Wild boar identification and monitoring system and method based on advanced vision algorithm

Abstract

The invention discloses a wild boar recognition and monitoring system and method based on an advanced vision algorithm, and belongs to the technical field of image recognition, and the system comprises an information collection device which is configured to collect an initial target image, and carries out the preprocessing and enhancement of the initial target image to obtain a second target image; the feature extraction module is configured to perform feature extraction on the second target image, optimize semantic expression and generate a multi-scale feature map; the lightweight detection head module comprises a classification task branch and a bounding box regression task branch; and the evaluation module is configured to optimize the detection result through a non-maximum value algorithm, and evaluate the model performance by taking the average precision mean value, the recall rate and the precision rate as evaluation indexes. According to the method, the rapid identification capability of the wild boars is remarkably improved, the method can be widely applied to real-time monitoring of wild boar invasion, farmland protection and emergency response scenes, and an efficient solution is provided for guaranteeing the safety of residents and preventing the farmland from being damaged.

Description

Technical Field

[0004] The present invention belongs to the technical field of image recognition, and particularly relates to a wild boar recognition and monitoring system and method based on an advanced vision algorithm. Background Art

[0005] With the increase in the number of wild boars, the damage they cause to farmland crops has become increasingly serious. Real-time monitoring and early warning reports of wild boar invasions have become an important requirement for agricultural protection. Traditional wild boar monitoring methods mostly rely on manual patrols or fixed monitoring devices, which are not only inefficient but also unable to meet the needs of large-scale farmland and complex mountain forest environments. Manual patrols also pose risks. Although some existing object detection algorithms have high detection accuracy, due to their high model complexity and large computational resource requirements, it is difficult to deploy them on unmanned aerial vehicle devices with limited computing power.

[0006] Therefore, a wild boar recognition and monitoring system and method based on an advanced vision algorithm are proposed to solve the above problems. Summary of the Invention

[0007] The purpose of the present invention is to improve the rapid recognition ability of wild boars, and a wild boar recognition and monitoring system and method based on an advanced vision algorithm are proposed.

[0008] To achieve the above purpose, the present invention adopts the following technical solutions:

[0009] A wild boar recognition and monitoring system based on an advanced vision algorithm, including

[0010] An information acquisition device, configured to acquire an initial target image, record geographical information in cooperation with a positioning device, and perform preprocessing and enhancement operations on the initial target image to obtain a second target image;

[0011] A feature extraction module, configured to extract features from the second target image, optimize semantic expressions, generate multi-scale feature maps, and introduce an attention mechanism to optimize the multi-scale feature maps;

[0012] A lightweight detection head module, including a classification task branch and a bounding box regression task branch. The classification task branch is configured to improve the accuracy of target category discrimination through depthwise separable convolution and activation function optimization; the bounding box regression task branch is configured to adopt a multi-scale prediction mechanism and a dynamic weight allocation strategy to ensure the target localization accuracy at different scales;

[0013] An evaluation module, configured to optimize the detection results through a non-maximum algorithm and evaluate the model performance using the mean average precision, recall rate, and precision rate as evaluation indicators.

[0014] Preferably, in the information acquisition device, preprocessing the target image includes adjusting the image size and normalizing the pixel values, and enhancing the target image includes using random rotation, cropping, and Mosaic data augmentation techniques.

[0015] Preferably, in the feature extraction module, the steps of feature extraction include: using a lightweight convolutional network as the backbone structure and introducing depthwise separable convolutions; extracting multi-scale feature maps layer by layer from the backbone network; where high-resolution features are used to capture local details of the target, and low-resolution features focus on the global semantic expression of the target, including the overall contour and population distribution features of the target.

[0016] Preferably, in the feature extraction module, the steps of optimizing the multi-scale feature maps include: combining channel and spatial lightweight ResCBAM attention mechanisms. The channel attention mechanism assigns different weights to each feature channel, highlighting features useful for target detection and suppressing irrelevant information; the spatial attention mechanism assigns weights to each position in the image, highlighting the region where the target is located and reducing background interference.

[0017] Preferably, in the lightweight detection head module, the classification task branch adopts an optimized feature compression mechanism, combines depthwise separable convolutions and a feature selection module, filters and reduces the dimensionality of redundant features in the feature map, and introduces a lightweight activation function.

[0018] Preferably, in the lightweight detection head module, the bounding box regression task branch introduces a dynamic weight allocation strategy, adaptively adjusts the feature extraction priority, and combines a multi-scale prediction mechanism to enhance the localization ability for targets of different scales.

[0019] Preferably, in the evaluation module, the steps of optimizing the detection results through the non-maximum algorithm specifically include: adopting an improved non-maximum suppression algorithm, which dynamically calculates the retention probability of each candidate box according to the target size and classification confidence through a dynamic threshold adjustment mechanism, and combines a context awareness mechanism to optimize the priority through the spatial distribution of the candidate boxes, introducing a multi-round confidence update strategy to eliminate redundant boxes and further optimize the confidence distribution of the retained boxes.

[0020] Preferably, in the evaluation module, the steps of detecting performance evaluation include: using the mean average precision as the core evaluation index, calculating the area under the precision-recall curve, and taking the average of all target classes to obtain the mean average precision.

[0021] A wild boar recognition and monitoring method based on advanced vision algorithms, including,

[0022] Step 1: Construct a dataset. Collect target images of multiple scenarios through information acquisition devices, record geographical information, perform preprocessing and enhancement operations on the target images, and construct a dataset;

[0023] Step 2: Build a backbone network using a lightweight convolutional network, perform efficient feature extraction on the target images, optimize the semantic expression of the feature maps, and generate multi-scale feature maps;

[0024] Step 3: Introduce an attention mechanism to optimize the multi-scale feature maps, align feature maps of different resolutions through upsampling and feature splicing operations, and strengthen channel and spatial attention characteristics;

[0025] Step 4: Design a lightweight detection head module, which includes a classification task branch and a bounding box regression task branch. The classification task branch is optimized through depthwise separable convolution and activation functions; the bounding box regression task branch uses a multi-scale prediction mechanism and a dynamic weight allocation strategy;

[0026] Step 5: Use an improved non-maximum suppression algorithm for the detection results to screen and optimize the candidate boxes; the improved non-maximum suppression algorithm calculates the retention probability of each candidate box dynamically according to the target size and classification confidence through a dynamic threshold adjustment mechanism; the detection performance evaluation uses the mean average precision as the core evaluation index to comprehensively measure the detection accuracy of the model.

[0027] In summary, the present invention has the following beneficial effects:

[0028] 1. Efficient target detection ability: By introducing a lightweight convolutional network and an attention mechanism, the present invention improves the target detection efficiency of the model. Especially in the real-time detection task of drones, it can quickly and accurately identify wild boar targets with fewer computing resources and adapt to the environment with limited hardware resources.

[0029] 2. Enhanced dataset construction: By using technologies such as GPS positioning, data augmentation, and fine annotation, a high-quality and diverse dataset covering multiple scenarios, different lighting conditions, and complex backgrounds is constructed. This dataset provides strong support for training a more robust model and can effectively improve the generalization ability of the model in wild boar detection.

[0030] 3. Optimized multi-scale feature extraction: Through a lightweight feature extraction network and the effective extraction of multi-scale feature maps, the present invention can better process the detailed information of wild boar targets at different scales. Compared with traditional methods, it has stronger expressiveness and accuracy in detecting large, medium, and small targets.

[0031] 4. Precise Target Location and Classification: Through the design of a lightweight detection head, the present invention not only effectively improves the accuracy of classification and regression tasks, but also reduces the computational complexity through a multi-task branch structure, ensuring the efficient output of detection results. This innovation can significantly enhance the performance of real-time drone detection and meet the actual application requirements in a large-scale environment.

[0032] 5. Enhanced Robustness and Real-time Performance: An improved non-maximum suppression algorithm is used to optimize the detection boxes, eliminating redundant boxes and adjusting the confidence distribution, further improving the accuracy and reliability of the detection results. In large-scale monitoring, it can quickly identify wild boars and respond, greatly enhancing the wild boar prevention and control capabilities in areas such as farmland and mountain forests.

[0033] 6. Ability to Adapt to Complex Scenarios: The introduced attention mechanism can optimize multi-scale feature maps, enhancing the model's adaptability in complex environments. For interference factors such as lighting, weather, and background changes in scenarios such as farmland and mountain forests, the model can still maintain a high detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a schematic flow chart of the method in the present invention;

[0035] Figure 2 is a schematic diagram of the network architecture in an embodiment of the present invention;

[0036] Figure 3 is a schematic diagram of the ResCBAM attention mechanism module in an embodiment of the present invention;

[0037] Figure 4 is a schematic diagram of the model training results in an embodiment of the present invention;

[0038] Figure 5 is a schematic diagram of the evaluation confusion matrix in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0040] A wild boar recognition and monitoring system based on an advanced vision algorithm includes

[0041] an information collection device configured to collect an initial target image, record geographical information in cooperation with a positioning device, and perform preprocessing and enhancement operations on the initial target image to obtain a second target image;

[0042] A feature extraction module, configured to extract features from a second target image, optimize semantic expressions, generate multi-scale feature maps, and introduce an attention mechanism to optimize the multi-scale feature maps;

[0043] A lightweight detection head module, including a classification task branch and a bounding box regression task branch. The classification task branch is configured to improve the accuracy of target category discrimination through depthwise separable convolution and activation function optimization; the bounding box regression task branch is configured to adopt a multi-scale prediction mechanism and a dynamic weight allocation strategy to ensure the target localization accuracy at different scales;

[0044] An evaluation module, configured to optimize the detection results through a non-maximum algorithm and evaluate the model performance using the mean average precision, recall rate, and precision rate as evaluation metrics.

[0045] In this embodiment, the information acquisition device includes a drone. Based on the original image data in scenarios such as farmland and mountain forests collected by the drone, combined with the geographical information recorded by GPS positioning, a basic dataset is formed; the dataset is processed, including preprocessing operations such as adjusting the image size and normalizing pixel values, and the data volume and sample complexity are expanded through techniques such as random rotation, cropping, and Mosaic data augmentation to enhance the diversity and applicability of the dataset.

[0046] By accurately annotating the enhanced image data, covering the target box information of wild boars under different scenarios and behavioral characteristics, a high-quality and diverse dataset is constructed to provide rich and accurate sample support for subsequent model training.

[0047] As Figure 2 shown, a backbone network using lightweight convolution is adopted for feature extraction. By optimizing the design of the convolutional layer and introducing depthwise separable convolution, the number of model parameters and computational complexity are significantly reduced, making it adaptable to the computing power limitations of drone hardware while maintaining efficient feature extraction capabilities.

[0048] This lightweight network can extract multi-scale feature maps layer by layer. The high-resolution feature maps are used to capture the fine details of wild boar targets, such as features like ears and tails; the low-resolution feature maps extract the global semantic information of the targets, such as the overall contour of wild boars or the distribution characteristics of groups of targets, thereby achieving balanced detection of targets of different sizes. In addition, the feature extraction network further optimizes the semantic expression of the feature maps in the convolutional layer. By hierarchically extracting target information, the accuracy and efficiency of feature extraction are effectively improved, providing high-quality feature inputs for subsequent multi-scale fusion and detection tasks, and ensuring the accuracy and robustness of the model.

[0049] In the feature extraction stage, a channel and spatial attention fusion mechanism based on the residual network is introduced, aiming to enhance the regional expression ability of wild boar targets, thereby improving the detection accuracy of the model. The channel attention mechanism weights each channel in the feature map, dynamically adjusts the contribution degree of different channels, and can automatically identify which feature channels contain important information related to wild boar targets, thus effectively suppressing the interference of irrelevant information. In this way, the network can give priority to features helpful for classification and localization, enhancing the discrimination ability for targets. The spatial attention mechanism weights each spatial position, focusing on the regions in the image containing wild boar targets. By learning the importance of different regions, this mechanism assigns higher weights to regions containing targets, enhancing the model's attention to the target positions, thereby improving the detection accuracy. Especially in complex scenarios, it can effectively reduce background interference and reduce the cases of false detection and missed detection. Through the above optimizations, the model performs more excellently when dealing with complex backgrounds, low-contrast images, or smaller targets, further improving the accuracy and robustness of wild boar visual detection. Especially in the dynamic environment captured by drones, it can ensure efficient and accurate target detection tasks.

[0050] As Figure 3 shown, combining the channel and spatial lightweight ResCBAM attention mechanisms, the multi-scale feature maps are dynamically optimized, and the specific formula is as follows:

[0051] F out = F + (M s (M c (F) · F) · (M c (F) · F))

[0052] Among them, F is the input feature map, M c (F) is the channel attention weight, and M s (F ′ ) is the spatial attention weight.

[0053] The channel attention mechanism assigns different weights to each feature channel, highlighting the features useful for wild boar target detection and suppressing irrelevant information, thereby enhancing the feature expression ability of the target region. The spatial attention mechanism assigns weights to each position in the image, focusing on the regions where wild boar targets are located, further reducing the impact of background interference on the detection performance.

[0054] The lightweight detection head module architecture achieves efficient object detection performance through the collaborative work of an independent classification task branch and a bounding box regression task branch: The classification task branch adopts an optimized feature compression mechanism, combines depthwise separable convolution and a feature selection module, screens and reduces the dimensionality of redundant features in the feature map, thereby efficiently extracting the semantic information of the object. In addition, by introducing a lightweight activation function, the computational overhead is further reduced, and the non-linear expression ability of the model is improved. The bounding box regression task branch introduces a dynamic weight allocation strategy for the accurate positioning of the object, adaptively adjusts the feature extraction priority, and combines a multi-scale prediction mechanism to enhance the positioning ability for objects of different scales, especially the detection effect for small objects is more significant. The detection head effectively reduces the number of parameters and computational complexity through grouped convolution and an optimized bottleneck design, significantly improving the detection speed and energy efficiency ratio, enabling the detection system to run in real time under the limited computing power of the drone while maintaining high detection accuracy.

[0055] Through an improved non-maximum suppression algorithm, redundant boxes are removed and the confidence distribution is adjusted, optimizing the detection results: This algorithm is based on dynamic threshold adjustment, comprehensively considering the object size, classification confidence, and scene context information, and dynamically calculates the retention probability of the candidate boxes. For small objects, the confidence weight is increased to avoid missed detections; for large objects, more attention is paid to suppressing redundant boxes to optimize the overall detection distribution. At the same time, the algorithm adds a context awareness mechanism to analyze the spatial position relationship of the candidate boxes, combines the semantic information of the multi-scale feature maps, and weights and optimizes the priority of the object boxes. This mechanism effectively reduces background interference and false detection phenomena in complex scenes. In addition, a multi-round confidence update strategy is introduced to optimize the confidence distribution of the remaining boxes while removing redundant boxes, further improving the reliability of the detection results. Experiments show that the improved NMS algorithm significantly improves the detection performance of small objects, reduces false detections and missed detections in complex scenes, and has a low computational overhead, enabling real-time processing of high-resolution images on the drone platform to meet the rapid response requirements of wild boar invasion monitoring.

[0056] The improved algorithm dynamically calculates the retention probability of each candidate box according to the object size and classification confidence through a dynamic threshold adjustment mechanism. The specific retention probability formula is:

[0057] P i =Conf i ·(1+IoU i,j )

[0058] Where P i represents the retention probability of candidate box i, Conf i is the classification confidence, and IoU i,j is the intersection over union of this box and other boxes.

[0059] In addition, the improved NMS algorithm combines a context awareness mechanism to optimize priorities through the spatial distribution of candidate bounding boxes and introduces a multi-round confidence update strategy. While removing redundant bounding boxes, it further optimizes the confidence distribution of the remaining bounding boxes, significantly improving detection accuracy and robustness.

[0060] The steps for evaluating detection performance include:

[0061] As Figure 5 shown, the mean average precision (mAP) is used as the core evaluation metric to comprehensively measure the detection accuracy of the model. mAP is calculated by computing the area under the precision-recall curve and taking the average of all target classes. The formula is as follows:

[0062]

[0063] where N is the number of target classes, and P(R) represents the precision at recall rate R. The mAP score of the model can comprehensively reflect the overall performance of object detection under different scenarios and target scales.

[0064] The experimental results of this embodiment are as Figure 4 shown. In the solution of this embodiment, both the training and validation losses decrease with the number of iterations, and the evaluation metrics increase synchronously, indicating that the model is not overfitting and the optimization process is stable. Among them, val / box_loss drops from 1.20 to about 0.6, and the localization performance on the validation set is consistent with that on the training set, without obvious overfitting. val / cls_loss drops from 2.5 to about 0.5, and finally converges well, with the precision basically remaining above 0.9 during the iteration. The data of mAP50(B) and mAP50-95(B) also demonstrate the excellent detection effect of this solution under different scenarios and target scales.

[0065] The technical solutions in the embodiments of the present application at least have the following technical effects or advantages: Through the design of a lightweight detection head, the present invention not only effectively improves the accuracy of classification and regression tasks, but also reduces the computational complexity through a multi-task branch structure, ensuring the efficient output of detection results. This innovation can significantly improve the performance of real-time drone detection and meet the actual application requirements in a large-scale environment.

[0066] As Figure 1 shown, the embodiments of the present application also provide a wild boar recognition and monitoring method based on advanced vision algorithms. The method includes the following steps:

[0067] Step 1: Collect wild boar target images in multiple scenarios such as farmland and mountain forests through drones, and construct a diverse and high-quality dataset in combination with data augmentation techniques.

[0068] Specifically, based on the original image data in scenarios such as farmland and mountain forests collected by drones, combined with the geographical information recorded by GPS positioning, a basic dataset is formed; the dataset is processed, including preprocessing operations such as adjusting the image size and normalizing pixel values, and the data volume and sample complexity are expanded through techniques such as random rotation, cropping, and Mosaic data augmentation to enhance the diversity and applicability of the dataset.

[0069] By accurately annotating the enhanced image data, covering the target box information of wild boars under different scenarios and behavioral characteristics, a high-quality and diverse dataset is constructed to provide rich and accurate sample support for subsequent model training. As Figure 2 shown.

[0070] Step 2: Use a lightweight convolutional network to efficiently extract multi-scale features of wild boars in complex backgrounds and enhance the ability to capture small wild boar targets in drone images.

[0071] Adopt a backbone network with lightweight convolution for feature extraction. By optimizing the design of the convolutional layer and introducing depthwise separable convolution, the number of model parameters and computational complexity are significantly reduced, making it adapt to the computing power limitations of drone hardware while maintaining efficient feature extraction capabilities.

[0072] This lightweight network can extract multi-scale feature maps layer by layer. The high-resolution feature maps are used to capture the fine details of wild boar targets, such as features like ears and tails; the low-resolution feature maps extract the global semantic information of the targets, such as the overall contour of wild boars or the distribution characteristics of grouped targets, thus achieving balanced detection of large and small targets. In addition, the feature extraction network further optimizes the semantic expression of the feature maps in the convolutional layer. By hierarchically extracting target information, the accuracy and efficiency of feature extraction are effectively improved, providing high-quality feature inputs for subsequent multi-scale fusion and detection tasks, and ensuring the accuracy and robustness of the model.

[0073] Step 3: Adopt multi-scale feature fusion technology to strengthen the semantic and detail expression of the target area and improve the detection adaptability in complex scenarios.

[0074] Adopt a backbone network with lightweight convolution for feature extraction. By optimizing the design of the convolutional layer and introducing depthwise separable convolution, the number of model parameters and computational complexity are significantly reduced, making it adapt to the computing power limitations of drone hardware while maintaining efficient feature extraction capabilities.

[0075] This lightweight network can extract multi-scale feature maps layer by layer. The high-resolution feature maps are used to capture the fine details of wild boar targets, such as features like ears and tails; the low-resolution feature maps extract the global semantic information of the targets, such as the overall contour of wild boars or the distribution features of groups of targets, thus achieving balanced detection of targets of different sizes. In addition, the feature extraction network further optimizes the semantic expression of the feature maps in the convolutional layer. By hierarchically extracting target information, it effectively improves the accuracy and efficiency of feature extraction, provides high-quality feature inputs for subsequent multi-scale fusion and detection tasks, and ensures the accuracy and robustness of the model.

[0076] Step 4: Design a lightweight detection head, combine the multi-task branches of classification and bounding box regression, and achieve accurate discrimination and positioning of wild boars and other targets.

[0077] In the design of the detection head, a lightweight structure is adopted to achieve efficient cooperation between classification and regression through task branch decoupling and feature optimization. The detection head includes independent classification and bounding box regression branches. Both share the lightweight backbone network during the feature extraction stage and are decoupled at the detection head part. The classification branch adopts an optimized feature compression mechanism, based on depthwise separable convolution (DSC) and the lightweight activation function (ReLU6). While reducing the computational overhead, it improves the semantic expression ability of the features, effectively enhancing the discrimination ability between wild boars and other target categories. The bounding box regression branch introduces a dynamic weight allocation strategy, adaptively adjusts the priority of feature extraction according to the size and position characteristics of the target, and at the same time adopts a multi-scale prediction mechanism, enabling accurate bounding box regression results to be generated for wild boar targets of different sizes. To further reduce the computational complexity and adapt to the computing power limitations of UAV hardware, the detection head module also adopts grouped convolution and an optimized bottleneck design, significantly reducing the number of parameters and floating-point operation counts. The detection head can efficiently complete the detection and positioning tasks, ensuring detection accuracy while achieving real-time operation.

[0078] Step 5: Apply an improved non-maximum suppression algorithm to the detection results to screen and optimize the candidate bounding boxes. The improved non-maximum suppression algorithm uses a dynamic threshold adjustment mechanism to dynamically calculate the retention probability of each candidate bounding box based on the object size and classification confidence. The detection performance evaluation uses the mean average precision as the core evaluation metric to comprehensively measure the detection accuracy of the model. The improved algorithm adaptively adjusts the retention probability of the candidate bounding boxes by comprehensively considering the object size, classification confidence, and scene context information. For small objects, the confidence weight is increased to avoid false deletion; for large objects, redundant bounding boxes are preferentially suppressed, thereby optimizing the overall distribution of the detection results. The algorithm also introduces a context awareness mechanism that uses the spatial position relationship of the candidate bounding boxes and the semantic information of the multi-scale feature maps to weighted optimize the priority of the object bounding boxes, effectively reducing the false detection phenomenon caused by background interference in complex scenes. When processing the candidate bounding boxes, a multi-round confidence update strategy is adopted, which not only eliminates redundant bounding boxes but also optimizes the confidence distribution of the retained bounding boxes, significantly improving the reliability of the detection results. In terms of performance evaluation, the mean average precision (mAP) calculated based on precision and recall is used as a comprehensive metric to comprehensively measure the performance of the model in classification and localization tasks. This algorithm performs excellently in small object detection and complex scene adaptability and can achieve efficient real-time detection with low computational overhead under the limited hardware resources of the drone, meeting the rapid response requirements of wild boar intrusion monitoring.

[0079] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages: significantly reducing the number of model parameters and computational complexity, making it adapt to the computing power limitations of the drone hardware, while maintaining an efficient feature extraction ability. Especially in complex scenes, it can effectively reduce background interference and reduce the cases of false detection and missed detection. Through the above optimizations, the model performs more excellently when dealing with complex backgrounds, low-contrast images, or smaller objects, further improving the accuracy and robustness of wild boar visual detection. Especially in the dynamic environment captured by the drone, it can ensure efficient and accurate completion of the object detection task.

[0080] The working principle is as follows: A lightweight convolutional network is used to efficiently extract features from images, layer by layer optimizing the semantic expression of feature maps, generating multi-scale feature maps that take into account both high and low resolutions. An attention mechanism is introduced to optimize the multi-scale feature maps. Different-resolution feature maps are aligned through upsampling and feature splicing operations, strengthening the channel and spatial attention characteristics, enhancing the feature expression ability of the wild boar target area. A lightweight detection head module is designed, and the detection efficiency is improved through independent classification and bounding box regression task branches. At the same time, the computational complexity is optimized to adapt to the computing power limitations of the drone hardware. The detection results are optimized through an improved non-maximum suppression algorithm, removing redundant boxes and adjusting the confidence distribution. Standard metrics such as the mean average precision (mAP) are used to comprehensively evaluate the model performance, ensuring the accuracy and robustness of the detection results.

[0081] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A wild boar identification and monitoring system based on advanced visual algorithms, characterized in that: include, The information collection device is configured to collect an initial target image, cooperate with the positioning device to record geographic information, and perform preprocessing and enhancement operations on the initial target image to obtain a second target image; A feature extraction module is configured to extract features from the second target image, optimize semantic expression, generate a multi-scale feature map, and introduce an attention mechanism to optimize the multi-scale feature map; A lightweight detection head module, including a classification task branch and a bounding box regression task branch. The classification task branch is configured to improve the accuracy of target category distinction through deep separable convolution and activation function optimization; the bounding box regression task branch is configured to adopt a multi-scale prediction mechanism and a dynamic weight allocation strategy to ensure the accuracy of target positioning at different scales; The evaluation module is configured to optimize the detection results through a non-maximum algorithm and evaluate the model performance using average precision, recall rate and precision as evaluation indicators.

2. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: In the information acquisition device, preprocessing the initial target image includes adjusting the image size and normalizing pixel values, and enhancing the initial target image includes adopting random rotation, cropping and Mosaic data enhancement technology.

3. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: In the feature extraction module, the feature extraction step includes: The main structure is a lightweight convolutional network, and depth-wise separable convolution is introduced; The backbone network is used to extract multi-scale feature maps layer by layer; the high-resolution features are used to capture the local detail information of the target, and the low-resolution features focus on the global semantic expression of the target, including the overall outline of the target and the group distribution characteristics.

4. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: In the feature extraction module, the step of optimizing the multi-size feature map includes: combining channel and spatial lightweight ResCBAM attention mechanisms, the channel attention mechanism assigns different weights to each feature channel, highlights the features useful for target detection and suppresses irrelevant information; the spatial attention mechanism assigns weights to each position in the image, highlights the target area, and reduces background interference.

5. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: The classification task branch in the lightweight detection head module adopts an optimized feature compression mechanism, combines deep separable convolution and feature selection modules, screens and reduces the dimensionality of redundant features in multi-size feature maps, and introduces a lightweight activation function.

6. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: The bounding box regression task branch in the lightweight detection head module introduces a dynamic weight allocation strategy, adaptively adjusts the feature extraction priority, and combines the multi-scale prediction mechanism to enhance the positioning capability of targets of different scales.

7. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: In the evaluation module, the step of optimizing the detection results through the non-maximum algorithm specifically includes: adopting an improved non-maximum suppression algorithm, which dynamically calculates the retention probability of each candidate box according to the target size and classification confidence through a dynamic threshold adjustment mechanism, and combines the context-aware mechanism to optimize the priority through the spatial distribution of the candidate boxes, introduces a multi-round confidence update strategy, eliminates redundant boxes and further optimizes the confidence distribution of the retained boxes.

8. The wild boar identification and monitoring system based on advanced visual algorithms according to claim 1 is characterized in that: In the evaluation module, the step of detecting performance evaluation includes: using the mean average precision as the core evaluation indicator, calculating the area under the precision-recall curve, and taking the average value of all target categories to obtain the mean average precision.

9. A wild boar identification and monitoring method based on advanced visual algorithms, characterized in that: include: Step 1: Construct a data set. Collect target images of multiple scenes and record geographic information through information collection equipment. Perform preprocessing and enhancement operations on the target images and construct a data set. Step 2: Use a lightweight convolutional network to build a backbone network, extract features of the target image efficiently, optimize the semantic expression of the feature map, and generate a multi-scale feature map; Step 3: Introduce the attention mechanism to optimize the multi-scale feature maps, align the feature maps of different resolutions through upsampling and feature concatenation operations, and strengthen the channel and spatial attention characteristics; Step 4: Design a lightweight detection head module, which includes a classification task branch and a bounding box regression task branch. The classification task branch is optimized by depthwise separable convolution and activation function; the bounding box regression task branch uses a multi-scale prediction mechanism and a dynamic weight allocation strategy. Step 5: Use an improved non-maximum suppression algorithm to screen and optimize the candidate boxes for the detection results; the improved non-maximum suppression algorithm dynamically calculates the retention probability of each candidate box according to the target size and classification confidence through a dynamic threshold adjustment mechanism; the detection performance evaluation uses the average precision mean as the core evaluation indicator to comprehensively measure the detection accuracy of the model.

Images