Wild animal target detection method based on lightweight structure and computing device
By introducing a sensitivity-aware attention module and a lightweight structure into the YOLO model, the problems of insufficient computing power and environmental noise interference in wildlife monitoring in the wild are solved, achieving efficient wildlife target detection and reducing computing costs and latency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CENTRAL SOUTH UNIVERSITY OF FORESTRY AND TECHNOLOGY
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-12
AI Technical Summary
Existing YOLO models have limited computing power and memory resources in wildlife monitoring scenarios, resulting in high latency, rapid power consumption, and high noise in complex environments. Wildlife target features are masked by secondary background features, limiting the accuracy of traditional models in learning to detect important targets.
We adopt a wildlife target detection method based on a lightweight structure, which introduces a sensitivity-aware attention module, a SlimNeck lightweight structure, and an adaptive statistical focus loss function. By integrating a SAAM attention module at the end of the backbone network, combined with the SlimNeck lightweight structure and the ASFL loss function, we optimize feature extraction and loss function to adapt to complex field environments.
It improves the model's ability to perceive wildlife targets, reduces computational burden, enhances detection accuracy, and meets the real-time detection needs of edge devices.
Smart Images
Figure CN122024326A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and computing device for detecting wild animal targets based on a lightweight structure. Background Technology
[0002] With the increasing global efforts to protect biodiversity, infrared camera technology, as a non-invasive survey method, has been widely deployed and applied in various nature reserves. This technology boasts advantages such as all-weather operation, high concealment, and minimal disturbance to animals, enabling long-term, unattended capture of images of rare and nocturnal wild animals. This has propelled wildlife monitoring from traditional field surveys into the era of "big data" monitoring, accumulating massive amounts of impact data. However, this explosive growth in data has also brought enormous challenges to data processing. Faced with millions of data transmissions from field monitoring, traditional manual screening and labeling methods have shown significant limitations: on the one hand, due to the complexity of the field environment, "aerial" images triggered by factors such as wind rustling and changes in light and shadow account for a very high proportion, leading to a significant waste of human resources; on the other hand, manual review is not only time-consuming, labor-intensive, and costly, but also prone to omissions or misjudgments due to visual fatigue from prolonged work, resulting in serious data delays. Therefore, utilizing advanced computer vision technology to achieve automated wildlife target detection is of paramount importance for reducing labor costs and improving the efficiency of biodiversity surveys.
[0003] Currently, single-stage object detection algorithms, represented by the YOLO (You Only Look Once) series, are widely used in fields such as industrial defect detection, autonomous driving, and facial recognition due to their end-to-end inference mode and good real-time performance. However, when directly transferring the general YOLO model to the field wildlife monitoring scenario, some problems still arise. Field equipment often has limited computing power and memory resources, and the model places an excessive burden on edge devices, resulting in high latency and rapid power consumption, making it difficult to meet the real-time monitoring needs in the field. The complex environment in the field is noisy, and the characteristics of wildlife targets are masked by secondary background features. Traditional models are dominated by a large number of simple samples in their learning direction, thus limiting the accuracy of important target detection.
[0004] Therefore, a technical solution is needed that can effectively extract the features of wild animal target images and meet the real-time detection needs of edge devices. Summary of the Invention
[0005] This invention aims to propose a wildlife target detection method and computing device based on a lightweight structure. To address the difficulties in feature extraction caused by complex backgrounds and strong environmental noise in the wild, this invention proposes an attention module, introduces a lightweight neck structure, and an adaptive statistical focus loss function. This solves the problems of limited computing power of edge computing devices in the wild, excessive computational burden of traditional models, and insufficient model learning of important features due to the extreme imbalance between easy and difficult samples in wildlife data.
[0006] According to one aspect of the present invention, a method for detecting wildlife targets based on a lightweight structure is provided, the method being executed on an edge computing device, the method comprising: Acquire image data captured by an outdoor infrared camera; The image data is preprocessed; The pre-trained target detection model is used to perform real-time target detection on the preprocessed image data, and the output includes the target category, predicted bounding box coordinates and prediction confidence. in, The target detection model is based on the YOLO11 model and includes a backbone network, a neck network, and a detection head. A sensitivity-aware attention module is integrated at the output of the backbone network to perform channel-level weighting on the input feature map. The sensitivity-aware attention module calculates the minimum energy value of a single neuron using an adaptive regularization coefficient based on channel variance, maps the reciprocal of the minimum energy value to attention weights using a Sigmoid activation function, and uses these attention weights to weight the sensitivity of the input feature map.
[0007] According to some embodiments, the attention weights are obtained using the following formula:
[0008] in, The input feature map, The feature map after attention weighting. E For all The energy matrix, where ⊙ denotes element-wise multiplication. For adaptive regularization coefficients, These are the preset base regularization parameters. The variance of neurons within the channel. To calculate the minimum energy value, t The value of the target neuron. This represents the average value of neurons within the channel.
[0009] According to some embodiments, the neck network is a SlimNeck lightweight structure, which includes a feature fusion path based on GSConv convolution and VoV-GSCSP module, used to retain multi-scale semantic information while reducing computational cost.
[0010] According to some embodiments, the GSConv convolution operation in the SlimNeck lightweight structure divides the input channel into two parts, performs standard convolution and depthwise separable convolution respectively, and then merges the output through channel splicing and shuffling operations.
[0011] According to some embodiments, the VoV-GSCSP module adopts a single-path dense connection structure, which contains multiple cascaded GSBottleneck units and runs in parallel with a direct connection branch that retains the original input. Finally, the two feature paths are concatenated and compressed by a 1×1 convolution before being output.
[0012] According to some embodiments, the detection head utilizes the ASFL adaptive statistical focus loss function.
[0013] According to some embodiments, the ASFL adaptive statistical focus loss function constructs an adaptive decision threshold using the statistical characteristics of the current batch:
[0014] in, This represents the mean of the predicted probabilities for the current batch. For the first batch The predicted probability of a sample. The total number of samples in the batch. The focusing parameters of the sample. Focusing on hyperparameters as a basis, For adjustment coefficients, For class balancing parameters, The predicted probability of the sample. This is the final loss value.
[0015] According to some embodiments, the edge computing device includes an infrared-triggered camera deployed in the field, an embedded AI box, or a low-power mobile terminal.
[0016] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method as described in any of the preceding claims.
[0017] According to another aspect of the present invention, a computing device is provided, comprising: Processor; and A memory that stores a computer program, which, when executed by the processor, implements the method as described in any of the preceding methods.
[0018] According to embodiments of the present invention, target detection of wild animals is performed by integrating a Sensitivity Awareness Attention (SAAM) module at the end of the backbone network. This addresses the common problems of low contrast, blurriness, and cluttered backgrounds in outdoor infrared images. The model can automatically identify and enhance feature channels that are sensitive to wild animal targets, suppress secondary features dominated by environmental noise, and improve the model's ability to perceive wild animal targets in complex environments.
[0019] Based on some implementations, the SlimNeck lightweight structure is introduced to reduce the number of parameters and computational burden while maintaining high accuracy. An Adaptive Statistical Focus Loss (ASFL) is proposed to address the problem of insufficient learning of the target's main features due to imbalance between easy and difficult samples. This solves the problems of traditional loss functions where secondary features dominate gradient updates, and the large number of model parameters and heavy computational burden that cannot meet the real-time detection requirements of edge devices.
[0020] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit the invention. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0022] Figure 1 A flowchart of a wildlife target detection method based on a lightweight structure, according to an example embodiment, is shown.
[0023] Figure 2 Images captured by an infrared camera for detecting wildlife are shown according to some embodiments.
[0024] Figure 3 This diagram illustrates the overall architecture of a wildlife target detection network based on an example embodiment.
[0025] Figure 4 A schematic diagram of the composition of the SAAM attention module according to an example embodiment is shown.
[0026] Figure 5 A schematic diagram of the structure of GSConv convolution according to an example embodiment is shown.
[0027] Figure 6 A schematic diagram of the structure of a VoV-GSCSP module according to an example embodiment is shown.
[0028] Figure 7A schematic diagram of the GSBottleneck module according to an example embodiment is shown.
[0029] Figure 8 A schematic diagram showing actual detection results according to an example embodiment.
[0030] Figure 9 A block diagram of a computing device according to an exemplary embodiment is shown. Detailed Implementation
[0031] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that the invention will be thorough and complete, and the concept of the exemplary embodiments will be fully conveyed to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.
[0032] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.
[0033] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.
[0034] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.
[0035] It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of the present invention. As used herein, the term "and / or" includes all combinations of any one and more of the associated listed items.
[0036] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily essential for implementing the present invention, and therefore cannot be used to limit the scope of protection of the present invention.
[0037] With the increasing global efforts to protect biodiversity, infrared camera technology, as a non-invasive survey method, has been widely deployed and applied in various nature reserves. This technology boasts advantages such as all-weather operation, high concealment, and minimal disturbance to animals, enabling long-term, unattended capture of images of rare and nocturnal wild animals. This has propelled wildlife monitoring from traditional field surveys into the era of "big data" monitoring, accumulating massive amounts of impact data. However, this explosive growth in data has also brought enormous challenges to data processing. Faced with millions of data transmissions from field monitoring, traditional manual screening and labeling methods have shown significant limitations: on the one hand, due to the complexity of the field environment, "aerial" images triggered by factors such as wind rustling and changes in light and shadow account for a very high proportion, leading to a significant waste of human resources; on the other hand, manual review is not only time-consuming, labor-intensive, and costly, but also prone to omissions or misjudgments due to visual fatigue from prolonged work, resulting in serious data delays. Therefore, utilizing advanced computer vision technology to achieve automated wildlife target detection is of paramount importance for reducing labor costs and improving the efficiency of biodiversity surveys.
[0038] Currently, single-stage object detection algorithms, represented by the YOLO (You Only Look Once) series, are widely used in fields such as industrial defect detection, autonomous driving, and facial recognition due to their end-to-end inference mode and good real-time performance. However, when directly transferring the general YOLO model to the field wildlife monitoring scenario, some problems still arise. Field equipment often has limited computing power and memory resources, and the model places an excessive burden on edge devices, resulting in high latency and rapid power consumption, making it difficult to meet the real-time monitoring needs in the field. The complex environment in the field is noisy, and the characteristics of wildlife targets are masked by secondary background features. Traditional models are dominated by a large number of simple samples in their learning direction, thus limiting the accuracy of important target detection.
[0039] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application.
[0040] Figure 1 A flowchart of a wildlife target detection method based on a lightweight structure, according to an example embodiment, is shown.
[0041] This embodiment is based on a lightweight wildlife target detection method executed on an edge computing device, which includes an infrared trigger camera deployed in the field, an embedded AI box, or a low-power mobile terminal.
[0042] See Figure 1 In S101, image data acquired by an outdoor infrared camera is obtained.
[0043] According to some embodiments, data collection is performed to acquire image data captured by field infrared cameras, resulting in a massive collection of unprocessed wildlife images (see [link]). Figure 2 ).
[0044] In S103, the image data is preprocessed.
[0045] According to some implementations, the original images are manually screened to remove blurry, overexposed, and negative sample images that do not contain valid detection targets, thus ensuring the quality of the dataset.
[0046] In S105, a pre-trained target detection model is used to perform real-time target detection on the pre-processed image data, and the output includes the target category, predicted bounding box coordinates and prediction confidence.
[0047] According to an example embodiment, the target detection model is based on the YOLO11 model and includes a backbone network, a neck network, and a detection head. A sensitivity-aware attention module is integrated at the output of the backbone network to perform channel-level weighting on the input feature map. The sensitivity-aware attention module calculates the minimum energy value of a single neuron using an adaptive regularization coefficient based on channel variance, maps the reciprocal of the minimum energy value to an attention weight using a Sigmoid activation function, and uses the attention weight to weight the sensitivity of the input feature map.
[0048] According to some embodiments, the attention weights are obtained using the following formula:
[0049] in, The input feature map, The feature map after attention weighting. E For all The energy matrix, where ⊙ denotes element-wise multiplication. For adaptive regularization coefficients, These are the preset base regularization parameters. The variance of neurons within the channel. To calculate the minimum energy value, t The value of the target neuron. This represents the average value of neurons within the channel.
[0050] The neck network in this embodiment is a SlimNeck lightweight structure. This SlimNeck structure includes a feature fusion path based on GSConv convolution and a VoV-GSCSP module, used to retain multi-scale semantic information while reducing computational cost. The GSConv convolution operation in the SlimNeck structure divides the input channel into two parts, performing standard convolution and depthwise separable convolution respectively, and then fusing the output through channel concatenation and shuffling operations. The VoV-GSCSP module adopts a single-path dense connection structure, containing multiple cascaded GSBottleneck units, and runs in parallel with a direct connection branch that retains the original input. Finally, the two feature paths are concatenated and compressed using a 1×1 convolution before output.
[0051] According to some embodiments, the detection head utilizes the ASFL adaptive statistical focus loss function. The ASFL adaptive statistical focus loss function constructs an adaptive decision threshold using the statistical characteristics of the current batch.
[0052] in, This represents the mean of the predicted probabilities for the current batch. For the first batch The predicted probability of a sample. The total number of samples in the batch. The focusing parameters of the sample. Focusing on hyperparameters as a basis, For adjustment coefficients, For class balancing parameters, The predicted probability of the sample. This is the final loss value.
[0053] Figure 3 This diagram illustrates the overall architecture of a wildlife target detection network based on an example embodiment.
[0054] According to some embodiments, a target detection network based on adaptive loss and lightweight structure is constructed; see [link / reference]. Figure 3The target detection network comprises a backbone, a neck, and a head. The backbone extracts multi-scale general features from the input image, integrating C3k2, SPPF, and C2PSA modules. The neck is responsible for fusing and enhancing the multi-scale features. Based on PANet structural components, it performs cross-layer aggregation and enhancement on the multi-scale features output from the backbone network to enrich the semantic information of the feature maps. The head is responsible for calculating the final target class probability and location coordinates from the fused features, and outputting these final target class probabilities and location coordinates.
[0055] The original images were screened, and blurry, overexposed, and negative sample images that did not contain valid detection targets were removed to ensure dataset quality. Animal targets in the images were manually labeled using image annotation software to generate corresponding label files, and the dataset was then formatted to conform to the standard input format of the YOLO algorithm. The cleaned dataset was then randomly divided into training, validation, and test sets according to a preset ratio (preferably 8:1:1) for model training, hyperparameter tuning, and performance evaluation, respectively.
[0056] This invention improves upon existing detection networks by introducing the proposed SAAM attention module at the tail of the backbone network (see [link]). Figure 4 The neck structure was replaced with a lightweight neck structure, SlimNeck, and the original CIoU loss was replaced with adaptive statistical focus loss (ASFL), thus obtaining a wildlife target detection model.
[0057] According to some embodiments, in order to dynamically balance the relationship between target enhancement and noise suppression in the energy function solution, the backbone constructs an adaptive regularization coefficient based on channel variance:
[0058] in, For self The adaptive regularization coefficient is 0.5, which is a constant. Using the preset basic regularization parameters, the minimum energy value of a single neuron is calculated using the adaptive regularization coefficients.
[0059] use To represent importance, the inverse of the energy value is ultimately mapped to attention weights through the Sigmoid activation function, thus completing the sensitivity-perceived weighted enhancement of the original features: , in, It is all The energy matrix; ⊙ represents element-wise multiplication; and These are the feature maps before and after attention weighting, respectively.
[0060] According to some embodiments, the lightweight neck structure SlimNeck includes GSConv convolutions and VoV-GSCSP modules. See GSConv convolutions for details. Figure 5 The specific process is as follows: The input and output feature maps are respectively and .
[0061] After half-channel standard convolution and depthwise separable convolution: , in, This represents the standard convolution operation. The feature map is the result of standard convolution; This indicates a depthwise separable convolution operation. This is the feature map after depthwise separable convolution.
[0062] Standard convolutional features that enhance semantics are concatenated with depthwise separable convolutional features along the channel dimension and then mixed across channels, as follows: , Where Concat represents splicing along the channel, X cat This is the spliced feature map; Shuffle indicates channel shuffling.
[0063] According to some embodiments, the VoV-GSCSP module is described below. Figure 6 The specific process is as follows: Input feature map Parallel through two independent convolutional layers and This is mapped into two branches, representing shallow mapping information and deep features respectively, as follows: , According to some embodiments, the output of the main branch is... Through stacked GSBottleneck modules (see) Figure 7 ): , Output branch one The output after processing branch two The feature maps are then concatenated along the channel dimension, and finally the concatenated feature maps are... After another convolutional layer The fusion will be carried out as follows: , According to some embodiments, the adaptive statistical focus loss (ASFL) of the detection head introduces an adaptive threshold mechanism based on batch statistical features and a segmented adaptive weighting strategy to adjust the weights of easy and difficult samples respectively, and uses the mean of the predicted probability of the current batch as the adaptive threshold for determining the ease or difficulty of samples.
[0064] Mean of the predicted probability for the current batch: , in, This represents the mean of the predicted probabilities for the current batch. For the first batch The predicted probability of a sample. This represents the total number of samples in the batch.
[0065] To specifically enhance the model's focus on hard-to-classify samples, a segmented focusing strategy based on dynamic thresholds is implemented, as follows: , in, The focusing parameters of the sample. Focusing on hyperparameters as a basis, For adjustment coefficients, The predicted probability of the sample is given; for difficult samples with predicted probabilities below the mean, the focusing parameter increases linearly with the prediction bias.
[0066] The adaptive statistical focus loss function is as follows: , in, For class balancing parameters, This represents the predicted probability of the sample.
[0067] Preprocessed image data is input into the model for iterative training. For example, the total number of training epochs is set to 100, the initial learning rate to 0.01, and the batch size to 32. The training process uses the backpropagation algorithm combined with a stochastic gradient descent (SGD) optimizer to update the network weights: in each training epoch, the gradient of the loss function for the current batch of data is calculated and the model parameters are updated, while the model performance is evaluated on the validation set simultaneously. An optimal model saving mechanism is introduced: if the validation loss of the current epoch is lower than the historical minimum, the current weights are overwritten and saved as the optimal weights; otherwise, the original weights remain unchanged. After training, the optimal weights throughout the entire process are loaded as the final model to ensure that the model has the best generalization ability and stability.
[0068] To comprehensively verify the superiority of the proposed method in wildlife detection tasks, this embodiment performs wildlife target detection based on adaptive loss and lightweight network architecture, and selects mainstream target detection models such as YOLOv5s, YOLOv8s, YOLO11s, YOLOv12s, and RT-DETR-R34 for comparison. Under the same wildlife image dataset and the same training and testing environment, the performance of each model was quantitatively analyzed. Evaluation metrics include precision, recall, and mean average precision (mAP), representing detection accuracy, and the number of parameters and computational cost (GFLOPs), representing model complexity. Detailed performance comparison data for each target detection model are shown in Table 1, and the detection results are as follows: Figure 8 As shown (the output box displays the animal category and confidence level).
[0069] Table 1
[0070] As shown in Table 1, compared to the baseline model YOLO11s, the method in this embodiment improves mAP50 by 0.66% (reaching 91.89%) and mAP50:95 by 0.45% (reaching 64.57%), demonstrating that its adaptive statistical focus loss (ASFL) and SAAM effectively improve the model training and tuning process, enhance the extraction of important features, and thus improve detection accuracy. Simultaneously, it reduces model complexity, decreasing the number of parameters by approximately 13.8% (9.4M to 8.1M) and effectively reducing computation by 25.7% (21.4 to 15.9), the lowest among all models. This indicates that the lightweight structure method will significantly outperform other models in terms of inference speed and power consumption on edge devices. Although YOLOv8s and RT-DETR-R34 models have slight advantages in precision and recall, respectively, their computational resource consumption is relatively high. The computational cost and parameter count of YOLOv8s are approximately 1.4 times and 1.8 times that of the method presented in this paper, respectively, while RT-DETR-R34's are even higher, at approximately 3.7 times and 5.6 times, severely limiting its practical deployment capabilities. As the above analysis shows, the wildlife target detection method based on a lightweight structure effectively reduces the computational cost burden while maintaining and improving detection accuracy, demonstrating the high practical value of this invention in resource-constrained real-time wildlife monitoring tasks.
[0071] Figure 9 A block diagram of a computing device according to an exemplary embodiment of the present invention is shown.
[0072] like Figure 9As shown, the computing device 30 includes a processor 12 and a memory 14. The computing device 30 may also include a bus 22, a network interface card 16, and an I / O interface 18. The processor 12, memory 14, network interface card 16, and I / O interface 18 can communicate with each other via the bus 22. The computing device of this invention can be applied to drones, autonomous vehicles, or monitoring equipment.
[0073] Processor 12 may include one or more general-purpose CPUs (Central Processing Units), microprocessors, or application-specific integrated circuits, for executing relevant program instructions. According to some embodiments, computing device 30 may also include a high-performance display adapter (GPU) 20 for accelerating processor 12.
[0074] Memory 14 may include a machine system readable medium in the form of volatile memory, such as random access memory (RAM), read-only memory (ROM), and / or cache memory. Memory 14 is used to store one or more programs containing instructions, as well as data. Processor 12 may read the instructions stored in memory 14 to perform the methods described above according to embodiments of the present invention.
[0075] The computing device 30 can also communicate with one or more networks via the network interface card 16. The network interface card 16 can be a DPU smart network card.
[0076] Bus 22 can include address bus, data bus, control bus, etc. Bus 22 provides a path for exchanging information between components.
[0077] It should be noted that, in specific implementations, the computing device 30 may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the device described above may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.
[0078] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), network storage devices, cloud storage devices, or any type of medium or device suitable for storing instructions and / or data.
[0079] This invention also provides a computer program product comprising a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments.
[0080] Those skilled in the art will clearly understand that the technical solutions of this application can be implemented using software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware capable of independently performing or cooperating with other components to perform a specific function, where the hardware may be, for example, a field-programmable gate array (FPGA), integrated circuit, etc.
[0081] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
[0082] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0083] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.
[0084] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0085] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0086] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.
[0087] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0088] The exemplary embodiments of this application have been specifically shown and described above. It should be understood that this application is not limited to the detailed structures, arrangements, or implementation methods described herein; rather, this application is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended provisions.
Claims
1. A wildlife target detection method based on a lightweight structure, the method being executed on an edge computing device, characterized in that, The method includes: Acquire image data captured by an outdoor infrared camera; The image data is preprocessed; The pre-trained target detection model is used to perform real-time target detection on the preprocessed image data, and the output includes the target category, predicted bounding box coordinates and prediction confidence. in, The target detection model is based on the YOLO11 model and includes a backbone network, a neck network, and a detection head. A sensitivity-aware attention module is integrated at the output of the backbone network to perform channel-level weighting on the input feature map. The sensitivity-aware attention module calculates the minimum energy value of a single neuron using an adaptive regularization coefficient based on channel variance, maps the reciprocal of the minimum energy value to attention weights using a Sigmoid activation function, and uses these attention weights to weight the sensitivity of the input feature map.
2. The method according to claim 1, characterized in that, The attention weights are obtained using the following formula: in, The input feature map, The feature map after attention weighting. E For all The energy matrix, where ⊙ denotes element-wise multiplication. For adaptive regularization coefficients, These are the preset base regularization parameters. Let V be the variance of neurons within the channel. To obtain the minimum energy value, t The value of the target neuron. This represents the average value of neurons within the channel.
3. The method according to claim 1, characterized in that, The neck network is a SlimNeck lightweight structure, which includes a feature fusion path based on GSConv convolution and VoV-GSCSP module, used to retain multi-scale semantic information while reducing computational cost.
4. The method according to claim 3, characterized in that, The GSConv convolution operation in the SlimNeck lightweight structure divides the input channel into two parts, performs standard convolution and depthwise separable convolution respectively, and then merges them into the output through channel splicing and shuffling operations.
5. The method according to claim 3, characterized in that, The VoV-GSCSP module adopts a single-path dense connection structure, which contains multiple cascaded GSBottleneck units and runs in parallel with the direct connection branch that retains the original input. Finally, the two feature paths are concatenated and compressed by 1×1 convolution.
6. The method according to claim 1, characterized in that, The detection head utilizes the ASFL adaptive statistical focus loss function.
7. The method according to claim 6, characterized in that, The ASFL adaptive statistical focus loss function utilizes the statistical characteristics of the current batch to construct an adaptive decision threshold: in, This represents the mean of the predicted probabilities for the current batch. For the first batch The predicted probability of a sample. The total number of samples in the batch. The focusing parameters of the sample. Based on focusing hyperparameters, For adjustment coefficients, For class balancing parameters, The predicted probability of the sample. This is the final loss value.
8. The method according to claim 1, characterized in that, The edge computing devices include infrared trigger cameras deployed in the field, embedded AI boxes, or low-power mobile terminals.
9. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-8.
10. A computing device, characterized in that, include: processor; as well as A memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1-8.