Method and system for recognizing dangerous behavior of personnel in aquaculture water area based on deep learning

Abstract

The application discloses a kind of based on deep learning's aquaculture water area personnel dangerous behavior identification method and system, belong to target detection technical field.The method includes: obtaining and pre-processing water area personnel image, to construct data set by pre-processing image;Dangerous behavior identification network is built;According to data set, dangerous behavior identification network is trained and tested, and the trained dangerous behavior identification network is obtained;Real-time acquisition target monitoring image of target aquaculture water area is carried out by monitoring system, and based on the trained dangerous behavior identification network, target monitoring image is handled, and dangerous behavior identification result is obtained.The dangerous behavior identification network of the application effectively improves the feature extraction ability of existing general detection algorithm in water area dangerous behavior identification, and then leads to the problems of false detection and missed detection.

Description

Technical Field

[0001] This application belongs to the field of target detection technology, specifically relating to a method and system for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning. Background Technology

[0002] With the large-scale development of the aquaculture industry, the aquaculture water area is large and highly open, and dangerous behaviors such as illegal swimming and illegal fishing occur frequently. This not only causes losses to aquatic products, but also easily leads to safety accidents such as drowning.

[0003] Currently, the identification of dangerous behaviors in water areas mainly relies on manual patrols and general detection algorithms. However, these general detection algorithms are not optimized for aquaculture water scenarios. Complex backgrounds such as water surface reflections, ripples, and changes in lighting can severely interfere with feature extraction, making it easy for human targets to be submerged in the background, resulting in a high rate of false positives and false negatives. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, which can solve one of the technical problems existing in the prior art or related technologies.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: Firstly, embodiments of this application provide a method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, the method comprising: Images of people in the water are acquired and preprocessed to obtain preprocessed images; a dataset is constructed based on the preprocessed images; A dangerous behavior recognition network is constructed, which includes a multi-scale feature extraction and interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; The dataset is divided into a training set and a test set according to a preset ratio, and the dangerous behavior recognition network is trained and tested based on the training set and the test set respectively to obtain a trained dangerous behavior recognition network. The system collects real-time monitoring images of the target aquaculture area and processes these images based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results.

[0006] As an optional implementation of the first aspect of this application, the step of processing the target surveillance image based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results includes: The target monitoring image is processed by the multi-scale feature extraction interaction module to obtain first-scale features, multi-scale interactive features and third-scale features respectively; The local enhancement module enhances the first-scale feature to obtain the local enhanced feature. The global semantic enhancement module is used to perform global semantic enhancement on the third-scale features to obtain globally semantically enhanced features. The feature fusion module fuses the local enhancement features, multi-scale interaction features, and global semantic enhancement features to obtain the target fusion feature. The detection head processes the target fusion features to obtain the dangerous behavior recognition result corresponding to the target monitoring image.

[0007] As an optional implementation of the first aspect of this application, the step of processing the target monitoring image according to the multi-scale feature extraction interaction module to obtain first-scale features, multi-scale interactive features, and third-scale features includes: The target monitoring image is subjected to two consecutive 3×3 convolutions to obtain basic information features, and the basic information features are then subjected to directional spatial attention enhancement to obtain directional spatial enhancement features. The directional spatial enhancement feature is processed by 3×3 convolution to obtain the first scale feature; the first scale feature is concatenated with the directional spatial enhancement feature to obtain the first concatenated and fused feature; The first splicing and fusion feature is processed by a 5×5 convolution to obtain the second scale feature; the second scale feature is spliced ​​with the first splicing and fusion feature to obtain the second splicing and fusion feature. The second splicing and fusion feature is convolved with 7×7 to obtain the third scale feature, and the second scale feature and the third scale feature are subjected to feature interaction processing to obtain multi-scale interactive features.

[0008] As an optional implementation of the first aspect of this application, the step of performing directional spatial attention enhancement on the basic information features to obtain directional spatially enhanced features includes: The basic information features are subjected to horizontal max pooling and average pooling respectively to obtain horizontal max pooling features and horizontal average pooling features; and the horizontal max pooling features and horizontal average pooling features are added together to obtain the horizontal pooling features; The basic information features are subjected to vertical max pooling and average pooling respectively to obtain vertical max pooling features and vertical average pooling features; and the vertical max pooling features and vertical average pooling features are added together to obtain vertical pooling features. The horizontal pooling features and vertical pooling features are concatenated to obtain a fused pooling feature; and the fused pooling feature is then subjected to 1×1 convolution and ReLU activation sequentially to obtain a first activation feature; The first activation feature is split along the channel dimension into a first segmentation feature and a second segmentation feature; The first segmentation feature is sequentially subjected to 1×1 convolution and Sigmoid activation to obtain the first weight; the second segmentation feature is sequentially subjected to 1×1 convolution and Sigmoid activation to obtain the second weight. The basic information features are multiplied sequentially by the first weight and the second weight to obtain the directional spatial enhancement features.

[0009] As an optional implementation of the first aspect of this application, the step of performing feature interaction processing on the second scale feature and the third scale feature to obtain multi-scale interactive features includes: The second scale feature is concatenated with the third scale feature to obtain the first concatenated feature; Spatial importance and channel importance models are performed on the first splicing feature respectively to obtain spatial importance features and channel importance features; The spatial importance feature and the channel importance feature are added together to obtain the fused importance feature; The fusion importance feature is combined with the first splicing feature to obtain the second splicing feature; The second spliced ​​feature is sequentially subjected to channel shuffling, 7×7 group convolution, and Sigmoid activation to obtain the fused attention weights; Multiply the fused attention weights by the second scale feature to obtain the second scale enhanced feature; The difference between 1 and the fused attention weight is calculated as a complementary weight. The complementary weight is then multiplied by the third-scale feature to obtain the third-scale enhanced feature. The second scale enhancement feature and the third scale enhancement feature are added together to obtain the multi-scale interactive feature.

[0010] As an optional implementation of the first aspect of this application, the step of enhancing the first-scale feature according to the local enhancement module to obtain the local enhanced feature includes: The first scale feature is sequentially subjected to 3×3 convolution, batch normalization and ReLU activation to obtain the first convolutional feature; The first convolutional feature is processed sequentially based on multiple concatenated residual blocks to obtain the residual enhancement feature output by the last residual block; The local enhancement features are obtained by performing 3×3 convolution, batch normalization, and ReLU activation on the residual enhancement features. The data processing procedure for each residual block includes: performing 3×3 convolution, batch normalization, ReLU activation, 3×3 convolution, and batch normalization sequentially on the input features of the residual block to obtain batch normalized features; adding the batch normalized features to the input features of the residual block to obtain residual fusion features; and performing ReLU activation on the residual fusion features to obtain the output features of the residual block.

[0011] As an optional implementation of the first aspect of this application, the step of performing global semantic enhancement on the third-scale features according to the global semantic enhancement module to obtain globally semantically enhanced features includes: The third-scale features are subjected to global max pooling and global average pooling respectively to obtain global max pooling features and global average pooling features respectively; The global max pooling feature and the global average pooling feature are added together to obtain the global fusion feature. The global fusion feature is then processed by fully connected processing and ReLU activation in sequence to obtain the global context feature. The global context features are sequentially processed with fully connected layers and activated with Sigmoid to obtain global semantic awareness weights; and the global semantic awareness weights are multiplied by the third-scale features to obtain the global semantic enhancement features.

[0012] Secondly, embodiments of this application provide a deep learning-based system for identifying dangerous behaviors of personnel in aquaculture waters, the system comprising: Data construction module: used to acquire images of people in the water and preprocess them to obtain preprocessed images; and to construct a dataset based on the preprocessed images; Network building module: used to build a dangerous behavior recognition network, which includes a multi-scale feature extraction interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; Network training module: used to divide the dataset into training set and test set according to a preset ratio, and to train and test the dangerous behavior recognition network based on the training set and the test set respectively, so as to obtain the trained dangerous behavior recognition network; Real-time monitoring module: used to collect target monitoring images of the target aquaculture water area in real time through the monitoring system, and process the target monitoring images based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results.

[0013] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0014] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0015] In the embodiments of this application, compared with the prior art, the following technical effects are achieved: This application enhances directional spatial attention by selectively extracting spatial feature weights in the horizontal and vertical directions, effectively suppressing background noise unique to water bodies such as ripples and reflections, improving the recognition accuracy of human target features, and solving the problem of targets being submerged in complex water backgrounds. It employs convolution with differentiated receptive fields to extract multi-scale features, and achieves complementary interaction of features at different scales through spatial and channel-based dual-dimensional importance modeling, balancing the action details of large near-field targets with the feature integrity of small far-field targets. The local enhancement module strengthens the details of features such as limb movements and tool textures through concatenated multiple residual blocks, improving the accuracy of distinguishing dangerous actions from normal behavior; the global semantic enhancement module strengthens the scene semantics of features through global context-aware weights, avoiding misjudgments caused by isolated local features. A weighted feature fusion mechanism with learnable parameters dynamically adjusts the fusion weights of local details, multi-scale interactions, and global semantic features, adapting to the recognition needs of different lighting and water environments. This application's dangerous behavior recognition network effectively improves the problem of insufficient feature extraction capabilities in existing general detection algorithms for dangerous behavior recognition in water bodies, leading to false positives and false negatives. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, provided in some embodiments of this application. Figure 2 This is a schematic diagram of the structure of a dangerous behavior recognition network in a deep learning-based method for recognizing dangerous behaviors of personnel in aquaculture waters, provided by some embodiments of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0019] The following description, in conjunction with the accompanying drawings, details a method and system for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, through specific embodiments and application scenarios.

[0020] Example A deep learning-based method for identifying hazardous behaviors of personnel in aquaculture waters includes the following steps: S100: Acquire images of people in the water and preprocess them to obtain preprocessed images; construct a dataset based on the preprocessed images; It should be noted that multiple images of people in the water are acquired based on existing data or surveillance cameras; these images are then preprocessed, including data cleaning, noise reduction, data augmentation, and pixel and size normalization, to obtain multiple preprocessed images; finally, these preprocessed images are used to construct a dataset for subsequent network training.

[0021] S200: Obtain the dangerous behavior recognition network, which includes a multi-scale feature extraction interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; S300: Divide the dataset into a training set and a test set according to a preset ratio, and train and test the dangerous behavior recognition network based on the training set and the test set respectively to obtain the trained dangerous behavior recognition network; It's important to understand that the dataset is divided into training and testing sets in a 4:1 ratio; a loss function is constructed, and the dangerous behavior recognition network is trained iteratively through multiple rounds using the training set; in each round of training, the loss is calculated using the loss function, and the Adam optimizer is used for backpropagation to optimize the network parameters; training stops when the network parameters converge; at this point, the trained network is tested using the dataset; and the test results are evaluated based on metrics such as precision and recall; if the evaluation results meet the recognition accuracy requirements, the training of the dangerous behavior recognition network is complete; if the evaluation results do not meet the requirements, the network parameters and training strategy are adjusted, and retraining is performed until the evaluation results meet the recognition accuracy requirements.

[0022] S400: The system collects real-time monitoring images of the target aquaculture area through the monitoring system, and processes the target monitoring images based on a trained dangerous behavior recognition network to obtain dangerous behavior recognition results. It should be noted that by deploying surveillance cameras in the target aquaculture area, target surveillance images of the target aquaculture area are collected based on these cameras. The collected target surveillance images need to undergo the same preprocessing process as in S100. The preprocessed target surveillance images are then input into the dangerous behavior recognition network for processing. The network outputs the dangerous behavior recognition result corresponding to the target surveillance image. When the dangerous behavior recognition result shows that there are dangerous actions such as swimming or fishing, the alarm system of the aquaculture area will be triggered to issue a warning.

[0023] As an optional implementation of this embodiment, in step S400, the target monitoring image is processed based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results, including: S410: The target monitoring image is processed by the multi-scale feature extraction interaction module to obtain the first-scale feature, the multi-scale interactive feature and the third-scale feature respectively; S420: Enhance the first-scale features using the local enhancement module to obtain locally enhanced features; S430: Perform global semantic enhancement on the third-scale features based on the global semantic enhancement module to obtain globally semantically enhanced features; S440: Based on the feature fusion module, local enhancement features, multi-scale interaction features, and global semantic enhancement features are fused to obtain the target fusion feature; S450: The detection head processes the target fusion features to obtain the dangerous behavior recognition result corresponding to the target monitoring image.

[0024] It should be noted that the feature fusion module in S440 uses three learnable parameters as weights for local enhancement features, multi-scale interaction features, and global semantic enhancement features, respectively. Then, the local enhancement features, multi-scale interaction features, and global semantic enhancement features are multiplied by their respective weights to obtain weighted local enhancement features, multi-scale interaction features, and global semantic enhancement features. Finally, the weighted local enhancement features, multi-scale interaction features, and global semantic enhancement features are fused by feature addition to obtain the final target fusion feature. The detection head of S450 processes the target fusion feature and outputs the dangerous behavior recognition result corresponding to the target monitoring image. The detection head will outline the specific location where the dangerous behavior occurs in the target monitoring image.

[0025] As an optional implementation of this embodiment, S410 includes: S411: Perform two consecutive 3×3 convolutions on the target monitoring image to obtain basic information features, and then perform directional spatial attention enhancement on the basic information features to obtain directional spatial enhancement features; S412: Perform 3×3 convolution on the directional spatial enhancement features to obtain the first scale feature; concatenate the first scale feature with the directional spatial enhancement features to obtain the first concatenated and fused feature; S413: Perform a 5×5 convolution on the first splicing and fusion feature to obtain the second scale feature; splice the second scale feature with the first splicing and fusion feature to obtain the second splicing and fusion feature; S414: Perform a 7×7 convolution on the second splicing and fusion feature to obtain the third scale feature, and perform feature interaction processing on the second scale feature and the third scale feature to obtain multi-scale interactive features.

[0026] It should be noted that the multi-scale feature extraction interaction module first performs two consecutive 3×3 convolutions on the target monitoring image using two concatenated 3×3 convolutional layers to extract basic information features. Then, directional spatial attention enhancement is applied to these basic information features to obtain directional spatial enhancement features. Next, a 3×3 convolution is performed on the directional spatial enhancement features to obtain the first-scale feature. The first-scale feature is concatenated with the directional spatial enhancement feature to obtain the first concatenated fusion feature. The first concatenated fusion feature is then concatenated with a 5×5 convolution to obtain the second-scale feature. The second-scale feature is then concatenated with the first concatenated fusion feature to obtain the second concatenated fusion feature. Similarly, a 7×7 convolution is performed on the second concatenated fusion feature to obtain the third-scale feature. Thus, three different scales of first-scale features, second-scale features, and third-scale features are obtained. Finally, the second-scale features and third-scale features are processed to achieve information interaction between features of different scales, ultimately obtaining multi-scale interactive features.

[0027] As an optional implementation of this embodiment, in S411, directional spatial attention enhancement is performed on the basic information features to obtain directional spatial enhancement features, including: S4111: Perform horizontal max pooling and average pooling on the basic information features respectively to obtain horizontal max pooling features and horizontal average pooling features; then add the horizontal max pooling features and horizontal average pooling features to obtain the horizontal pooling features; S4112: Perform vertical max pooling and average pooling on the basic information features respectively to obtain vertical max pooling features and vertical average pooling features; then add the vertical max pooling features and vertical average pooling features to obtain the vertical pooling features. S4113: Concatenate the horizontal pooling features and the vertical pooling features to obtain the fused pooling features; and perform 1×1 convolution and ReLU activation on the fused pooling features in sequence to obtain the first activation features; S4114: Split the first activation feature along the channel dimension into a first segmentation feature and a second segmentation feature; S4115: Perform 1×1 convolution and Sigmoid activation on the first segmentation feature in sequence to obtain the first weight; perform 1×1 convolution and Sigmoid activation on the second segmentation feature in sequence to obtain the second weight; S4116: Multiply the basic information features sequentially by the first weight and the second weight to obtain the directional spatial enhancement features.

[0028] It should be noted that, firstly, horizontal max pooling and average pooling are performed on the basic information features to obtain horizontal max pooling features and horizontal average pooling features, respectively. These features are then fused together to obtain the horizontal pooling feature. Similarly, vertical max pooling and average pooling are performed on the basic information features to obtain vertical max pooling features and vertical average pooling features, respectively. These features are then fused together to obtain the vertical pooling feature. Next, the horizontal and vertical pooling features are fused together through feature concatenation to obtain the fused pooling feature. Finally, a series of 1×1 convolutional layers and... The ReLU activation layer sequentially performs 1×1 convolution and ReLU activation on the fused pooling features to obtain the first activation feature. Then, the first activation feature is split into a first segmentation feature and a second segmentation feature by equal channel segmentation. The first segmentation feature is then sequentially subjected to 1×1 convolution and sigmoid activation by a concatenated 1×1 convolution layer and sigmoid activation layer to obtain the first weight. Similarly, the second segmentation feature is sequentially subjected to 1×1 convolution and sigmoid activation by a concatenated 1×1 convolution layer and sigmoid activation layer to obtain the second weight. Finally, the basic information feature is multiplied by the obtained first and second weights to obtain the directional spatial enhancement feature.

[0029] As an optional implementation of this embodiment, in S414, feature interaction processing is performed on the second-scale features and the third-scale features to obtain multi-scale interactive features, including: S4141: Concatenate the second-scale feature with the third-scale feature to obtain the first concatenated feature; S4142: Perform spatial importance and channel importance modeling on the first splicing feature respectively to obtain spatial importance features and channel importance features respectively; S4143: Add the spatial importance feature and the channel importance feature to obtain the fused importance feature; S4144: Combine the importance feature with the first concatenation feature to obtain the second concatenation feature; S4145: The second concatenated feature is sequentially subjected to channel shuffling, 7×7 group convolution, and Sigmoid activation to obtain the fused attention weights; S4146: Multiply the fused attention weights by the second-scale features to obtain the second-scale enhanced features; S4147: Calculate the difference between 1 and the fusion attention weight as the complementary weight, multiply the complementary weight by the third-scale feature to obtain the third-scale enhanced feature; S4148: Add the second-scale enhancement features and the third-scale enhancement features to obtain multi-scale interactive features.

[0030] It should be noted that the second-scale features and the third-scale features are first fused by concatenation to obtain the first concatenated feature. Then, spatial importance and channel importance models are performed on the first concatenated feature to obtain spatial importance features and channel importance features, respectively. Specifically, spatial importance modeling of the first concatenated feature involves performing global average pooling and global max pooling on the channel dimension to obtain the first average pooling feature and the first max pooling feature, respectively. These first average pooling features and the first max pooling features are then concatenated to obtain the pooled concatenated feature. Finally, a 7×7 convolutional layer is applied to the pooled concatenated feature to obtain the spatial importance feature. Channel importance modeling of the first concatenated feature involves first performing global average pooling on the first concatenated feature to obtain the second average pooling feature. Then, a series of 1×1 convolutional layers, ReLU activation, and 1×1 convolutional layers are applied sequentially to the second average pooling feature to obtain the channel importance feature. Then, the spatial importance feature and the channel importance feature are added together to obtain the fused importance feature; this fused importance feature is then concatenated with the first concatenated feature to obtain the second concatenated feature. Subsequently, the second concatenated feature is subjected to channel shuffling, 7×7 convolutions, and sigmoid activation sequentially to obtain the fused attention weight; this fused attention weight is then multiplied with the second-scale feature to obtain the second-scale enhancement feature. Finally, the difference between 1 and the fused attention weight is calculated as the complementary weight, and this complementary weight is multiplied with the third-scale feature to obtain the third-scale enhancement feature; finally, the second-scale enhancement feature and the third-scale enhancement feature are added together to obtain the multi-scale interactive feature.

[0031] As an optional implementation of this embodiment, S420 includes: S421: Perform 3×3 convolution, batch normalization and ReLU activation on the first scale feature in sequence to obtain the first convolution feature; S422: The first convolutional feature is processed sequentially based on multiple concatenated residual blocks to obtain the residual enhancement feature output by the last residual block; S423: Perform 3×3 convolution, batch normalization and ReLU activation on the residual enhancement features to obtain local enhancement features; The data processing for each residual block includes: performing 3×3 convolution, batch normalization, ReLU activation, 3×3 convolution, and batch normalization sequentially on the input features of the residual block to obtain batch normalized features; adding the batch normalized features to the input features of the residual block to obtain residual fusion features; and performing ReLU activation on the residual fusion features to obtain the output features of the residual block.

[0032] It should be noted that the first convolutional feature is obtained by sequentially performing 3×3 convolution, batch normalization (BN) layer, and ReLU activation layer on the first scale feature. Then, the first convolutional feature is processed by multiple concatenated residual blocks. The input of the first residual block is the first convolutional feature, the output of the first residual block is used as the input of the second residual block, and the output of the previous residual block is used as the output of the next residual block, thus obtaining the residual enhancement feature output by the last residual block. Finally, the residual enhancement feature is subjected to 3×3 convolution, batch normalization layer, and ReLU activation layer again to obtain the local enhancement feature.

[0033] As an optional implementation of this embodiment, S430 includes: S431: Perform global max pooling and global average pooling on the third-scale features respectively to obtain global max pooling features and global average pooling features. S432: Add the global max pooling feature and the global average pooling feature to obtain the global fused feature, and then perform fully connected processing and ReLU activation on the global fused feature in sequence to obtain the global context feature; S433: Perform fully connected processing and Sigmoid activation on the global context features in sequence to obtain global semantic awareness weights; then multiply the global semantic awareness weights with the third-scale features to obtain global semantic enhancement features.

[0034] It should be noted that, firstly, global max pooling and global average pooling are performed on the third-scale features to obtain global max pooling features and global average pooling features, respectively. Then, the global max pooling features and global average pooling features are fused by feature addition to obtain global fused features. Next, the global fused features are processed by fully connected layers and ReLU activation layers with the first parameter to obtain global context features. Then, the global context features are processed by fully connected layers and Sigmoid activation layers with the second parameter to obtain global semantic awareness weights. Finally, the global semantic awareness weights are multiplied by the third-scale features to obtain global semantic enhancement features.

[0035] This embodiment presents a deep learning-based method for identifying dangerous behaviors of people in aquaculture waters. Through directional spatial attention enhancement, it selectively extracts spatial feature weights in the horizontal and vertical directions, effectively suppressing background noise specific to aquatic environments such as ripples and reflections, thus improving the recognizability of human target features and solving the problem of targets being submerged in complex aquatic backgrounds. It employs convolution with differentiated receptive fields to extract multi-scale features, and simultaneously achieves complementary interaction of features at different scales through spatial and channel-based dual-dimensional importance modeling, taking into account both the action details of large near-field targets and the feature integrity of small far-field targets. The local enhancement module strengthens the details of features such as limb movements and tool textures through multi-residual block concatenation, improving the accuracy of distinguishing dangerous actions from normal behaviors. The global semantic enhancement module strengthens the scene semantics of features through global context-aware weights, avoiding misjudgments caused by isolated local features. A learnable parameter-based weighted feature fusion mechanism dynamically adjusts the fusion weights of local details, multi-scale interactions, and global semantic features, adapting to the recognition needs of different lighting and aquatic environments. The hazardous behavior recognition network in this embodiment effectively improves the problem of insufficient feature extraction capability in the recognition of hazardous behaviors in water areas by existing general detection algorithms, which leads to false detections and false negatives.

[0036] It should be noted that the method for identifying dangerous behaviors of personnel in aquaculture areas based on deep learning provided in this application can be executed by a system for identifying dangerous behaviors of personnel in aquaculture areas based on deep learning, or by a control module within that system for executing the method. This application uses the execution of the method by a system for identifying dangerous behaviors of personnel in aquaculture areas based on deep learning to illustrate this method.

[0037] A deep learning-based system for identifying hazardous behaviors of personnel in aquaculture waters includes: Data construction module: used to acquire images of people in water areas and preprocess them to obtain preprocessed images; and to construct a dataset based on the preprocessed images; Network building module: used to build a dangerous behavior recognition network, which includes a multi-scale feature extraction interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; Network training module: used to divide the dataset into training set and test set according to a preset ratio, and to train and test the dangerous behavior recognition network based on the training set and test set respectively, so as to obtain the trained dangerous behavior recognition network; Real-time monitoring module: Used to collect target monitoring images of the target aquaculture water area in real time through the monitoring system, and process the target monitoring images based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results.

[0038] The deep learning-based human hazardous behavior identification system in aquaculture waters, as described in this application embodiment, can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), etc. This application embodiment does not impose specific limitations.

[0039] The deep learning-based human hazardous behavior recognition system in aquaculture waters, as described in this application embodiment, can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit its implementation.

[0040] This application provides a deep learning-based system for identifying dangerous behaviors of personnel in aquaculture waters, which can achieve... Figures 1-2 The various processes and technical effects of the deep learning-based method for identifying dangerous behaviors of personnel in aquaculture waters, as described in the method embodiments, will not be repeated here to avoid duplication.

[0041] Optionally, this application embodiment also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of a method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0042] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of a method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0043] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0044] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0045] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0046] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for identifying hazardous behaviors of personnel in aquaculture waters based on deep learning, characterized in that, The method includes: Images of people in the water are acquired and preprocessed to obtain preprocessed images; a dataset is constructed based on the preprocessed images; A dangerous behavior recognition network is constructed, which includes a multi-scale feature extraction and interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; The dataset is divided into a training set and a test set according to a preset ratio, and the dangerous behavior recognition network is trained and tested based on the training set and the test set respectively to obtain a trained dangerous behavior recognition network. The system collects real-time monitoring images of the target aquaculture area and processes these images based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results.

2. The method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning according to claim 1, characterized in that, The process of processing the target surveillance image based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results includes: The target monitoring image is processed by the multi-scale feature extraction interaction module to obtain first-scale features, multi-scale interactive features and third-scale features respectively; The local enhancement module enhances the first-scale feature to obtain the local enhanced feature. The global semantic enhancement module is used to perform global semantic enhancement on the third-scale features to obtain globally semantically enhanced features. The feature fusion module fuses the local enhancement features, multi-scale interaction features, and global semantic enhancement features to obtain the target fusion feature. The detection head processes the target fusion features to obtain the dangerous behavior recognition result corresponding to the target monitoring image.

3. The method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning according to claim 2, characterized in that, The step of processing the target monitoring image according to the multi-scale feature extraction interaction module to obtain first-scale features, multi-scale interactive features, and third-scale features includes: The target monitoring image is subjected to two consecutive 3×3 convolutions to obtain basic information features, and the basic information features are then subjected to directional spatial attention enhancement to obtain directional spatial enhancement features. The directional spatial enhancement feature is processed by 3×3 convolution to obtain the first scale feature; the first scale feature is concatenated with the directional spatial enhancement feature to obtain the first concatenated and fused feature; The first splicing and fusion feature is processed by a 5×5 convolution to obtain the second scale feature; the second scale feature is spliced ​​with the first splicing and fusion feature to obtain the second splicing and fusion feature. The second splicing and fusion feature is convolved with 7×7 to obtain the third scale feature, and the second scale feature and the third scale feature are subjected to feature interaction processing to obtain multi-scale interactive features.

4. The method for identifying hazardous behaviors of personnel in aquaculture waters based on deep learning according to claim 3, characterized in that, The step of performing directional spatial attention enhancement on the basic information features to obtain directional spatially enhanced features includes: The basic information features are subjected to horizontal max pooling and average pooling respectively to obtain horizontal max pooling features and horizontal average pooling features; and the horizontal max pooling features and horizontal average pooling features are added together to obtain the horizontal pooling features; The basic information features are subjected to vertical max pooling and average pooling respectively to obtain vertical max pooling features and vertical average pooling features; and the vertical max pooling features and vertical average pooling features are added together to obtain vertical pooling features. The horizontal pooling features and vertical pooling features are concatenated to obtain a fused pooling feature; and the fused pooling feature is then subjected to 1×1 convolution and ReLU activation sequentially to obtain a first activation feature; The first activation feature is split along the channel dimension into a first segmentation feature and a second segmentation feature; The first segmentation feature is sequentially subjected to 1×1 convolution and Sigmoid activation to obtain the first weight; the second segmentation feature is sequentially subjected to 1×1 convolution and Sigmoid activation to obtain the second weight. The basic information features are multiplied sequentially by the first weight and the second weight to obtain the directional spatial enhancement features.

5. The method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning according to claim 4, characterized in that, The step of performing feature interaction processing on the second-scale features and the third-scale features to obtain multi-scale interactive features includes: The second scale feature is concatenated with the third scale feature to obtain the first concatenated feature; Spatial importance and channel importance models are performed on the first splicing feature respectively to obtain spatial importance features and channel importance features; The spatial importance feature and the channel importance feature are added together to obtain the fused importance feature; The fusion importance feature is combined with the first splicing feature to obtain the second splicing feature; The second spliced ​​feature is sequentially subjected to channel shuffling, 7×7 group convolution, and Sigmoid activation to obtain the fused attention weights; Multiply the fused attention weights by the second scale feature to obtain the second scale enhanced feature; The difference between 1 and the fused attention weight is calculated as a complementary weight. The complementary weight is then multiplied by the third-scale feature to obtain the third-scale enhanced feature. The second scale enhancement feature and the third scale enhancement feature are added together to obtain the multi-scale interactive feature.

6. The method for identifying dangerous behaviors of personnel in aquaculture waters based on deep learning according to claim 5, characterized in that, The step of enhancing the first-scale features according to the local enhancement module to obtain locally enhanced features includes: The first scale feature is sequentially subjected to 3×3 convolution, batch normalization and ReLU activation to obtain the first convolutional feature; The first convolutional feature is processed sequentially based on multiple concatenated residual blocks to obtain the residual enhancement feature output by the last residual block; The local enhancement features are obtained by performing 3×3 convolution, batch normalization, and ReLU activation on the residual enhancement features. The data processing procedure for each residual block includes: performing 3×3 convolution, batch normalization, ReLU activation, 3×3 convolution, and batch normalization sequentially on the input features of the residual block to obtain batch normalized features; adding the batch normalized features to the input features of the residual block to obtain residual fusion features; and performing ReLU activation on the residual fusion features to obtain the output features of the residual block.

7. A method for identifying hazardous behaviors of personnel in aquaculture waters based on deep learning, as described in claim 5, is characterized in that... The step of performing global semantic enhancement on the third-scale features based on the global semantic enhancement module to obtain globally semantically enhanced features includes: The third-scale features are subjected to global max pooling and global average pooling respectively to obtain global max pooling features and global average pooling features respectively; The global max pooling feature and the global average pooling feature are added together to obtain the global fusion feature. The global fusion feature is then processed by fully connected processing and ReLU activation in sequence to obtain the global context feature. The global context features are sequentially processed with fully connected layers and activated with Sigmoid to obtain global semantic awareness weights; and the global semantic awareness weights are multiplied by the third-scale features to obtain the global semantic enhancement features.

8. A deep learning-based system for identifying dangerous behaviors of personnel in aquaculture waters, capable of implementing the deep learning-based method for identifying dangerous behaviors of personnel in aquaculture waters as described in any one of claims 1-7, characterized in that, The system includes: Data construction module: used to acquire images of people in the water and preprocess them to obtain preprocessed images; and to construct a dataset based on the preprocessed images; Network building module: used to build a dangerous behavior recognition network, which includes a multi-scale feature extraction interaction module, a local enhancement module, a global semantic enhancement module, a feature fusion module, and a detection head; Network training module: used to divide the dataset into training set and test set according to a preset ratio, and to train and test the dangerous behavior recognition network based on the training set and the test set respectively, so as to obtain the trained dangerous behavior recognition network; Real-time monitoring module: used to collect target monitoring images of the target aquaculture water area in real time through the monitoring system, and process the target monitoring images based on the trained dangerous behavior recognition network to obtain dangerous behavior recognition results.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of a deep learning-based method for identifying dangerous behaviors of personnel in aquaculture waters as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of a deep learning-based method for identifying dangerous behaviors of personnel in aquaculture waters as described in any one of claims 1-7.

Images