Cervus nippon image automatic labeling method, system and device based on iterative target detection model

Abstract

The application discloses a kind of sika deer image automatic labeling method, system and equipment based on iterative target detection model, belong to computer vision, deep learning and wisdom livestock technology field.The method includes the following steps: S1: obtaining the original image of sika deer to be labeled, and constructing initial dataset containing a small part of samples;S2: input the image into SAL-YOLO improved target detection model, and obtain target prediction result;S3: input the prediction result into automatic labeling script, and convert it into label file in standard format according to preset rule;S4: manually check and modify the label file, and incorporate the checked data into the original dataset to realize data expansion;S5: load the model weight of last generation, and use the expanded dataset to iteratively fine-tune the improved target detection model, and obtain the final version of automatic labeling model after three iterations.The application suppresses false detection from the source by deep integration of algorithm architecture and active learning closed loop, and improves labeling efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning, and smart animal husbandry, specifically to an automatic image annotation method, system, and device for sika deer based on an iterative target detection model. Background Technology

[0002] With the rapid development of smart agriculture and precision animal husbandry, animal individual recognition and behavior monitoring based on computer vision has become a research hotspot. In large-scale sika deer farming, object detection algorithms have shown great potential for industrial automation, but training highly robust visual models relies on a large amount of high-quality image annotation data. In real farm environments, traditional manual annotation is not only time-consuming and labor-intensive, but also prone to missed or incorrect annotations due to subjective fatigue. To alleviate this data bottleneck, using pre-trained models for automatic annotation to accelerate the production of proprietary datasets has become an inevitable trend. This paradigm can achieve high theoretical accuracy on ideal test sets with simple backgrounds, which to some extent speeds up the construction of initial datasets. However, when deployed in the complex and ever-changing captive environment of sika deer farming, existing technologies expose serious pain points in industrial implementation. First, existing models are trapped in the "high evaluation index trap," resulting in a fundamental "index misalignment" in actual automatic annotation projects. Due to the high coupling of background features in sika deer farms, conventional baseline networks are prone to "shortcut learning," fitting false background association features such as background objects of specific shapes in order to improve recall during the training phase. When deployed in real-world applications, this leads to numerous false positives and severe "pseudo-label contamination," necessitating manual review and significant time investment to correct these errors, thus diminishing the efficiency advantages of automatic labeling. Secondly, limited by the computing power of edge computing devices in livestock farms, commonly used lightweight network solutions (such as direct use of depthwise separable convolutions) often abruptly sever the implicit information connections between feature channels. When dealing with complex areas such as overlapping sika deer herds and occlusion in enclosures, this break in channel information severely degrades the model's feature representation capabilities, easily resulting in missed detections or the inability to output high-quality independent bounding boxes. Thirdly, existing object detection networks generally employ deterministic logic at the prediction end, directly regressing the absolute values ​​of bounding boxes. When faced with motion blur and dense overlap caused by the rapid movement of sika deer, this leads to severe localization shifts. This existing structure has poor robustness, generating bounding boxes that do not conform to the target contour, resulting in a large influx of low-quality labels into the dataset. Finally, most current automatic labeling schemes remain at the open-loop stage of single prediction, lacking a "human-in-the-loop" active learning closed-loop mechanism that efficiently feeds back high-quality data after manual fine-tuning and error correction to feed back into the model. This results in the model being unable to adaptively evolve with the accumulation of data, leading to high maintenance costs in the later stages. Summary of the Invention

[0003] The purpose of this invention is to address the problems of high false positives, serious pseudo-label contamination, location drift under dense occlusion, and high cost of manual cleaning in existing technologies. This invention provides an automatic labeling method, system, and device for sika deer images based on an iterative target detection model. Through the deep integration of algorithm architecture and active learning closed loop, false detections are suppressed from the source, thereby improving labeling efficiency.

[0004] To achieve the above objectives, this invention first provides an automatic image annotation method for sika deer based on an iterative target detection model, comprising the following steps:

[0005] S1: Obtain raw image data of sika deer under different weather conditions, lighting patterns and shooting angles to construct an initial dataset containing manually labeled samples;

[0006] S2: Based on the YOLOv9s architecture, a SAL-YOLO model is constructed to extract features of sika deer and output prediction boxes. The SAL-YOLO model includes a backbone feature extraction network, a neck feature fusion network, and a LightBlurDetect detection head.

[0007] S2.1: The SAL-YOLO model uses YOLOv9s as the baseline network architecture. The specific implementation method is as follows: Based on the deep features extracted by the backbone network RepNCSPELAN4, a joint enhancement module composed of SPPELAN and LSKA is embedded at the output end, which aims to purify the background noise by constructing a super large global receptive field.

[0008] The LSKA module employs a four-stage decomposition mechanism, which decomposes a... The standard convolution kernel is decomposed into cascaded horizontal depthwise convolution, vertical depthwise convolution, horizontal depthwise dilation convolution, and vertical depthwise dilation convolution.

[0009] The mathematical formula for generating the attention map is as follows:

[0010]

[0011]

[0012]

[0013]

[0014] in, Represents the original feature map output by SPPELAN. Represents depthwise convolution. This represents a depthwise convolution with a dilation rate. For the generated spatial attention weights, Represents element-wise multiplication;

[0015] S2.2: The neck feature fusion network introduces a GSConv lightweight convolutional module for fusing multi-scale features. The processing procedure of the GSConv lightweight convolutional module includes:

[0016] The input feature map is fed into a standard convolutional layer for channel compression and feature mapping to generate the first intermediate feature.

[0017] The first intermediate feature is input into a depthwise separable convolutional layer to extract local spatial features and generate a second intermediate feature;

[0018] The first intermediate feature and the second intermediate feature are concatenated along the channel dimension;

[0019] S2.3: Replace the native YOLOv9s detection head with the LightBlurDetect lightweight detection head to improve the decoupling accuracy of bounding boxes in motion-blurred scenes. The processing of the LightBlurDetect lightweight detection head includes:

[0020] The input feature map is globally mapped using shared regular convolutional layers, and shared intermediate features are generated through batch normalization and activation function processing.

[0021] Construct parallel classification and regression branches, and input the shared intermediate features into depthwise separable convolutions to extract fine-grained features;

[0022] In the classification branch, the classification confidence score for the target sika deer is output through a 1×1 convolution;

[0023] In the regression branch, the bounding box regression distribution vector of the target is output through 1×1 convolution;

[0024] S3: An iterative active learning closed-loop mechanism is adopted to train the SAL-YOLO model using the initial dataset to obtain weights;

[0025] S4: Use optimal weights to perform automated inference pre-labeling on unlabeled sika deer images and automatically convert them into preliminary label files in standard format;

[0026] S5: Manually verify and modify the preliminary label file, and merge the verified data into the initial dataset to achieve iterative data expansion.

[0027] Furthermore, in step S2.3, the regression branch employs a distributed expectation regression mechanism to handle fuzzy boundaries. The specific processing steps include:

[0028] The edge of the target falls on the first Discrete probabilities of each interval The calculation is performed using the Softmax function, as shown in the following formula:

[0029]

[0030] Predicted values ​​of final bounding box coordinates Calculated using the following formula:

[0031]

[0032] in, This represents the number of equally spaced discrete intervals.

[0033] Furthermore, the iterative active learning closed-loop mechanism described in step S3 specifically includes three expansion stages:

[0034] Cold start phase: Construct an initial dataset using a small number of images with completely manual annotations, and train the model to obtain initial weights;

[0035] First expansion stage: Load the initial weights to perform inference pre-labeling on unlabeled sika deer images. After manual fine-tuning of missed and false detection samples, the corrected data is incorporated into the training set for incremental training to obtain the second generation weights.

[0036] The second expansion stage involves loading the second-generation weights to pre-label a large number of unlabeled sika deer images. After manual fine-tuning of complex samples, these images are incorporated into the training set to complete the final dataset expansion and train the complete model.

[0037] Furthermore, the cost of human-machine collaborative correction is evaluated using an automatic annotation performance score. The specific method is as follows:

[0038] The penalty coefficients for correcting three types of errors—missed detection, false detection, and correct but non-standard prediction boxes—are set to 5, 3, and 1, respectively.

[0039] The final automatic annotation performance score is calculated using the following formula:

[0040]

[0041] in, This represents the number of missed detections. This represents the number of false positives. To determine the number of correct but not standard prediction boxes within the preset range for the intersection-union ratio (IU). The system's theory received full marks.

[0042] In addition, the present invention also provides an automatic image annotation system for sika deer based on an iterative target detection model, comprising:

[0043] The image processing module is used to acquire the original images of sika deer to be labeled, and to construct the dataset through preprocessing and data augmentation;

[0044] The object detection encoder consists of a backbone feature extraction network that incorporates the LSKA attention mechanism and a neck network that incorporates the GSConv lightweight convolution to extract multi-level object fusion features.

[0045] A lightweight detection decoder, including the LightBlurDetect detection head, is used to output the classification confidence and bounding box prediction results of the target;

[0046] The annotation and auditing module is used to automatically convert the prediction results into preliminary label files and provides a human-computer collaborative interface to fine-tune the boundary boxes of missed and false detections.

[0047] The closed-loop iteration module is used to load the previous generation weight file and iteratively fine-tune the object detection encoder and lightweight detection decoder using the expanded high-quality dataset.

[0048] Furthermore, the lightweight detection decoder includes:

[0049] The feature sharing unit is used to perform channel compression and global mapping of input features through regular convolution;

[0050] Fine-grained extraction units are used for parallel feature extraction via depthwise separable convolutions;

[0051] Discrete Expectation Regression (DRR) units are used to treat bounding box coordinates as a probability distribution and output the final bounding box localization result by calculating the mathematical expectation.

[0052] Finally, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described automatic image annotation method for sika deer.

[0053] Beneficial effects of the present invention

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] (1) To reveal the pain point of misalignment of indicators in automatic annotation engineering, a SPPELAN and LSKA joint enhancement module is embedded in the backbone network. Through the large separable convolutional kernel attention mechanism of four-stage decomposition, the model constructs a huge global receptive field, which can accurately extract spatial context, purify the background noise of the enclosure from the spatial logic, block "shortcut learning", reduce the false positive rate, and significantly reduce the number of background false detections in complex scenarios.

[0056] (2) The neck network introduces the GSConv module, which significantly reduces the number of parameters while retaining implicit feature connections through the channel shuffling mechanism, thus achieving efficient and lossless fusion of multi-scale features. Combined with the LightBlurDetect detection head, it transforms hard coordinate regression into flexible distribution expectation regression, solving the problem of localization drift in dense occlusion and motion blur scenarios with low computing power cost, thereby improving the accuracy of the model.

[0057] (3) Constructing a closed-loop and intelligent iterative control for engineering efficiency, a three-stage iterative active learning mechanism of "model pre-labeling - human-in-the-loop fine-tuning - data feedback" is proposed, and an automatic labeling efficiency score (ALES) is introduced. This score objectively quantifies the cost of human-machine collaborative correction and serves as a technical indicator for system state feedback. When ALES is lower than a preset threshold, the system automatically determines that the current model is not adaptable to complex scenarios and triggers the closed-loop iterative module to start the next round of incremental training. At the same time, the system always selects the weight with the highest ALES score as the deployed model, realizing automated closed-loop control from the evaluation system to model evolution, and significantly reducing the cost of manual label cleaning. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the automatic annotation method described in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the overall architecture of the SAL-YOLO model described in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the LSKA module described in an embodiment of the present invention;

[0061] Figure 4 This is a schematic diagram of the GSConv module described in an embodiment of the present invention;

[0062] Figure 5 This is a schematic diagram of the structure of the LightBlurDetect detection head according to an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0064] Example 1

[0065] First, this invention provides an automatic image annotation method for sika deer based on an iterative target detection model, the principle of which is shown in the following block diagram. Figure 1 As shown, it includes the following steps:

[0066] S1: Acquire raw image data of sika deer under different weather conditions (sunny, rainy, snowy), lighting patterns, and shooting angles to construct an initial dataset including manually labeled samples. Specifically, high-definition surveillance cameras were used to conduct all-weather, non-contact image acquisition of the sika deer farm. The cameras were fixedly installed diagonally inside the enclosures and at a high altitude, covering the deer pens, activity fences, and feeding areas. The acquisition process was conducted in an undisturbed, natural environment, covering different weather conditions such as sunny, cloudy, and rainy days, full-color images during the day, infrared images at night, and various perspectives including slight overhead and high-altitude overhead views. Open-source annotation tools were used to accurately annotate the bounding boxes of the sika deer in the images, constructing the initial baseline dataset.

[0067] S2: A SAL-YOLO model based on the YOLOv9s architecture is built to extract features from sika deer and output predicted bounding boxes. For example... Figure 2 As shown, the SAL-YOLO model uses YOLOv9s as the baseline network architecture and mainly consists of three core components: a backbone feature extraction network, a neck feature fusion network, and a lightweight detection head. Addressing challenges such as high false positive rates, severe background feature coupling, and limited computational resources in complex aquaculture environments, this invention deeply reconstructs the baseline architecture, forming a logical chain of "feature purification - lightweight fusion - precise decoupling."

[0068] In the backbone feature extraction network, based on the deep features extracted by RepNCSPELAN4, a joint enhancement module consisting of SPPELAN and LSKA is embedded at the output. The feature stream is first input to the SPPELAN module, which simulates the effect of large-size pooling kernels by cascading multiple small max-pooling layers. Utilizing cross-layer branching and feature concatenation strategies, it captures multi-scale semantic information of different spatial dimensions while reducing computational redundancy, providing a spatial feature base map for identifying individual sika deer in different body positions and depths of field.

[0069] Subsequently, as Figure 3 As shown, features that have undergone preliminary spatial aggregation are entered into the LSKA module. In complex sika deer farming scenarios, background disturbances (such as protruding parts on the top of the shed, non-uniform lighting shadows, and feeding and watering troughs of similar shapes) exhibit strong spatial correlations. Traditional models, due to their limited receptive fields, are prone to "shortcut learning," incorrectly fitting background regularities as the target. The core design of the LSKA module lies in capturing long-range spatial dependencies through ultra-large convolutional kernels, thereby achieving an understanding of the global environmental logic. To avoid the exponential parameter growth and memory overhead caused by directly using two-dimensional large kernel convolutions, LSKA adopts a four-stage decomposition mechanism, decomposing a k×k standard convolutional kernel into cascaded horizontal depthwise convolutions, vertical depthwise convolutions, horizontal depthwise dilation convolutions, and vertical depthwise dilation convolutions. The mathematical formula for generating the attention map is as follows:

[0070]

[0071]

[0072]

[0073]

[0074] in, Represents the original feature map output by SPPELAN. Represents depthwise convolution. This represents a depthwise convolution with a dilation rate. For the generated spatial attention weights, Represents element-wise multiplication;

[0075] Through this decomposition strategy, the model obtains a global receptive field equivalent to a super-large kernel, and further expands the feature response range by using dilated convolution, enabling it to distinguish between fences with straight line features and sika deer individuals with curved outlines from a spatial logic perspective, thus powerfully purifying background noise in the enclosures.

[0076] like Figure 4 As shown, a lightweight convolutional module, GSConv, is introduced into the neck feature fusion network. Efficient fusion of multi-scale features is crucial in automatic annotation tasks for complex aquaculture scenarios. Traditional feature fusion typically relies on a large number of standard convolutions, resulting in significant computational overhead when processing high-resolution feature maps. While conventional lightweight solutions tend to directly employ depthwise separable convolutions, which significantly reduce the number of parameters, they sever implicit information connections between feature channels. When dealing with complex regions such as sika deer partially obscured by fences or in densely overlapping areas, this break in channel information leads to a severe degradation in the model's feature representation capabilities.

[0077] The GSConv module aims to preserve the non-linear dependencies between channels through a sophisticated topological structure. Its processing includes: inputting the input feature map into a standard convolutional layer for channel compression and feature mapping to generate a first intermediate feature; inputting the first intermediate feature into a depthwise separable convolutional layer to extract local spatial features and generate a second intermediate feature; concatenating the first and second intermediate features along the channel dimension; and applying a channel shuffling mechanism to the concatenated features, ensuring that the local spatial information generated by the depthwise separable convolutional branch is evenly and deeply integrated into the global channel information generated by the standard convolutional branch. This compensates for the representational loss in feature extraction caused by lightweight convolutions, outputting a fused multi-scale feature map. By deploying GSConv, the model not only effectively reduces the burden of redundant parameters in the feature propagation process but also ensures the efficient and lossless fusion of multi-scale semantic information of the sika deer.

[0078] like Figure 5 As shown, in the detection head section, the native YOLOv9s detection head is replaced with the LightBlurDetect lightweight detection head. In large-scale farming environments, frequent interactions and gatherings of deer herds lead to severe physical overlap between individuals, and the motion blur caused by the rapid movement of animals makes the boundaries of targets exhibit non-deterministic features. The original prediction structure has poor robustness in handling such blurred edges. Traditional high-computing-power convolutional branches often experience localization shifts when processing motion-blurred edge features due to insufficient feature abstraction granularity. LightBlurDetect abandons parallel independent dimensionality reduction and instead adopts a "shared mapping - lightweight decoupling" design logic. The processing steps include: globally mapping the input feature map using shared regular convolutional layers, and generating shared intermediate features through batch normalization and activation functions; constructing parallel classification and regression branches, and inputting the shared intermediate features into depthwise separable convolutions to extract fine-grained features; in the classification branch, outputting the classification confidence score of the target as a sika deer through a 1×1 convolution; and in the regression branch, outputting the bounding box regression distribution vector of the target through a 1×1 convolution.

[0079] Furthermore, the regression branch employs a distributed expectation regression mechanism to handle fuzzy boundaries. The model no longer directly regresses the absolute values ​​of the bounding box coordinates, but instead treats the coordinates as a probability distribution variable. For each dimension of the bounding box, the vector output by the regression branch is transformed into a discrete probability distribution. Let the prediction range be divided into... If there are n equally spaced discrete intervals, then the target edge falls on the nth interval. Discrete probabilities of each interval The formula for calculation using the Softmax function is as follows:

[0080]

[0081] Based on the above discrete probability distribution, the formula for calculating the predicted values ​​of the final bounding box coordinates is as follows:

[0082]

[0083] Through this distribution expectation mechanism, the model transforms rigid coordinate regression into flexible probability integral, mathematically eliminating the feature mutual exclusion between classification and regression tasks, accurately measuring the uncertainty of fuzzy edges, and significantly enhancing the regression quality of bounding boxes when facing motion fuzziness and extremely overlapping targets.

[0084] S3: An iterative active learning closed-loop mechanism is adopted to train the SAL-YOLO model using the training set to obtain weights. Specifically, this mechanism includes three expansion stages: Cold start stage: an initial dataset is constructed using a small number of images with complete manual annotation, and the model is trained to obtain initial weights; First expansion stage: the initial weights are loaded to perform inference pre-annotation on a certain scale of unlabeled images. After manual fine-tuning of missed and false detection samples, the corrected data is incorporated into the training set for incremental training to obtain the second-generation weights; Second expansion stage: the second-generation weights are loaded to perform pre-annotation on a large scale of unlabeled images. After manual fine-tuning of complex samples, the data is incorporated into the training set to complete the final dataset expansion and train to generate the complete model.

[0085] S4: Use optimal weights to perform automated inference pre-labeling on unlabeled sika deer images and automatically convert them into preliminary label files in standard format.

[0086] S5: Manually verify and modify the initial label files, and merge the verified data into the original dataset to achieve iterative data expansion.

[0087] Furthermore, the automatic annotation method for sika deer images uses the Automatic Annotation Performance Score (ALES) to evaluate the cost of human-machine collaborative correction, and uses it as a control signal for model iteration and deployment. In the automatic annotation process, there is a fundamental "metric misalignment" between the evaluation target and the offline validation set. Correcting missed detections (FN) requires redrawing bounding boxes, which has the highest cost; correcting false detections (FP) only requires deletion, which has a moderate cost; correcting correct but non-standard predicted boxes (CNS) only requires fine-tuning, which has the lowest cost. Therefore, the penalty coefficients for these three types of errors are set to 5, 3, and 1, respectively. The final automatic annotation performance score calculation formula is as follows:

[0088]

[0089] This formula exhibits strict boundary convergence: when the model fails to detect any target, its score is precisely reduced to 0. The closer the score is to 100, the higher the quality of the pseudo-labels. In practical application, the system calculates the ALES score of the current weights on the validation set in real time. When the ALES score falls below a preset threshold, the system determines that the risk of pseudo-label contamination has increased, indicating that the current model is insufficient to adapt to complex scenarios. It then automatically triggers the closed-loop iteration module to initiate the next round of active learning data expansion and incremental training. Simultaneously, the system always retains the weight with the highest ALES score on the validation set as the deployed model, thereby achieving closed-loop control from the evaluation system to model evolution, transforming the labeling work from arduous "labeling from scratch" to efficient "fine-tuning review."

[0090] In addition, this invention provides an automatic image annotation system for sika deer based on an iterative object detection model. This system corresponds to the aforementioned automatic image annotation method for sika deer and includes: an image processing module for acquiring the original images of sika deer to be annotated and constructing a dataset through preprocessing and data augmentation; an object detection encoder comprising a backbone feature extraction network incorporating an LSKA attention mechanism and a neck network incorporating a GSConv lightweight convolution, used to extract multi-level object fusion features; an LSKA module constructing a large global receptive field through a four-stage decomposition mechanism to purify background noise; a lightweight detection decoder comprising a LightBlurDetect detection head, used to output the object classification confidence and bounding box prediction results; an annotation and auditing module for automatically converting the prediction results into preliminary label files and providing a human-computer collaborative interface to fine-tune the bounding boxes of missed and false detections; and a closed-loop iteration module for loading the previous generation weight file and iteratively fine-tuning the object detection encoder and lightweight detection decoder using the expanded high-quality dataset.

[0091] Furthermore, the lightweight detection decoder includes: a feature sharing unit for channel compression and global mapping of input features using regular convolution; a fine-grained extraction unit for parallel feature extraction using depthwise separable convolution; and a discrete expectation regression unit for treating bounding box coordinates as a probability distribution and outputting the final bounding box localization result through a calculation formula.

[0092] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0093] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0094] 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.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic image annotation method for sika deer based on an iterative target detection model, characterized in that, Includes the following steps, S1: Obtain raw image data of sika deer under different weather conditions, lighting patterns and shooting angles to construct an initial dataset containing manually labeled samples; S2: Based on the YOLOv9s architecture, a SAL-YOLO model is constructed to extract features of sika deer and output prediction boxes. The SAL-YOLO model includes a backbone feature extraction network, a neck feature fusion network, and a LightBlurDetect detection head. S2.1: The SAL-YOLO model uses YOLOv9s as the baseline network architecture. The specific implementation method is as follows: Based on the deep features extracted by the backbone network RepNCSPELAN4, a joint enhancement module composed of SPPELAN and LSKA is embedded at the output end, which aims to purify the background noise by constructing a super large global receptive field. The LSKA module employs a four-stage decomposition mechanism, which decomposes a... The standard convolution kernel is decomposed into cascaded horizontal depthwise convolution, vertical depthwise convolution, horizontal depthwise dilation convolution, and vertical depthwise dilation convolution. The mathematical formula for generating the attention map is as follows: ； ； ； ； in, Represents the original feature map output by SPPELAN. Represents depthwise convolution. This represents a depthwise convolution with a dilation rate. For the generated spatial attention weights, Represents element-wise multiplication; S2.2: The neck feature fusion network introduces a GSConv lightweight convolutional module for fusing multi-scale features. The processing procedure of the GSConv lightweight convolutional module includes: The input feature map is fed into a standard convolutional layer for channel compression and feature mapping to generate the first intermediate feature. The first intermediate feature is input into a depthwise separable convolutional layer to extract local spatial features and generate a second intermediate feature; The first intermediate feature and the second intermediate feature are concatenated along the channel dimension; S2.3: Replace the native YOLOv9s detection head with the LightBlurDetect lightweight detection head to improve the decoupling accuracy of bounding boxes in motion-blurred scenes. The processing of the LightBlurDetect lightweight detection head includes: The input feature map is globally mapped using shared regular convolutional layers, and shared intermediate features are generated through batch normalization and activation function processing. Parallel classification and regression branches are constructed, and the shared intermediate features are respectively input into depthwise separable convolutions to extract fine-grained features; In the classification branch, the classification confidence score for the target sika deer is output through a 1×1 convolution; In the regression branch, the bounding box regression distribution vector of the target is output through 1×1 convolution; S3: An iterative active learning closed-loop mechanism is adopted to train the SAL-YOLO model using the initial dataset to obtain weights; S4: Use optimal weights to perform automated inference pre-labeling on unlabeled sika deer images and automatically convert them into preliminary label files in standard format; S5: Manually verify and modify the preliminary label file, and merge the verified data into the initial dataset to achieve iterative data expansion.

2. The automatic image annotation method for sika deer based on an iterative target detection model according to claim 1, characterized in that, In step S2.3, the regression branch uses a distributed expectation regression mechanism to handle fuzzy boundaries. The specific processing steps include: The edge of the target falls on the first Discrete probabilities of each interval The calculation is performed using the Softmax function, as shown in the following formula: ； Predicted values ​​of final bounding box coordinates Calculated using the following formula: ； in, This represents the number of equally spaced discrete intervals.

3. The automatic image annotation method for sika deer based on an iterative target detection model according to claim 1, characterized in that, The iterative active learning closed-loop mechanism described in step S3 specifically includes three expansion stages: Cold start phase: Construct an initial dataset using a small number of images with completely manual annotations, and train the model to obtain initial weights; First expansion stage: Load the initial weights to perform inference pre-labeling on unlabeled sika deer images. After manual fine-tuning of missed and false detection samples, the corrected data is incorporated into the training set for incremental training to obtain the second generation weights. The second expansion stage involves loading the second-generation weights to pre-label a large number of unlabeled sika deer images. After manual fine-tuning of complex samples, these images are incorporated into the training set to complete the final dataset expansion and train the complete model.

4. The automatic image annotation method for sika deer based on an iterative target detection model according to claim 3, characterized in that, The cost of human-machine collaborative correction is evaluated using an automatic annotation performance score. The specific method is as follows: The penalty coefficients for correcting three types of errors—missed detection, false detection, and correct but non-standard prediction boxes—are set to 5, 3, and 1, respectively. The final automatic annotation performance score is calculated using the following formula: ； in, This represents the number of missed detections. This represents the number of false positives. To determine the number of correct but not standard prediction frames within a preset range for the intersection-union ratio (IU). The system's theory received full marks.

5. An automatic image annotation system for sika deer based on an iterative target detection model, characterized in that, include: The image processing module is used to acquire the original images of sika deer to be labeled, and to construct the dataset through preprocessing and data augmentation; The object detection encoder consists of a backbone feature extraction network that incorporates the LSKA attention mechanism and a neck network that incorporates the GSConv lightweight convolution, used to extract multi-level object fusion features. A lightweight detection decoder, including the LightBlurDetect detection head, is used to output the classification confidence and bounding box prediction results of the target; The annotation and auditing module is used to automatically convert the prediction results into preliminary label files and provides a human-computer collaborative interface to fine-tune the boundary boxes of missed and false detections. The closed-loop iteration module is used to load the previous generation weight file and iteratively fine-tune the object detection encoder and lightweight detection decoder using the expanded high-quality dataset.

6. The automatic image annotation system for sika deer based on an iterative target detection model according to claim 5, characterized in that, The lightweight detection decoder includes: The feature sharing unit is used to perform channel compression and global mapping of input features through regular convolution; Fine-grained extraction units are used for parallel feature extraction via depthwise separable convolutions; Discrete Expectation Regression (DRR) units are used to treat bounding box coordinates as a probability distribution and output the final bounding box localization result by calculating the mathematical expectation.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

Images