Automatic identification method for andrias davidianus lung breathing behavior

Abstract

The application relates to the technical field of automatic identification of animal behavior, in particular to an automatic identification method for lung breathing behavior of giant salamanders, which comprises the following steps: acquiring an initial video image of a target area; pre-processing the initial video image to obtain a standard video image; performing feature enhancement on the standard video image to obtain an enhanced feature image; performing multi-scale breathing behavior feature extraction on the enhanced feature image to obtain deep feature maps of different scales; performing multi-level feature aggregation on the deep feature maps of different scales to generate an aggregated feature map; and generating an identification result based on the aggregated feature map. The application helps to solve the problems that the artificial monitoring of the breathing behavior of giant salamanders in a low-contrast scene at night is prone to high missing detection rate and long-term efficient and objective monitoring is difficult to realize due to the nocturnal and cave-dwelling habits of the giant salamanders.

Description

Technical Field

[0001] This application relates to the field of automatic identification technology of animal behavior, and in particular to an automatic identification method for the lung breathing behavior of the giant salamander. Background Technology

[0002] The nocturnal and burrowing habits of the giant salamander pose significant challenges to traditional monitoring: low contrast between the salamander and its environment in nighttime videos, coupled with weak respiratory movements, leads to high false negative rates in manual observation or traditional detection methods, making long-term, high-frequency, and continuous monitoring difficult. Early studies focused primarily on daytime lung respiration, neglecting nighttime active behavior. Although recent studies have filled gaps in the study of circadian rhythms, overall research on its respiratory behavior remains scarce.

[0003] Currently, computer vision and deep learning technologies have yielded significant results in the field of aquatic animal monitoring. Algorithms such as the YOLO series, Faster R-CNN, and SSD have been applied to scenarios such as fish disease detection and crab sex identification. Combined with image enhancement technology, they can also alleviate interference from low contrast and blur underwater. However, the application of these technologies in the study of giant salamander behavior is almost non-existent. At present, monitoring of giant salamander respiratory behavior still relies on manual video playback and counting, which is time-consuming, labor-intensive, and susceptible to subjective interference. The data integrity and objectivity are insufficient, which not only hinders in-depth research on its respiratory rhythm and environmental response mechanisms but also limits the application of AI in the protection of rare amphibians. Therefore, there is an urgent need for targeted automatic identification methods to break through the bottleneck. Summary of the Invention

[0004] To help solve the problem of high false negative rates and difficulty in long-term, efficient, and objective monitoring of the breathing behavior of giant salamanders in low-contrast nighttime scenes due to their nocturnal and burrowing habits, this application provides an automatic identification method for the lung breathing behavior of giant salamanders.

[0005] Firstly, this application provides an automatic identification method for the lung breathing behavior of a giant salamander, employing the following technical solution:

[0006] An automatic identification method for the lung breathing behavior of a giant salamander includes:

[0007] Acquire the initial video image of the target area;

[0008] The initial video image is preprocessed to obtain a standard video image;

[0009] The standard video image is enhanced to obtain an enhanced feature image;

[0010] Multi-scale respiratory behavior feature extraction is performed on the enhanced feature image to obtain depth feature maps at different scales;

[0011] Multi-level feature aggregation is performed on the depth feature maps at different scales to generate aggregated feature maps;

[0012] Based on the aggregated feature map, a recognition result is generated.

[0013] Optionally, the preprocessing of the initial video image to obtain a standard video image includes:

[0014] The initial video image is extracted frame by frame and converted into a single-channel grayscale image;

[0015] Obtain the mean and variance of gray levels for different frames of images;

[0016] Determine whether the mean gray level and the variance gray level both meet the corresponding gray level threshold requirements;

[0017] If both the mean gray level and the variance gray level meet the corresponding gray level threshold requirements, then the corresponding frame image is taken as the target video image.

[0018] The target video image is standardized to obtain a standard video image.

[0019] Optionally, the step of performing feature enhancement on the standard video image to obtain an enhanced feature image includes:

[0020] The standard video image is input into the small target restoration and enhancement module, which includes a super-resolution reconstruction sub-network SRCNN and a lightweight dimensionality reduction structure SimpleStem.

[0021] The lightweight dimensionality reduction structure performs spatial compression on the standard video image and outputs global contour features;

[0022] The super-resolution reconstruction subnetwork performs super-resolution reconstruction on the standard video image and outputs detail enhancement features;

[0023] The detailed enhancement features are fused with the global contour features to output an enhanced feature image.

[0024] Optionally, the lightweight dimensionality reduction structure spatially compresses the standard video image and outputs global contour features including:

[0025] Perform a Conv1 convolution operation with kernel size K=3 and stride=3 on the standard video image to obtain the first convolution feature map;

[0026] The first convolutional feature map is batch normalized to obtain the first standardized feature map;

[0027] Perform ReLU activation function operation based on the first standardized feature map to obtain the first activated feature map;

[0028] Perform a Conv2 convolution operation with kernel size K=3 and stride=3 on the first activation feature map to obtain the second convolution feature map;

[0029] Batch normalization is performed on the second convolutional feature map to obtain the second standardized feature map;

[0030] Perform ReLU activation function operation based on the second standardized feature map to obtain the second activation feature map;

[0031] The visual cue fusion module filters key information from the second activated feature map, retaining visual cues related to the giant salamander's outline.

[0032] Channel integration is performed on the filtered feature maps to output global contour features.

[0033] Optionally, the super-resolution reconstruction subnetwork performs super-resolution reconstruction on the standard video image, and outputs detail enhancement features including:

[0034] Conv1 convolution operation was performed on standard video images, with the kernel size set to 64×9×9, padding parameter=4, stride=1, to obtain a preliminary detailed feature map with 64 channels.

[0035] Perform ReLU activation function operation on the preliminary detailed feature map to generate an activated intermediate feature map;

[0036] Based on the activated intermediate feature map, Conv2 convolution operation is performed with the kernel size set to 32×5×6, padding=2 and stride=1 to extract depth detail information and obtain a 32-channel depth detail feature map.

[0037] Perform ReLU activation function operation on the deep detail feature map to improve feature discrimination and generate enhanced detail feature map;

[0038] Based on the enhanced detail feature map, perform Conv3 convolution operation with kernel size set to 32×5×5, padding=2 and stride=1 to obtain high-resolution feature map.

[0039] Channel integration and feature calibration are performed on the high-resolution feature map to output detailed enhanced features.

[0040] Optionally, the step of fusing the detail enhancement features with the global contour features to output an enhanced feature image includes:

[0041] The detail enhancement features are transformed to ensure that their dimensions and scale are consistent with those of the global contour features, resulting in aligned detail features.

[0042] The global contour features and the aligned detail features are weighted and fused to obtain intermediate output features. The fusion formula is as follows: ;

[0043] The intermediate output features are added element-wise to the original output features to achieve residual fusion. Then, an activation function is used to enhance the feature representation ability to obtain the final fused features, as shown in the formula:

[0044] ;

[0045] The final fused features are channel integrated and spatial dimension standardized to output an enhanced feature image that simultaneously contains the global contour of the giant salamander's respiratory behavior and high-frequency details;

[0046] Where α and β are learnable weights used to adaptively balance the contributions of the two types of features, F merge It is an intermediate output feature, F backbone It is the global contour feature F s It is a high-frequency detail enhancement feature after processing by the super-resolution reconstruction sub-network, u(F s ) indicates that for F s Operations that perform spatial or channel-level transformations, F SRC F represents the enhanced feature image output by the final module. srcnn σ represents the original features directly output by the super-resolution reconstruction subnetwork, and σ is the activation function.

[0047] Optionally, the step of extracting multi-scale respiratory behavior features from the enhanced feature image to obtain depth feature maps at different scales includes:

[0048] The enhanced feature image is normalized to generate a standard feature map;

[0049] The standard feature map is divided into a complementary first feature subset and a second feature subset;

[0050] Extract local features from the first feature subset using the left path and output the local feature sub-map;

[0051] Extract global features from the second feature subset via the right path and output the global feature submap.

[0052] The local feature sub-map and the global feature sub-map are concatenated to generate a fused feature map that simultaneously contains local details and global context;

[0053] Based on the fused feature map, depth feature maps of different scales are obtained.

[0054] Optionally, the step of extracting local features from the first feature subset via a left path and outputting a local feature sub-map includes:

[0055] The first feature subset is input into a standard convolutional layer, and the original convolutional feature map is output.

[0056] The original convolutional feature map is input into the batch normalization layer, and the normalized convolutional feature map is output.

[0057] Apply the SILU activation function to the normalized convolutional feature map and output the activated local feature map;

[0058] The activated local feature map is input into the visual state space module, and the optimized feature map is output.

[0059] The optimized feature map is concatenated with the original convolutional feature map to generate a concatenated local feature map;

[0060] The stitched local feature map is input into a standard convolutional layer for channel integration, then the feature distribution is stabilized by a batch normalization layer, and finally the nonlinear expression is enhanced by the SILU activation function to generate an enhanced local feature sub-map.

[0061] Optionally, the step of extracting global features from the second feature subset via a right path and outputting a global feature subgraph includes:

[0062] The second feature subset is input into the linear transformation layer, and the output is a dimension-adjusted feature map after linear transformation.

[0063] The dimension-adjusted feature map is input into the normalization layer for normalization processing, and the output is a normalized global feature map;

[0064] The normalized global feature map is input into the two-dimensional selective module, and the output is the global feature map;

[0065] Input the global feature map into a depthwise separable convolutional layer and output a lightweight global feature map;

[0066] Apply the SILU activation function to the lightweight global feature map and output the activated global feature map.

[0067] The activated global feature map is input into a linear transformation layer to further adjust the feature dimension and expressive power. Then, a normalization layer is used to stabilize the feature distribution and generate an enhanced global feature sub-map.

[0068] In summary, this application includes the following beneficial technical effects:

[0069] 1. The SRC module is designed specifically for this purpose. By integrating the SRCNN super-resolution branch and the SimpleStem backbone network, it can accurately recover high-frequency details of the image in the early stage of feature extraction, greatly enhance the target perception ability in low-contrast scenes, effectively solve the problems of dim lighting and blurred boundaries between the target and the environment caused by the nocturnal and burrowing habits of the giant salamander, and significantly reduce the false detection rate.

[0070] 2. The C2f module in YOLOv8 is replaced with a multi-scale visual spatial module (MVSM). By integrating multi-scale convolution with spatial attention mechanism, the rich posture change features during the breathing process of giant salamander can be fully captured, further improving the accuracy and robustness of behavior recognition. Attached Figure Description

[0071] Figure 1 This is a main flowchart of an automatic identification method for the lung breathing behavior of a giant salamander according to an embodiment of this application;

[0072] Figure 2 This is a flowchart of the steps to perform feature enhancement on a standard video image and obtain an enhanced feature image.

[0073] Figure 3 This is a flowchart of the steps for extracting multi-scale respiratory behavior features from enhanced feature images to obtain depth feature maps at different scales.

[0074] Figure 4 This is the architecture diagram of the Mamba-YOLO-SRC model;

[0075] Figure 5 This is the SRC module architecture diagram;

[0076] Figure 6 This is a diagram of the MVSM module architecture;

[0077] Figure 7 This is a Grad-CAM heatmap. Detailed Implementation

[0078] In the first aspect, this application discloses an automatic identification method for the lung breathing behavior of giant salamanders.

[0079] Reference Figures 1 to 5 An automatic identification method for the lung breathing behavior of a giant salamander includes steps S101 to S106:

[0080] Step S101: Obtain the initial video image of the target area.

[0081] Specifically, the initial video images refer to the raw video data directly acquired from the acquisition device without any processing, containing complete behavioral images of the giant salamander in the target area (such as the breeding cave and surrounding waters). In this embodiment, infrared cameras (such as the HIKvision DS-2CD3T35D-I5, a dedicated camera device capable of capturing clear images in low-light environments) deployed at the observation entrance and above the giant salamander breeding cave continuously acquire video of the target area 24 hours a day. The acquisition parameters are set to a resolution of 2560x1040 pixels, a frame rate of 50 frames per second, and stored in HEVC format to ensure the capture of complete respiratory behavior details of the giant salamander, including diving (head submerged in water), raising its head (head above water), inhalation (inhalation through relaxation of the oral cavity and pharynx), and exhalation (exhalation in water accompanied by bubbles).

[0082] Step S102: Preprocess the initial video image to obtain a standard video image.

[0083] Specifically, preprocessing refers to the initial processing operations of filtering, transforming, and standardizing the initial video images, with the aim of eliminating low-quality data and reducing noise interference. Standard video images refer to standardized image data whose size, brightness, contrast, and other parameters meet the model input requirements after preprocessing. In this embodiment, through frame-by-frame extraction and grayscale conversion, grayscale filtering, and standardization, low-quality frames (such as overly dark or blurry frames) are eliminated, and image specifications are standardized, laying the foundation for subsequent feature extraction.

[0084] Step S103: Perform feature enhancement on the standard video image to obtain the enhanced feature image.

[0085] Specifically, feature enhancement refers to the process of strengthening key features (such as contours and detailed textures) related to the breathing behavior of giant salamanders in an image using specific algorithms, while suppressing irrelevant background noise; enhanced feature images refer to feature representation data that makes the target features more prominent and has higher recognizability after feature enhancement. In this embodiment, the Small Target Recovery and Enhancement (SRC) module is used to integrate super-resolution reconstruction and lightweight dimensionality reduction structures, simultaneously enhancing the global contour (overall shape of the target) and high-frequency details (such as mouth and nose movements and bubble textures) of the image, solving the problem of difficulty in identifying giant salamander targets under low contrast at night.

[0086] Step S104: Extract multi-scale respiratory behavior features from the enhanced feature image to obtain depth feature maps at different scales.

[0087] Specifically, multi-scale respiratory behavior feature extraction refers to the operation of extracting features related to the breathing behavior of the giant salamander from different spatial dimensions (such as local details and global morphology). A deep feature map refers to a feature matrix containing abstract semantic information of the target, extracted through a deep learning model. Deep feature maps at different scales are multi-level feature data covering small-scale details (such as air bubbles), medium-scale movements (such as head lifting), and large-scale postures (such as body floating). In this embodiment, the Multi-Scale Visual Spatial Module (MVSM) is used to extract local details (such as subtle movements of the mouth and nose) and global contextual features (such as body posture and movement trajectory) through dual paths. These are then fused to generate a multi-scale deep feature map, accurately capturing the posture changes of the giant salamander's breathing behavior.

[0088] Step S105: Perform multi-level feature aggregation on depth feature maps of different scales and generate aggregated feature maps.

[0089] Specifically, in this embodiment, multi-level feature aggregation refers to the process of integrating deep feature maps of different scales and levels of abstraction to strengthen the correlation between features; aggregated feature maps refer to comprehensive feature representations that contain multi-scale and multi-level feature information after aggregation, which can more comprehensively reflect the target behavior features. Through the PAFPN neck network, deep feature maps of different scales are aggregated hierarchically, strengthening the reverse injection of semantic information into lower-level features and improving the completeness and correlation of feature expression.

[0090] Step S106: Generate recognition results based on the aggregated feature map.

[0091] Specifically, the recognition results refer to the categories of giant salamander breathing behavior (diving, raising head, inhaling, exhaling) and their corresponding confidence scores (the model's credibility of the recognition results) output by the model. This data can be directly used to determine the state of giant salamander breathing behavior. In this embodiment, a decoupled detection head (a network component that separates bounding box regression and classification tasks to avoid interference between tasks) is used to perform bounding box regression (predicting the position and size of the target in the image) and classification tasks (determining the target's behavior category) respectively, outputting the category of giant salamander breathing behavior and its corresponding confidence score, thus achieving accurate recognition and labeling of breathing behavior.

[0092] The automatic identification method for giant salamander lung breathing behavior provided in this embodiment filters and standardizes image data through preprocessing steps and removes low-quality frames to reduce noise interference. It enhances key features related to giant salamander breathing in low-contrast environments to improve target recognition by using feature enhancement steps. Then, it comprehensively captures local details and global contextual information of breathing behavior through multi-scale feature extraction and multi-level feature aggregation to ensure accurate capture of subtle movements and concealed postures. Finally, it replaces the traditional manual video playback counting method with a systematic technical link, which not only greatly reduces the missed detection rate, but also realizes long-term, high-frequency, and continuous monitoring of giant salamander breathing behavior, taking into account both monitoring efficiency and data objectivity.

[0093] In one embodiment of this example, step S102 preprocesses the initial video image to obtain a standard video image, including steps S201 to S205:

[0094] Step S201: Extract the initial video image frame by frame and convert the initial video image into a single-channel grayscale image.

[0095] Specifically, a single-channel grayscale image refers to a black and white image that contains only luminance information and has no color channels (such as RGB). It has only one channel, which reduces the amount of data and focuses on luminance features. In this embodiment, the video file is read frame by frame using OpenCV, and the RGB three-channel image is converted into a single-channel grayscale image using a weighted average method. The formula is: Gray = 0.299R + 0.587G + 0.114B, reducing data dimensionality while retaining key luminance information.

[0096] Step S202: Obtain the mean gray level and gray level variance corresponding to different frame images.

[0097] Specifically, the grayscale mean refers to the average brightness value of all pixels in a single frame of grayscale image, reflecting the overall brightness of the image; the grayscale variance refers to the degree to which the pixel brightness value in a single frame of grayscale image deviates from the grayscale mean, reflecting the image contrast. Both are used together to evaluate image quality. In this embodiment, the pixel mean μ and variance σ of each frame of grayscale image are calculated. 2 Where μ = (1 / MN)ΣΣGray(i,j) (M and N are the height and width of the image, respectively, and Gray(i,j) is the gray value of the pixel in the i-th row and j-th column), σ 2 =(1 / MN)ΣΣ(Gray(i,j)-μ) 2 Quantify the brightness distribution and contrast of the image.

[0098] Step S203: Determine whether the grayscale mean and grayscale variance both meet the corresponding grayscale threshold requirements.

[0099] Specifically, the grayscale threshold requirement refers to the preset grayscale mean and variance range used to screen qualified image frames based on the giant salamander farming scenario (such as a nighttime cave environment). Frames exceeding the range are rejected due to insufficient brightness or contrast. In this embodiment, the preset grayscale mean threshold range is [50, 200] and the variance threshold range is [300, 1500], which filters out image frames with moderate brightness and meeting the contrast requirements, and rejects invalid frames that are too dark, too bright, or blurry.

[0100] Step S204: If the grayscale mean and grayscale variance both meet the corresponding grayscale threshold requirements, then the corresponding frame image is taken as the target video image.

[0101] Specifically, the target video image refers to a valid image frame whose brightness and contrast meet the requirements after being filtered by a grayscale threshold. It serves as the input data for subsequent standardization processing. In this embodiment, valid frames are filtered out by threshold judgment to ensure the reliability of subsequent processing data and reduce noise interference (such as image quality degradation caused by ambient light fluctuations).

[0102] Step S205: Standardize the target video image to obtain a standard video image.

[0103] Specifically, standardization refers to the operation of uniformly adjusting the size and pixel value range of the target video image, with the aim of eliminating the impact of data specification differences on model training; a standard video image refers to a standardized image whose size and pixel value range conform to the model input standard. In this embodiment, the target video image is uniformly cropped to a size of 640×640 pixels (unifying the image input specification to adapt to the model input requirements), and the pixel values ​​are mapped to the [0,1] interval using the Min-Max normalization method, with the formula: Norm(x)=(x-min(x)) / (max(x)-min(x)), (where min(x) and max(x) are the minimum and maximum values ​​of pixels in the image, respectively), eliminating the impact of scale differences on model training.

[0104] In one embodiment of this example, step S103 performs feature enhancement on the standard video image to obtain the enhanced feature image, which includes steps S301 to S304:

[0105] Step S301: Input the standard video image into the small target restoration and enhancement module.

[0106] Specifically, in this embodiment, the Small Target Recovery and Enhancement (SRC) module is a feature processing module specifically designed for low-contrast, small-sized targets (such as giant salamanders in cave environments). It can recover target details and enhance feature clarity. The SRC module includes a super-resolution reconstruction sub-network and a lightweight dimensionality reduction structure. The super-resolution reconstruction sub-network (SRCNN) is a deep learning sub-network that improves image resolution and recovers high-frequency details (such as textures and edges) through convolutional operations. The lightweight dimensionality reduction structure (SimpleStem) is a network structure that uses simple convolutional operations to compress the spatial dimension of the image (reduce the width and height) and reduce the computational load, while preserving the global contour features of the target. The SRC module has a two-branch structure: SRCNN is responsible for recovering high-frequency details, and SimpleStem is responsible for spatial dimensionality reduction (reducing the image spatial size and computational load) and global feature extraction. The two work together to improve the target perception capability in low-contrast scenes.

[0107] Step S302: The lightweight dimensionality reduction structure performs spatial compression on the standard video image and outputs global contour features.

[0108] Specifically, spatial compression refers to reducing the width and height of an image through convolution operations, which can reduce computational overhead and focus on core features; global contour features refer to feature data that reflects the overall shape of the target (such as the body contour and head position of the giant salamander), without including local minor details. Spatial dimension compression is achieved through two convolution operations, and key contour information is filtered by the visual cue fusion module to preserve global features such as the giant salamander's body posture.

[0109] Step S303: The super-resolution reconstruction sub-network performs super-resolution reconstruction on the standard video image and outputs detail enhancement features.

[0110] Specifically, super-resolution reconstruction refers to the process of improving image resolution through algorithms to restore high-frequency details (such as edges and textures) lost due to low contrast and low pixel count; detail enhancement features refer to feature data containing subtle local features of the target (such as the movement of the giant salamander's snout and the texture of exhaled air bubbles), which can improve the recognition of target details. In this embodiment, high-frequency details of the image are extracted step by step through three layers of convolution to restore key features of breathing behavior such as snout movements and air bubble textures, thereby improving the recognition of details.

[0111] Step S304: Perform feature fusion between the detail enhancement features and the global contour features to output the enhanced feature image.

[0112] Specifically, feature fusion refers to the process of integrating different types of features to form a more comprehensive and representative integrated feature. In this embodiment, feature fusion is performed through the vision clue merge module. Enhanced feature image refers to feature data that, after fusion, simultaneously contains the global contour and local details of the target, which can effectively improve the target recognition accuracy in low-contrast scenes. In this embodiment, by using weighted fusion and residual connection, the advantages of the two types of features are integrated to generate an enhanced feature image that simultaneously contains global pose and local details, thereby strengthening the feature representation of the giant salamander's breathing behavior.

[0113] In one embodiment of this example, step S302, where the lightweight dimensionality reduction structure spatially compresses the standard video image and outputs global contour features, includes steps S401 to S408:

[0114] Step S401: Perform Conv1 convolution operation on the standard video image with kernel size K=3 and stride=3 to obtain the first convolution feature map.

[0115] Specifically, the convolution kernel refers to the matrix window used to extract image features. K=3 indicates a 3×3 kernel, which can capture feature information within a local 3×3 area. The stride refers to the number of pixels the convolution kernel moves each time it slides across the image. stride=3 indicates that it slides 3 pixels each time, determining the degree of spatial compression. The Conv1 convolution operation refers to the first convolution operation in the module. The first convolution feature map refers to the feature matrix output after the first convolution, containing the initially extracted contour features. In this embodiment, by setting a 3×3 convolution kernel and a stride of 3, the image spatial dimension is compressed to 1 / 3 of its original size while extracting basic features, reducing computational overhead (reducing model computation and improving efficiency).

[0116] Step S402: Perform batch normalization on the first convolutional feature map to obtain the first standardized feature map.

[0117] Specifically, batch normalization refers to the standardization operation performed on the convolutional features of a batch (e.g., 32 images), improving training stability by adjusting the feature value distribution. The first standardized feature map refers to the feature matrix with a more concentrated and stable feature value distribution after batch normalization. In this embodiment, batch normalization eliminates the feature value distribution shift after convolution, thus improving training stability.

[0118] Step S403: Perform ReLU activation function operation based on the first normalized feature map to obtain the first activation feature map.

[0119] Specifically, the first activated feature map refers to the feature matrix that retains effective features and suppresses ineffective features after ReLU activation. In this embodiment, the ReLU activation function introduces nonlinearity (enabling the model to fit complex feature relationships) and enhances the expressive power of effective features.

[0120] Step S404: Perform a Conv2 convolution operation with kernel size K=3 and stride=3 on the first activation feature map to obtain the second convolution feature map.

[0121] Specifically, the Conv2 convolution operation refers to the second convolution operation in the module; the second convolution feature map refers to the feature matrix output after the second convolution, which focuses the contour features more. In this embodiment, the spatial dimension is further compressed to 1 / 9 of the original size by setting a 3×3 convolution and a stride of 3, focusing on the core contour features (the most critical overall shape features of the target).

[0122] Step S405: Perform batch normalization on the second convolutional feature map to obtain the second standardized feature map.

[0123] Specifically, the second normalized feature map refers to the feature matrix whose feature value distribution is stable after batch normalization of the second convolutional feature map. In this embodiment, the batch normalization operation is repeated to continuously stabilize the feature distribution, providing an optimization basis for subsequent activation operations.

[0124] Step S406: Perform ReLU activation function operation based on the second normalized feature map to obtain the second activation feature map.

[0125] Specifically, the second activated feature map refers to the feature matrix that makes the effective contour features more prominent after ReLU activation of the second normalized feature map. In this embodiment, ReLU activation is used to further enhance the nonlinear feature representation and filter out more distinctive contour-related features.

[0126] Step S407: The visual cue fusion module filters key information from the second activation feature map and retains visual cues related to the giant salamander outline.

[0127] Specifically, the visual cue fusion module refers to the module used to filter key visual information related to the target (giant salamander) in the image and remove background noise; giant salamander contour-related visual cues refer to feature information related to target recognition, such as the giant salamander's body contour, head shape, and torso posture. In this embodiment, drawing on the ZNN framework's principle of "prioritizing the processing of dominant feature components under interference," key visual cues related to the giant salamander's body contour and posture are filtered out, and background noise interference is removed to effectively preserve crucial high-frequency details in low-contrast scenes.

[0128] Step S408: Perform channel integration based on the filtered feature maps to output global contour features.

[0129] Specifically, channel integration refers to adjusting the number of channels in the feature map through convolution operations to adapt the number of output channels to the input requirements of subsequent modules; global contour features refer to the integrated feature data that fully reflects the overall morphology of the giant salamander. In this embodiment, the channel dimensions of the filtered feature map are integrated through 1×1 convolution to output global contour features with unified dimensions and prominent contour information.

[0130] In one embodiment of this example, step S303, the super-resolution reconstruction sub-network performs super-resolution reconstruction on the standard video image and outputs detail enhancement features, including steps S501 to S506:

[0131] Step S501: Perform Conv1 convolution operation based on standard video images, with the convolution kernel size set to 64×9×9, padding parameter=4, stride=1, to obtain a preliminary detail feature map with 64 channels.

[0132] Specifically, the 64×9×9 convolution kernel specification refers to the presence of 64 convolution kernels of size 9×9. Each kernel extracts one type of detail feature and outputs 64 feature channels. The padding parameter refers to the number of pixels added to the image edges; padding=4 means adding 4 pixels to each side of the image to ensure that the image size remains unchanged after convolution. The preliminary detail feature map refers to the 64-channel feature matrix of basic details (such as edges and textures) extracted after the first convolution. In this embodiment, the 64×9×9 convolution kernel expands the receptive field, and combined with padding=4 to maintain the image size, it extracts basic image detail information and outputs a 64-channel feature map to enhance the detail representation capability.

[0133] Step S502: Perform ReLU activation function operation on the preliminary detailed feature map to generate an activated intermediate feature map.

[0134] Specifically, the intermediate feature map after activation refers to the intermediate feature matrix that retains effective detailed features and suppresses ineffective features after ReLU activation. In this embodiment, nonlinearity is introduced through ReLU activation to enhance the nonlinear expressive power of the features and highlight effective detailed features (such as the skin texture and mouth and nose edges of the giant salamander).

[0135] Step S503: Perform Conv2 convolution operation based on the activated intermediate feature map, with the convolution kernel size set to 32×5×6, padding parameter=2, stride=1, and extract depth detail information to obtain a 32-channel depth detail feature map.

[0136] Specifically, deep detail information refers to more abstract and critical detailed features than preliminary details (such as the movement trends of the mouth and nose); the deep detail feature map refers to the 32-channel feature matrix extracted after the second convolution to reveal deeper details. In this embodiment, a 5×6 asymmetric convolution kernel is used to further mine deeper details, reducing the number of channels to 32 to balance computational load and feature representation, capturing key respiratory details such as the mouth and nose movements of the giant salamander.

[0137] Step S504: Perform ReLU activation function operation on the deep detail feature map to improve feature discrimination and generate enhanced detail feature map.

[0138] Specifically, feature discrimination refers to the degree of difference between different types of detailed features (such as the details of inhalation and exhalation); enhanced detail feature map refers to a feature matrix where the differences between different detailed features are more obvious after activation. In this embodiment, ReLU activation is used to further filter effective detailed features and improve the discrimination of different breathing behavior details.

[0139] Step S505: Perform Conv3 convolution operation based on the enhanced detail feature map, with the convolution kernel size set to 32×5×5, padding parameter=2, stride=1, to obtain a high-resolution feature map.

[0140] Specifically, a high-resolution feature map refers to a feature matrix with clearer details and higher resolution after a third convolution, reflecting pixel-level subtle features. In this embodiment, a 5×5 convolution kernel is used to refine the detailed features, while maintaining padding=2 to ensure the image size remains unchanged, outputting a high-resolution detailed feature map.

[0141] Step S506: Perform channel integration and feature calibration on the high-resolution feature map to output detailed enhancement features.

[0142] Specifically, feature calibration refers to adjusting the weights of features from different channels through an attention mechanism to enhance key detail channels and suppress irrelevant channels; detail enhancement features refer to the final detail feature data that, after integration and calibration, has more prominent key detail feature responses. In this embodiment, a channel attention mechanism is used to perform feature calibration on the high-resolution feature map, enhancing the feature responses of key details and outputting detail enhancement features.

[0143] In one embodiment of this example, step S304, which fuses the detail enhancement features with the global contour features to output an enhanced feature image, includes steps S601 to S604:

[0144] Step S601: Perform a transformation operation on the detail enhancement feature to ensure that the dimension and scale of the detail enhancement feature are consistent with those of the global contour feature, thus obtaining the aligned detail feature.

[0145] Specifically, transformation operations refer to spatial or channel-level transformation operations, such as interpolation scaling and 1×1 convolution; consistent dimension and scale mean that the width, height, and number of channels of the detail enhancement features and the global contour features are exactly the same, ensuring that the features can be directly fused; aligned detail features refer to detail feature data that, after transformation, matches the specifications of the global contour features. In this embodiment, through interpolation scaling and channel adjustment, the spatial scale and number of channels of the detail enhancement features are made to perfectly match the global contour features, ensuring the effectiveness of the fusion.

[0146] Step S602: Perform weighted fusion of global contour features and aligned detail features to obtain intermediate output features.

[0147] Specifically, in this embodiment, weighted fusion refers to assigning learnable weights (α, β) to the global contour features and the aligned detail features, adjusting the contribution ratio of the two types of features through weights, and then performing a linear superposition fusion method; the intermediate output feature refers to the transitional feature matrix obtained after weighted fusion, which initially integrates the two types of feature information. The fusion formula is: , where α and β are learnable weights used to adaptively balance the contributions of the two types of features (parameters automatically adjusted during model training, with initial values ​​set to 0.5), F backbone For global contour features, F merge It is an intermediate output feature, μ(F) S ) indicates that for F S Operations that perform spatial or channel-level transformations.

[0148] Step S603: Add the intermediate output features to the original output features element by element to achieve residual fusion, and then improve the feature representation ability through the activation function to obtain the final fused features.

[0149] Specifically, the original output features refer to the untransformed detail features directly output by the super-resolution reconstruction subnetwork (F...). srcnn The activation function (here, the Sigmoid function) is used to introduce non-linearity and enhance feature representation. The final fused feature refers to the feature matrix that integrates global contour and detail information after residual fusion and activation. In this embodiment, the fusion formula is: , of which F SRC For the final fused features, σ is the activation function (mapping the output to the [0,1] interval to enhance feature nonlinearity). The original feature information is preserved through residual connection, and the activation function enhances the nonlinear expression of the features.

[0150] Step S604: Perform channel integration and spatial dimension standardization on the final fused features to output an enhanced feature image that simultaneously contains the global contour of the giant salamander's respiratory behavior and high-frequency details.

[0151] In one embodiment of this example, step S104 involves multi-scale respiratory behavior feature extraction from the enhanced feature image to obtain depth feature maps at different scales, including steps S701 to S706:

[0152] Step S701: Normalize the enhanced feature image to generate a standard feature map.

[0153] Specifically, normalization refers to the process of standardizing the feature values ​​of the enhanced feature image to make its distribution more concentrated and stable (here, layer normalization). A standard feature map refers to a feature matrix with a uniform range and stable distribution of feature values ​​after normalization, which can improve model training efficiency. In this embodiment, layer normalization (standardizing the features of a single image, as opposed to batch processing in batch normalization) is used to standardize the enhanced feature image, thereby improving the consistency of feature distribution.

[0154] Step S702: Divide the standard feature map into a complementary first feature subset and a second feature subset.

[0155] Specifically, in this embodiment, the first feature subset refers to the feature data used for subsequent local feature extraction of the left path; the second feature subset refers to the feature data used for subsequent global feature extraction of the right path. In this embodiment, the standard feature map is divided according to the average channel dimension (e.g., a 64-channel feature map is divided into two 32-channel subsets), resulting in two feature subsets with the same dimension and complementary information, which are used for local feature extraction and global feature extraction, respectively.

[0156] Step S703: Extract local features from the first feature subset through the left path and output the local feature sub-map.

[0157] Specifically, the left path refers to the network branch in the module specifically used to extract local detail features, including components such as convolution, normalization, and activation; local features refer to features reflecting small-scale, subtle movements of the target (such as the mouth and nose movements of the giant salamander, and air bubbles); the local feature subgraph refers to the feature matrix output by the left path that focuses on local details. In this embodiment, the left path adopts a combined structure of convolution-normalization-activation-visual state space modules to focus on extracting local detail features of the giant salamander's breathing behavior (such as mouth and nose movements, air bubbles, and other small-scale, subtle features).

[0158] Step S704: Extract global features from the second feature subset through the right path and output the global feature sub-map.

[0159] Specifically, the right path refers to the network branch in the module dedicated to extracting global contextual features, including components such as linear transformation, normalization, and depthwise separable convolution; global features refer to features reflecting the overall shape of the target, its motion trajectory, and its relationship with the surrounding environment (such as the giant salamander's body posture and floating / sinking state); the global feature subgraph refers to the feature matrix output by the right path that focuses on the global context. In this embodiment, the right path adopts a combined structure of linear transformation-normalization-two-dimensional selective module-depthwise separable convolution, focusing on capturing global contextual features such as the giant salamander's body posture and motion trajectory.

[0160] Step S705: Concatenate the local feature sub-map and the global feature sub-map to generate a fused feature map that contains both local details and global context.

[0161] Specifically, the fused feature map refers to a comprehensive feature matrix that, after being stitched together, simultaneously contains local details and global context, thus providing a more complete reflection of the giant salamander's respiratory behavior characteristics. In this embodiment, local feature sub-maps and global feature sub-maps are stitched together along the channel dimension to integrate the two types of feature information and improve the completeness of feature representation.

[0162] The output feature maps of the two pathways are concatenated along the channel dimension, and cross-channel information fusion and dimensional alignment are achieved through subsequent standard convolutional layers. This design not only promotes deep fusion of local and global features but also enhances the model's ability to perceive multi-scale targets. Furthermore, MVSM introduces a residual connection structure, which passes input features to the output through a shortcut path, thereby alleviating the gradient vanishing problem and improving training stability. The entire module significantly enhances the richness and robustness of feature representations while maintaining low computational overhead.

[0163] The output feature maps of the two branches are concatenated along the channel dimension and then processed by a standard convolutional layer to achieve cross-branch information fusion and dimension alignment. This process can be described as follows:

[0164]

[0165] Where F left and F right These represent the output features of the left local branch and the right global modeling branch, respectively. Furthermore, MVSM introduces residual connections through shortcut paths, allowing input features to be directly passed to the output, effectively mitigating the vanishing gradient problem and enhancing training stability. The final output of MVSM is defined as:

[0166] This design not only promotes the deep fusion of local and global features, but also enhances the model's ability to perceive multi-scale targets. The entire module significantly improves the overall performance of feature representation in terms of richness and robustness while maintaining low computational overhead.

[0167] Step S706: Based on the fused feature map, obtain deep feature maps at different scales.

[0168] Specifically, depth feature maps at different scales refer to depth feature matrices extracted using convolutional kernels of different sizes (1×1, 3×3, 5×5), corresponding to features of different spatial dimensions. Small convolutional kernels (1×1) capture small-scale details (such as bubbles), while large convolutional kernels (5×5) capture large-scale global features (such as body posture). In this embodiment, multi-scale convolutional kernels (1×1, 3×3, 5×5, with different kernel sizes corresponding to different scale features) are used to perform convolution operations on the fused feature map to generate depth feature maps at different scales, corresponding to fine-grained details (such as bubble size) and coarse-grained global features (such as body buoyancy) of the giant salamander's respiratory behavior, respectively.

[0169] In one embodiment of this example, step S703 extracts local features from the first feature subset through the left path and outputs a local feature sub-map, including steps S801 to S806:

[0170] Step S801: Input the first feature subset into the standard convolutional layer and output the original convolutional feature map.

[0171] Specifically, a standard convolutional layer refers to a convolutional layer that uses a conventional 3×3 kernel, stride 1, and padding=1, as opposed to special convolutions such as depthwise separable convolution. The original convolutional feature map refers to the feature matrix that initially extracts local basic features (such as the edges of the mouth and nose, and the outline of bubbles) after standard convolution. In this embodiment, a standard convolutional layer with a 3×3 kernel, stride 1, and padding=1 is used to extract local basic features (such as the edges of the mouth and nose, and the outline of bubbles) while maintaining the feature map size.

[0172] Step S802: Input the original convolutional feature map into the batch normalization layer and output the normalized convolutional feature map.

[0173] Specifically, a batch normalization layer refers to a network layer that performs batch normalization operations, which can stabilize the feature value distribution; a normalized convolutional feature map refers to a feature matrix whose feature value distribution is more concentrated and whose training is more stable after batch normalization. In this embodiment, batch normalization stabilizes the feature distribution, reduces the risk of gradient vanishing, and improves training efficiency and feature representation stability.

[0174] Step S803: Apply the SILU activation function to the normalized convolutional feature map and output the activated local feature map.

[0175] Specifically, the SILU activation function refers to the Sigmoid Linear Unit, with the formula SILU(x) = x·σ(x) (where σ is the Sigmoid function), which can introduce smooth nonlinearity; the activated local feature map refers to the feature matrix with stronger local feature discrimination after SILU activation.

[0176] Step S804: Input the activated local feature map into the visual state space module and output the optimized feature map.

[0177] Specifically, in this embodiment, the Visual State Space Module (VSSM) refers to a visual feature processing module designed based on a state space model, which can focus on key regions through a selective scanning mechanism; the optimized feature map refers to a feature matrix that, after VSSM processing, makes key local features (such as mouth and nose movements) more prominent. VSSM focuses on key local regions through a selective scanning mechanism (focusing on key regions in the image and ignoring irrelevant backgrounds), enhancing the feature response of core details of the giant salamander's mouth and nose movements, air bubbles, and other respiratory behaviors.

[0178] Step S805: Concatenate the optimized feature map with the original convolutional feature map to generate a concatenated local feature map.

[0179] Specifically, concatenating local feature maps refers to the process of superimposing the optimized feature map and the original convolutional feature map along the channel dimension to obtain a feature matrix that can simultaneously retain the original basic features and the optimized key detail features. In this embodiment, the optimized feature map and the original convolutional feature map are concatenated along the channel dimension, retaining the original feature information while incorporating the optimized key detail features.

[0180] Step S806: Input the spliced ​​local feature map into a standard convolutional layer for channel integration, then pass it through a batch normalization layer to stabilize the feature distribution, and finally enhance the nonlinear expression through the SILU activation function to generate an enhanced local feature sub-map.

[0181] Specifically, the enhanced local feature sub-map refers to the final feature matrix that, after channel integration, normalization, and activation, makes local details more prominent and more discriminative. In this embodiment, channel integration is performed through 1×1 convolution (unifying the number of channels to avoid channel redundancy), batch normalization is used to stabilize the distribution, SILU activation is used to enhance nonlinearity, and the output is an enhanced local feature sub-map with prominent local details (such as clear mouth and nose movement features).

[0182] In one embodiment of this example, step S704 extracts global features from the second feature subset through the right path and outputs a global feature sub-map, including steps S901 to S906:

[0183] Step S901: Input the second feature subset into the linear transformation layer, and output the dimension-adjusted feature map after linear transformation.

[0184] Specifically, the linear transformation layer refers to a network layer that adjusts the feature channel dimensions through matrix multiplication without changing the width and height of the feature map; the dimension-adjusted feature map refers to a feature matrix whose number of channels is adapted to the input requirements of subsequent modules after linear transformation. In this embodiment, the linear transformation layer adjusts the feature channel dimensions (e.g., from 64 channels to 32 channels) to adapt to the input requirements of subsequent modules, while preserving global feature information (e.g., overall body posture).

[0185] Step S902: Input the dimension-adjusted feature map into the normalization layer for normalization processing, and output the normalized global feature map.

[0186] Specifically, in this embodiment, the normalization layer refers to the network layer that performs layer normalization operations to standardize a single feature map; the normalized global feature map refers to the feature matrix with stable feature value distribution and more consistent global features after layer normalization.

[0187] Step S903: Input the normalized global feature map into the two-dimensional selective module and output the global feature map.

[0188] Specifically, in this embodiment, the global feature map refers to the feature matrix that contains global information such as the overall pose and motion trajectory of the target after SS2D processing.

[0189] Step S904: Input the global feature map into the depthwise separable convolutional layer and output a lightweight global feature map.

[0190] Specifically, in this embodiment, a depthwise separable convolutional layer refers to a network layer that performs depthwise separable convolution operations, splitting standard convolution into depthwise convolution (extracting spatial features) and pointwise convolution (integrating channel features), which can significantly reduce the amount of computation; a lightweight global feature map refers to a feature matrix that, after being processed by depthwise separable convolution, has fewer parameters and lower computational overhead, but still retains the ability to express global features.

[0191] Step S905: Apply the SILU activation function to the lightweight global feature map and output the activated global feature map.

[0192] Specifically, in this embodiment, the activated global feature map refers to the feature matrix that has stronger nonlinear representation of global features and higher distinguishability of different pose features after SILU activation.

[0193] Step S906: Input the activated global feature map into the linear transformation layer to further adjust the feature dimension and expressive power, and then pass it through the normalization layer to stabilize the feature distribution and generate an enhanced global feature sub-map.

[0194] Specifically, the enhanced global feature subgraph refers to the final global feature matrix that, after final linear transformation and normalization, has clearer global context information and is more suitable for subsequent fusion operations. In this embodiment, the feature dimension is optimized through linear transformation and the distribution is normalized to be stable, outputting an enhanced global feature subgraph with clear global context information.

[0195] It is worth noting that, to gain a deeper understanding of the model's decision-making mechanism in recognizing the respiratory behavior of the giant salamander, this application uses Grad-CAM to visualize the feature responses of the Mamba-YOLO-SRC model. For example... Figure 7 As shown, the heatmap reveals that the model's attention is primarily focused on the head and snout region of the giant salamander. Notably, the heatmap distribution closely matches the target contour, with minimal response from the background region, indicating strong anti-interference capabilities and accurate target localization. Overall visualization results demonstrate that Mamba-YOLO-SRC does not rely on global texture or background information for judgment, but rather makes decisions based on key local features of the individual target, validating the model's good interpretability and behavioral consistency in complex underwater environments.

[0196] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for automatically identifying the respiratory behavior of a giant salamander, characterized in that, The method comprises the following steps: acquiring an initial video image of a target region; preprocessing the initial video image to obtain a standard video image; enhancing features of the standard video image to obtain an enhanced feature image; extracting multi-scale respiratory behavior features from the enhanced feature image to obtain deep feature maps of different scales; aggregating multi-level features of the deep feature maps of different scales to generate an aggregated feature map; generating a recognition result based on the aggregated feature map; wherein the preprocessing of the initial video image to obtain a standard video image comprises: extracting each frame of the initial video image and converting the initial video image into a single-channel grayscale image; acquiring the mean and variance of the grayscale of different frames of images; determining whether the mean and variance of the grayscale satisfy the corresponding grayscale threshold requirements; if the mean and variance of the grayscale satisfy the corresponding grayscale threshold requirements, the corresponding frame of image is taken as a target video image; standardizing the target video image to obtain a standard video image; the enhancing features of the standard video image to obtain an enhanced feature image comprises: inputting the standard video image into a small target recovery and enhancement module, which comprises a super-resolution reconstruction subnetwork and a lightweight dimension reduction structure; the lightweight dimension reduction structure performs spatial compression on the standard video image to output global contour features; the super-resolution reconstruction subnetwork performs super-resolution reconstruction on the standard video image to output detail enhancement features; the detail enhancement features and the global contour features are fused to output an enhanced feature image; the extracting multi-scale respiratory behavior features from the enhanced feature image to obtain deep feature maps of different scales comprises: normalizing the enhanced feature image to generate a standard feature map; dividing the standard feature map into a first feature subset and a second feature subset; extracting local features from the first feature subset through a left path to output a local feature submap; extracting global features from the second feature subset through a right path to output a global feature submap; splicing the local feature submap and the global feature submap to generate a fusion feature map containing local details and global context; based on the fusion feature map, deep feature maps of different scales are obtained.

2. The automatic identification method of the lung breathing behavior of Andrias davidianus according to claim 1, characterized in that, the spatial compression of the standard video image by the lightweight dimension reduction structure to output global contour features comprises: performing a Conv1 convolution operation with a kernel size K=3 and a stride stride=3 on the standard video image to obtain a first convolution feature map; performing batch normalization processing on the first convolution feature map to obtain a first standardized feature map; performing a ReLU activation function operation based on the first standardized feature map to obtain a first activated feature map; performing a Conv2 convolution operation with a kernel size K=3 and a stride stride=3 on the first activated feature map to obtain a second convolution feature map; performing batch normalization processing on the second convolution feature map to obtain a second standardized feature map; Performing a ReLU activation function operation based on the second standardized feature map to obtain a second activated feature map; Performing key information screening on the second activated feature map through a visual cue fusion module to retain the anduan salamander contour-related visual cues; Performing channel integration based on the screened feature map to output a global contour feature.

3. The automatic identification method of the lung breathing behavior of Andrias davidianus according to claim 1, characterized in that, The super-resolution reconstruction sub-network performs super-resolution reconstruction on the standard video image to output a detail enhancement feature, including: Performing a Conv1 convolution operation based on the standard video image, with a convolution kernel specification of 64x9x9, padding parameters of padding=4, and a step length of stride=1, to obtain a 64-channel preliminary detail feature map; Performing a ReLU activation function operation on the preliminary detail feature map to generate an activated intermediate feature map; Performing a Conv2 convolution operation based on the activated intermediate feature map, with a convolution kernel specification of 32x5x6, padding parameters of padding=2, and a step length of stride=1, to extract deep detail information and obtain a 32-channel deep detail feature map; Performing a ReLU activation function operation on the deep detail feature map to improve feature discrimination and generate a reinforced detail feature map; Performing a Conv3 convolution operation based on the reinforced detail feature map, with a convolution kernel specification of 32x5x5, padding parameters of padding=2, and a step length of stride=1, to obtain a high-resolution feature map; Performing channel integration and feature calibration on the high-resolution feature map to output a detail enhancement feature.

4. The automatic identification method of the lung breathing behavior of Andrias davidianus according to claim 1, characterized in that, The feature fusion of the detail enhancement feature and the global contour feature to output an enhanced feature image includes: Performing a transformation operation on the detail enhancement feature to make the dimensions and scales of the detail enhancement feature and the global contour feature consistent, obtaining an aligned detail feature; The global contour feature and the aligned detail feature are weighted and fused to obtain an intermediate output feature, and a fusion formula is as follows: ; Element-wise addition of the intermediate output feature and the original output feature to realize residual fusion, and then improving the feature expression capability through an activation function to obtain the final fusion feature, with the formula being: ； Performing channel integration and spatial dimension standardization on the final fusion feature to output an enhanced feature image containing both the anduan salamander breathing behavior global contour and high-frequency details; wherein α , β are learnable weights for adaptive balancing the contribution of two types of features, F merge is the intermediate output feature, F backbone is the global contour feature, F s is the high-frequency detail enhancement feature processed by the super-resolution reconstruction subnetwork, u(F s ) represents the operation of spatial or channel-level transformation on F s , F SRC represents the enhanced feature map output by the final module, F srcnn is the original feature directly output by the super-resolution reconstruction subnetwork, and σ is an activation function.

5. The automatic identification method of the lung breathing behavior of Andrias davidianus according to claim 1, characterized in that, The left path extraction of the first feature subset to output a local feature subgraph includes: Inputting the first feature subset into a standard convolution layer and outputting an original convolution feature map; Inputting the original convolution feature map into a batch normalization layer to output a normalized convolution feature map; Applying a SILU activation function to the normalized convolution feature map to output an activated local feature map; Inputting the activated local feature map into a visual state space module to output an optimized feature map; Concatenating the optimized feature map and the original convolution feature map to generate a concatenated local feature map; Inputting the concatenated local feature map into a standard convolution layer for channel integration, then stabilizing the feature distribution through a batch normalization layer, and finally enhancing the non-linear expression through a SILU activation function to generate an enhanced local feature subgraph.

6. The automatic identification method of the lung breathing behavior of Andrias davidianus according to claim 1, characterized in that, The right path extraction of the second feature subset to output a global feature subgraph includes: The second feature subset is input into a linear transformation layer, and the output dimension is adjusted after linear transformation to output a dimension-adjusted feature map; The dimension-adjusted feature map is input into a normalization layer for normalization processing, and a normalized global feature map is output; The normalized global feature map is input into a two-dimensional selective module, and a global feature map is output; The global feature map is input into a depth separable convolution layer, and a lightweight global feature map is output; An SILU activation function is applied to the lightweight global feature map, and an activated global feature map is output; The activated global feature map is input into a linear transformation layer, and the feature dimension and expression ability are further adjusted, and then the feature distribution is stabilized by a normalization layer to generate an enhanced global feature sub-map.

Images