the same time position, the bait response features corresponding to each spatial response region are spatially sorted to form spatial response feature groups for the corresponding sliding windows, i.e.: For each time step According to the regional identifier order of the spatial response regions, the bait response feature vectors of each spatial response region are arranged into spatial response feature groups for that time step. Based on the spatial relationship of the response regions, the regional adjacency identifiers of each spatial response region are determined, and these regional adjacency identifiers are written into the bait response features of the corresponding spatial response regions, i.e.: The adjacency matrix is ​​obtained from the spatial response region positional relationships. For each spatial response region, the corresponding row vector in this matrix is ​​extracted as its region adjacency identifier. This identifier is a vector with a length equal to the total number of spatial response regions. The binary vector indicates the adjacency relationship between this region and other spatial response regions, where the adjacency identifier with itself is always 0; The region adjacency identifier is appended to the bait response feature vector of that region to form an extended feature vector with adjacency identifier; According to the time sequence of each sliding window, the spatial response feature groups corresponding to different sliding windows are continuously arranged to form a spatiotemporal bait response feature sequence (that is, the spatial response feature groups of each time step are arranged in ascending order of t). This ensures that each time position in the spatiotemporal bait response feature sequence contains the bait response features of each spatial response region and the corresponding region adjacency identifier. The spatiotemporal bait response feature sequence completely preserves the spatial distribution relationship and temporal evolution trajectory of the group's response to bait stimuli.

[0029] The steps for jointly encoding the spatial interaction and temporal dependence of the spatiotemporal bait response feature sequence are as follows: Feature mapping is performed on the bait response features of each spatial response region within the same sliding window in the spatiotemporal bait response feature sequence to obtain the spatial response feature vector, i.e.: For each time step For each spatial response region within the region, its extended feature vector is linearly mapped using a trainable feature mapping matrix, thus mapping the feature dimension from the original dimension to the hidden state dimension. Obtain the spatial response feature vector and the hidden state dimension. The preferred setting is 128 to 512; Based on the regional adjacency identifiers of each spatial response region, the spatial association weights between the current spatial response region and its neighboring spatial response regions are calculated. Then, the spatial response feature vectors of the neighboring spatial response regions are weighted and aggregated according to these spatial association weights to obtain the spatial interaction feature vector of the current spatial response region. Extract the region adjacency identifier from the extended feature vector to determine the set of neighboring regions of the current spatial response region. Since the corresponding bit of the region adjacency identifier is 0, the set of neighboring regions only includes other spatial response regions that are adjacent to the current region, and does not include itself. For each neighboring region, calculate the spatial association weight between the current region and the neighboring region. After concatenating the spatial response feature vectors of the current region and the neighboring region, process them through linear mapping and activation function, and then perform an inner product with the trainable attention vector to obtain the association score. The correlation scores of all adjacent regions are normalized to obtain the spatial correlation weights of each adjacent region. The spatial response feature vectors of each adjacent region are weighted and aggregated according to spatial correlation weights, and then summed element by element with the spatial response feature vector of the current region itself to obtain the spatial interaction feature vector of the current spatial response region in the time step. Spatial interaction feature vectors are modeled temporally using a gating mechanism to generate historical evolution features of each spatial response region within each sliding window, i.e.: Set a learnable initial historical evolution feature vector with dimension . The initial values ​​were randomly sampled from a normal distribution with a mean of zero and a standard deviation of 0.01. For each time step The current spatial interaction feature vector is compared with the historical evolution features of the previous time step (when...). At that time, the historical evolution features of the previous time step (which are the initial historical evolution features) are fused through a gating mechanism to generate the historical evolution features of the current time step; The gating mechanism adopts a gated loop unit structure, and its specific parameter configuration is as follows: after concatenating the current spatial interaction feature vector with the historical evolution features of the previous time step, the updated gate vector is obtained by activating it through a trainable first linear mapping matrix and a logistic function, and the reset gate vector is obtained by activating it through a trainable second linear mapping matrix and a logistic function. The first linear mapping matrix and the second linear mapping matrix are independent trainable parameter matrices. The historical evolution features of the previous time step are multiplied element-wise using the reset gate vector. The result is concatenated with the current spatial interaction feature vector and then activated by a trainable third linear mapping matrix and hyperbolic tangent to obtain candidate historical evolution features. Using the updated gate vector as the weight, the historical evolution features and candidate historical evolution features of the previous time step are weighted and combined to obtain the historical evolution features of the current time step. By iterating step by step, the historical evolution features of each spatial response region at each time step are obtained. The historical evolution features corresponding to each spatial response region within the same sliding window are arranged and merged according to their regional location to generate the latent state features of the bait response process for each sliding window, namely: For each time step The historical evolution features of each spatial response region are concatenated into a long vector according to the region identifier order, and linearly mapped through a trainable fusion matrix to obtain the hidden state features of the bait response process corresponding to that time step. This latent state feature integrates the state information of each spatial response region after its evolution in the time dimension, and fully characterizes the spatiotemporal response process of the population to bait stimuli.

[0030] The specific steps for generating the bait response evolution vector corresponding to each sliding window are as follows: The matching response values ​​between the latent state features of the bait response process and each bait response evolutionary primitive are calculated separately, and each matching response value is normalized to obtain the evolutionary response weight of the sliding window corresponding to each bait response evolutionary primitive, that is: Let a set of learnable bait response evolution primitives be denoted as Each dimension is vector, number of primitives The preferred setting is 8 to 32; For each time step The hidden state characteristics of the bait response process are used to calculate the cosine value of the angle between the bait and each primitive as the matching response value. The M matching response values ​​are subjected to exponential normalization to obtain the evolutionary response weights of each evolutionary motif at this time step. The sum of all weights is 1. The evolutionary response weights represent the degree of matching between the latent state characteristics of the bait response process at the current time step and the potential response change patterns represented by each evolutionary motif. Based on the evolutionary response weights, a set of learnable bait response evolution primitives are weighted and combined to obtain the primitive aggregation feature corresponding to the sliding window, i.e.: Using the evolution response weight as the weighting coefficient, the evolutionary primitives of each bait response are weighted and summed to obtain the primitive aggregation feature of this time step. This feature integrates all potential response change pattern information matched by the population response trajectory within the current time step. Calculate the change in evolutionary response weights between the sliding window and the adjacent time-bound sliding windows, and generate evolutionary migration features characterizing the direction and degree of change in the bait response process based on the weight changes. Set the current time step The evolution response weight vector and the previous time step The evolutionary response weight vectors are subtracted element-wise, and the resulting difference vector is the evolutionary transfer feature, whose dimension and number of primitives are related. same; For the first time step, the evolutionary transfer feature is set as a zero vector. Positive elements in the evolutionary transfer feature indicate that the matching degree of the corresponding evolutionary primitive increases between adjacent time steps, while negative elements indicate a decrease. The sign and magnitude of each dimension together describe the evolutionary direction and the magnitude of change. By fusing the latent state features, primitive aggregation features, and evolutionary transfer features of the bait response process corresponding to the sliding window, the bait response evolution vector corresponding to the sliding window is obtained, i.e.: The latent state features, primitive aggregation features, and evolutionary transfer features of the bait response process are concatenated into a long vector, which is then mapped using a trainable linear mapping matrix and a bias vector to obtain a vector with dimension [missing information]. The bait response evolution vector, The preferred setting is 128 to 256; The bait response evolution vector condenses the latent state information of the group's behavioral response to bait stimuli in the current time step, the comprehensive pattern information matching each evolutionary motif, and the evolutionary migration information between adjacent time steps.

[0031] Specifically, the pre-training steps of the spatiotemporal process evolution perception network are as follows: Obtain historical bait response image sequences with feeding status annotations, and generate historical spatiotemporal bait response feature sequences, i.e.: Historical feed response image sequences corresponding to different feeding batches in multiple breeding cycles were collected. Breeding professionals labeled the feeding status of each sliding window based on the group's behavior. The feeding status label included three categories: enhanced feeding response, maintained feeding response, and weakened feeding response. The feeding state of each sliding window is marked with the group feeding state corresponding to its last frame, that is, marked with the state category at the end of the window; Following the steps described above, extract the historical spatiotemporal bait response feature sequence and the corresponding feeding state annotations; The historical feeding state is output based on the historical bait response evolution vector sequence, i.e.: The historical spatiotemporal bait response feature sequences labeled with feeding states are used as training samples and input into the spatiotemporal process evolution perception network to be trained. During the forward propagation of each training iteration, based on the spatiotemporal encoder parameters, process evolution attention layer parameters, and bait response evolution primitive parameters corresponding to the current iteration, the hidden state features of the historical bait response process, the historical bait response evolution vector, and the historical feeding state probability distribution are generated sequentially. The classification loss is determined based on the difference between historical feeding status and the feeding status labels, i.e.: Classification loss The cross-entropy loss function is used to calculate and compare the probability distribution of historical feeding states with the corresponding feeding state labels. The deviation between the predicted distribution and the true label is measured in the form of negative log-likelihood. The evolutionary continuity relationship between the bait response evolution vectors corresponding to adjacent sliding windows in time is analyzed, and the evolutionary continuity loss is determined, i.e.: The evolutionary continuity loss applies continuity constraints only to time-adjacent sliding windows with the same feeding state label, thus avoiding suppression of true state transitions; Set a sliding window and The corresponding bait response evolution vectors are respectively and Feeding status is marked as and Evolutionary continuous loss The calculation formula is: ; in, The total number of sliding windows in the same feeding event. This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The Euclidean norm of a vector. To prevent small constants with zero denominators, this loss promotes a smooth transition in evolutionary representations under the same conditions; The state is determined by analyzing the state change relationship between feeding states corresponding to adjacent sliding windows in time, and the constraint loss is transformed accordingly. Define the state transition weight matrix , where matrix elements Indicates a change from feeding state Switch to feeding state The degree of reasonableness, These correspond to the states of enhanced feeding response, maintained feeding response, and weakened feeding response, respectively. The state transition weight matrix is ​​set according to the feeding behavior pattern of shrimp and crab larvae. Preferably, the transition weight from enhancement to maintenance is set to 0.9, the transition weight from maintenance to weakening is set to 0.9, and the transition weight from weakening to maintenance is set to 0.8; the transition weight from enhancement to weakening is set to 0.2, and the transition weight from weakening to enhancement is set to 0.2; the transition weight for maintaining the same state is set to 1.0. The values ​​of each element in the state transition weight matrix are empirically set based on prior knowledge of the feeding behavior of shrimp and crab larvae. Their purpose is to provide a reasonable initial bias for model training, guide the network output to conform to the gradual law of biological feeding process, and avoid the prediction result from a direct jump from the state of enhanced feeding response to the state of weakened feeding response. The above weight values ​​can also be fine-tuned according to the actual data distribution during the training process. Set a sliding window and The corresponding feeding state probability distributions are as follows: and State transition constraint loss The calculation formula is: ; in, Represents a sliding window In the first The probability of a feeding state. Represents a sliding window In the first The probability of a feeding state; An initial joint optimization objective is constructed based on classification loss, evolutionary continuity loss, and state transition constraint loss. Evolutionary primitives are then expanded to form a comprehensive joint optimization objective. This objective is then iteratively optimized on the spatiotemporal encoder, process evolution attention layer, state classifier, and a set of learnable bait response evolutionary primitives until a preset training termination condition is met, resulting in a pre-trained spatiotemporal process evolution perception network. Joint optimization objective Defined as: ; Among them, the weighting coefficients are preferably set as follows: , , , , , , ; An adaptive moment estimation algorithm (e.g., the Adam optimizer) is used to perform gradient descent optimization on the above joint optimization objective, with the initial learning rate preferably set to... The number of training iterations is preferably set to 500 to 2000. In each training iteration, the gradient of the joint optimization objective with respect to each trainable parameter is calculated by automatic differentiation, and the parameters of the spatiotemporal encoder, process evolution attention layer, state classifier and each bait response evolution primitive are updated synchronously. The preset training termination condition is that the relative change of the joint optimization objective over 50 consecutive iterations is less than 1%. Or, it may reach the maximum number of iterations; After training, different bait response evolution primitives form characteristic representations of different potential response change patterns. The potential response change patterns are not pre-established with a one-to-one correspondence with the feeding response enhancement state, feeding response maintenance state, or feeding response weakening state. Different feeding states can be characterized by different combinations of response weights of multiple bait response evolution primitives.

[0032] The specific steps of evolutionary primitive expansion processing are as follows: Before training begins, a predetermined number of initial bait response evolution primitives are set, and each initial bait response evolution primitive is set as a trainable parameter in the process evolution attention layer, i.e.: Set the number of primitives to , The value range is usually from 8 to 32, and it is preferably set to 16; Each component of the initial bait response evolution primitive has a mean of zero and a standard deviation of... Random sampling is performed in a truncated normal distribution to control parameter scaling and reduce gradient fluctuations in the early stages of training; The differences between the primitives in the feature space are adjusted by subsequent primitive separation constraints; During each forward propagation, the matching response value between the hidden state features of the bait response process corresponding to each sliding window and each initial bait response evolutionary primitive is calculated. Based on the matching response value, the evolutionary response weight of each sliding window corresponding to each initial bait response evolutionary primitive is determined, i.e.: In the During the forward propagation of each training iteration, for each time step The latent state features of the bait response process are analyzed, and the matching response values ​​with each primitive are calculated based on cosine similarity. The results are then normalized to obtain the evolutionary response weight vector. The calculation method is consistent with the steps described above for generating the bait response evolution vector; The feature similarity between each initial bait response evolution primitive is extracted, and the primitive separation constraint is calculated. The primitive separation constraint increases with the feature similarity between any two initial bait response evolution primitives, that is: First, for the first Each primitive vector in each training iteration is L2 normalized to obtain the unit primitive vector. Using the cosine similarity between pairwise unit primitive vectors as a measure of feature similarity between corresponding initial bait response evolution primitives, the cosine similarity of all primitive pairs is calculated to obtain a cosine similarity matrix. The element in the i-th row and j-th column of this matrix is... and Cosine similarity; In this matrix, the diagonal elements represent the similarity between each primitive and itself, which is always 1; the off-diagonal elements represent the similarity between different primitives, which are the objects to which the primitive separation constraint is applied. Element Separation Constraints Defined as all off-diagonal lines with a similarity exceeding a preset threshold. The sum of squares of the parts is calculated using the following formula: ; in, The similarity threshold typically ranges from 0.3 to 0.7, and is preferably set to [value missing]. ; When the cosine similarity between two primitives exceeds a threshold, the primitive separation constraint increases, pushing them away from each other in the feature space through gradient backpropagation. The lumped constraint quantity of the primitive response is calculated based on the distribution of the evolution response weights corresponding to each sliding window, i.e.: Elementary response lumped constraint quantity The entropy calculation based on the evolutionary response weight vector is as follows: ; in, This represents the total number of sliding windows. For the first Sliding window during each training iteration Corresponding to the The evolutionary response weights of each primitive. To prevent the logarithmic function input from being a small constant that is zero, it is preferable to set it to... ; The closer the distribution of evolutionary response weights is to a uniform distribution, the greater the constraint. When the weights are concentrated on the primitives that are highly matched with the current window, the constraint decreases, thereby prompting each window to form a concentrated differentiated response. Based on the average evolutionary response weight of each bait response evolution primitive within the same training batch, the equilibrium constraint amount used by the primitive is calculated, i.e.: For the current training batch, the average value of the evolution response weights of each bait response evolution primitive over all sliding windows is calculated and used as the batch average response weight of each primitive. Calculate the sum of squares of the deviations between the batch average response weight of each primitive and the ideal uniform distribution value (i.e., the reciprocal of the total number of primitives), and use it as the equilibrium constraint for the primitive. The primitive evolution constraint is calculated based on the relationship between the changes in evolutionary response weights corresponding to temporally adjacent sliding windows and the corresponding feeding state labels. Specifically, for temporally adjacent sliding windows with the same feeding state label, the corresponding evolutionary response weights remain continuous; for temporally adjacent sliding windows where the feeding state label changes, the corresponding evolutionary response weights undergo a shift corresponding to the change in feeding state. Primitive Evolution Constraints Using a contrastive loss method, sliding windows with adjacent time intervals are selected from the training batch. Define the Euclidean distance between the evolution response weight vectors of adjacent windows. ; If feeding status is marked Then the evolutionary response weights corresponding to the constraints remain continuous, making the distance... As small as possible; if Then the evolutionary response weights corresponding to the constraints undergo migration in accordance with the changes in feeding state, causing the distance to... At least greater than the preset interval The formula for calculating the primitive evolution constraint is: ; in, and Label the corresponding feeding status. For indicator functions, The preset interval typically ranges from 0.3 to 0.7, and is preferably set to... ; The constraints of primitive separation, primitive response concentration, primitive usage equilibrium, and primitive evolution are added to the joint optimization objective to form a comprehensive joint optimization objective.

[0033] Specifically, the feeding state probability distribution includes the probability of a feeding response enhancement state, the probability of a feeding response maintenance state, and the probability of a feeding response weakening state. The specific steps for generating feeding decisions are as follows: By applying time weights to the probability distributions of various feeding states corresponding to multiple recent sliding windows, a cumulative set of feeding states is obtained (including cumulative values ​​for increased feeding, sustained feeding, and decreased feeding). The sliding window with the closest time position to the current moment has a larger time weight. Let Q be the number of recent sliding windows. The value of Q is usually between 5 and 15, and it is preferred to set it to 15. ; Define a time-weighted sequence that satisfies the constraints that the weight of a timeframe increases with the timeframe closest to the current time and that the sum of all weights is 1. The time weights are set in a linear or exponential increasing manner. In the linear increasing manner, the weight of the q-th nearest window is... ; The probability of enhanced feeding response, maintained feeding response, and weakened feeding response in the most recent Q sliding windows are weighted and summed with time weights to obtain the cumulative value of enhanced feeding response, the cumulative value of maintained feeding response, and the cumulative value of weakened feeding response. These three constitute the cumulative set of feeding states and their sum is 1. And calculate the overall trend of the weakened feeding response state, that is: For each sliding window Calculate the probability of its weakened feeding response state compared with the previous sliding window. The difference in the probability of a weakened feeding response state This represents the weakening trend of the window. To reduce single-window fluctuations, take the values ​​corresponding to the Q most recent sliding windows. The overall trend of the weakening feeding response at the current moment is obtained by weighting the average according to the time weight. A positive overall trend value indicates an increased probability of weakening, while a negative value indicates a decreased probability of weakening. The cumulative set of feeding states, the combined trend of weakened feeding response states, and the probability distribution of feeding states are compared with preset feeding conditions, and corresponding feeding decisions are generated based on the comparison results. As the highest priority, when the probability of a weakened feeding response in any sliding window exceeds a preset forced stop threshold, a stop-feeding decision is immediately generated, independent of any other conditions. The preset forced stop threshold typically ranges from 0.9 to 0.95, preferably set to 0.92. When a stop-feeding decision is generated, bait delivery is stopped. The following multi-level decisions are executed only if the above-mentioned mandatory stopping conditions are not triggered. Multi-level decisions are executed in the order of priority: stop, decrease, increase, maintain. When multiple decision conditions are met simultaneously, only the decision with the highest priority is executed: When the cumulative value of feeding enhancement exceeds a preset enhancement threshold, and the probability of enhanced feeding response shows a non-decreasing trend within a preset number of consecutive sliding windows, and the probability of enhanced feeding response in the most recent sliding window exceeds a preset activation threshold, a decision to increase feeding is generated. The preset enhancement threshold typically ranges from 0.5 to 0.7, with a preferred value of 0.6. The preset number of continuous sliding windows is set to 3 to 5, with a preferred value of 3. A non-decreasing trend refers to each subsequent value in the sequence being no less than the previous value. The preset activation threshold ranges from 0.3 to 0.5, with a preferred value of 0.4. When generating an increased feeding decision, the amount of bait given per unit time is increased by a preset percentage based on the current amount, with a preferred increase percentage ranging from 10% to 30%. When the cumulative value of food intake maintenance is greater than the preset maintenance threshold, and the absolute value of the overall trend of the current weakening food intake response is less than the stability threshold, a maintenance feeding decision is generated. The preset maintenance threshold ranges from 0.5 to 0.7, preferably set to 0.6, and the stability threshold typically ranges from 0.05 to 0.15, preferably set to 0.1. When generating a maintenance feeding decision, the amount of feed given per unit time remains unchanged. When the probability distribution of feeding states in the most recent consecutive sliding windows satisfies the condition that the probability of the feeding response weakens is non-decreasing, and the cumulative value of the weakening feeding is greater than a preset weakening threshold, a decision to reduce feeding is generated. The preset weakening threshold ranges from 0.5 to 0.7, with 0.6 being the preferred setting. A non-decreasing trend means that each subsequent value in the sequence is not less than the previous value. When generating a decision to reduce feeding, the amount of bait given per unit time is reduced by a preset percentage based on the current amount. The preferred range for the reduction percentage is 15% to 35%. When the probability of a weakened feeding response is greater than the preset stopping probability within a preset number of consecutive sliding windows, and the absolute value of the overall trend of the weakened feeding response at the current moment is lower than the stopping response threshold, a stop feeding decision is generated. The preset stop probability ranges from 0.7 to 0.9, preferably 0.8. The preset number of continuous sliding windows is set to 5 to 8, preferably 5. The stop response threshold is used to determine whether the trend of the weakening state has become flat, and its value ranges from 0.02 to 0.08, preferably 0.05. When a stop feeding decision is generated, the feeding is stopped.

[0034] Please see Figure 2 This invention provides a technical solution: a shrimp and crab seedling feeding decision system based on feed response features, comprising: a data acquisition unit, used to divide the image acquisition range corresponding to the shrimp and crab seedling breeding pond into several spatial response regions, continuously acquire feed response image sequences during the shrimp and crab seedling feeding process, and obtain spatial feed response image sequences corresponding to each spatial response region; a feed response feature construction unit, used to perform continuous behavioral response analysis on the spatial feed response image sequences, extract feed response features, and construct a spatiotemporal feed response feature sequence; and a spatiotemporal process evolution perception unit, used to input the spatiotemporal feed response feature sequence into a pre-trained spatiotemporal process evolution perception network. Furthermore, the spatiotemporal process evolution perception unit includes a spatiotemporal coding module, used to jointly encode the spatiotemporal bait response feature sequence with spatial interaction and temporal dependence to generate latent state features of the bait response process; a process evolution module, used to match and interact the latent state features of the bait response process with a set of learnable bait response evolution primitives to generate bait response evolution vectors; a state classification module, used to map the bait response evolution vectors to output the feeding state probability distribution; and a feeding decision unit, used to generate feeding decisions based on the feeding state probability distribution.

[0035] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0036] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for feeding decision-making of shrimp and crab fry based on feed response characteristics, characterized in that, Includes the following steps: The image acquisition range corresponding to the shrimp and crab seedling breeding pond is divided into several spatial response regions. The feed response image sequence during the feeding process of shrimp and crab seedlings is continuously acquired, and the spatial feed response image sequence corresponding to each spatial response region is obtained. Continuous behavioral response analysis was performed on spatial bait response image sequences to extract bait response features and construct a spatiotemporal bait response feature sequence. The spatiotemporal bait response feature sequence is input into a pre-trained spatiotemporal process evolution perception network; The spatiotemporal bait response feature sequence is jointly encoded with spatial interaction and temporal dependence to generate latent state features of the bait response process; The latent state features of the bait response process are matched and interacted with a set of learnable bait response evolution primitives to generate a bait response evolution vector. The bait response evolution vector is mapped and processed to output the probability distribution of feeding states; Feeding decisions are generated based on the probability distribution of feeding states.

2. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 1, characterized in that, The steps for dividing the image acquisition area into several spatial response regions and obtaining the spatial bait response image sequence corresponding to each spatial response region are as follows: The bait placement locations within the image acquisition range are obtained, and several spatial response areas are set within the image acquisition range, using the bait placement locations as the basis for area division. And record the spatial response region positional relationship of each spatial response region within the image acquisition range; Based on the regional boundaries in the spatial response region location relationship, the bait response image sequence is cropped to obtain the spatial bait response image sequence corresponding to each spatial response region.

3. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 2, characterized in that, The specific steps for extracting bait response features are as follows: According to the set window length, the spatial bait response image sequence is slidably cropped to obtain several sliding window image groups; Inter-frame motion estimation and seedling response region segmentation statistics are performed on the image groups of each sliding window to obtain the population motion vector field and population response density distribution; Dynamic evolution analysis is performed on the group motion vector field and the group response density distribution to extract a set of change features, including motion state change features, aggregation state change features, diffusion state change features, and direction change features. Arrange the set of changing features of the same spatial response region within the same sliding window to form bait response features.

4. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 3, characterized in that, The specific steps for constructing the spatiotemporal bait response feature sequence are as follows: Based on the temporal position of each sliding window in the spatial bait response image sequence, the bait response features of the same spatial response region in different sliding windows are sorted temporally. For sliding windows with the same time position, the bait response features are spatially sorted to form spatial response feature groups corresponding to the sliding windows; Based on the spatial relationship of the response regions, the regional adjacency identifiers of each spatial response region are determined and written into the bait response features respectively; The spatial response feature groups corresponding to different sliding windows are arranged continuously to form a spatiotemporal bait response feature sequence.

5. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 4, characterized in that, The steps for jointly encoding the spatial interaction and temporal dependence of the spatiotemporal bait response feature sequence are as follows: Feature mapping is performed on the bait response features in the spatiotemporal bait response feature sequence to obtain the spatial response feature vector; Based on the region adjacency identifier, spatial association weights are calculated, and spatial response feature vectors are weighted and aggregated to obtain spatial interaction feature vectors. The spatial interaction feature vectors are modeled for temporal evolution through a gating mechanism to generate historical evolution features corresponding to each sliding window; By arranging and fusing historical evolutionary features, latent state features of the bait response process corresponding to each sliding window are generated.

6. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 5, characterized in that, The specific steps for generating the bait response evolution vector are as follows: The matching response values ​​between the latent state features of the bait response process and each bait response evolution primitive are calculated and normalized to obtain the evolution response weights. Based on the evolutionary response weights, a set of learnable bait response evolution primitives are weighted and combined to obtain primitive aggregation features; Calculate the weight change between the evolutionary response weight and the evolutionary response weight corresponding to the adjacent sliding window, and generate evolutionary migration features based on the weight change. The latent state features, primitive aggregation features, and evolutionary migration features of the bait response process are fused to obtain the bait response evolution vector.

7. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 1, characterized in that, The pre-training steps for the spatiotemporal process evolution perception network are as follows: Acquire historical bait response image sequences with feeding status annotations and generate historical spatiotemporal bait response feature sequences; Output the corresponding historical feeding state based on the historical bait response evolution vector sequence; Classification loss is determined based on the difference between historical feeding status and the feeding status label; Analyze the evolution continuity relationship between the bait response evolution vectors corresponding to adjacent sliding windows, and determine the evolution continuity loss; Analyze the state change relationship between the feeding states corresponding to adjacent sliding windows to determine the state, and transform the constraint loss; An initial joint optimization objective is constructed based on classification loss, evolutionary continuity loss, and state transition constraint loss. Evolutionary primitives are then expanded to form a comprehensive joint optimization objective. Joint iterative optimization is performed until the preset training termination condition is met, resulting in a pre-trained spatiotemporal process evolution perception network.

8. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 7, characterized in that, The specific steps of evolutionary primitive expansion processing are as follows: Before starting training, a predetermined number of initial bait response evolution primitives are set, and each initial bait response evolution primitive is set as a trainable parameter; The evolution response weights are calculated for each forward propagation process; Extract the feature similarity between each initial bait response evolution primitive and calculate the primitive separation constraint. Calculate the lumped constraint quantity of the primitive response based on the distribution of the evolution response weights; The primitive evolution constraint is calculated based on the relationship between the changes in evolution response weights and the corresponding feeding state labels; The constraints of primitive separation, primitive response concentration, and primitive evolution are added to the joint optimization objective to form a comprehensive joint optimization objective.

9. The shrimp and crab fry feeding decision-making method based on feed response characteristics according to claim 1, characterized in that, The probability distribution of feeding states includes the probability of a state with enhanced feeding response, the probability of a state with maintained feeding response, and the probability of a state with weakened feeding response. The specific steps for generating feeding decisions are as follows: Time-weighted summation is performed on the probabilities of various feeding states in the probability distributions of the feeding states corresponding to the most recent multiple sliding windows to obtain the cumulative set of feeding states; And calculate the overall trend of the weakened feeding response; The cumulative set of feeding states, the combined trend of the weakening feeding response, and the probability distribution of feeding states are compared with the preset feeding conditions, and corresponding feeding decisions are generated based on the comparison results.

10. A shrimp and crab fry feeding decision system based on feed response characteristics, employing the shrimp and crab fry feeding decision method based on feed response characteristics as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire spatial bait response image sequences corresponding to each spatial response region; The bait response feature construction unit is used to construct a spatiotemporal bait response feature sequence; The spatiotemporal process evolution sensing unit includes a spatiotemporal coding module, which is used to generate latent state features of the bait response process; The process evolution module is used to generate bait response evolution vectors; The state classification module is used to output the probability distribution of feeding states; The feeding decision unit is used to generate feeding decisions.

Images