Small target detection method based on association rules

Abstract

The invention discloses a small target detection method based on association rules, and relates to the technical field of target detection. The system is realized based on a small target detection system, and comprises a data preprocessing module which is responsible for the preprocessing of an input image, including image denoising, scale normalization and enhancement processing, so as to improve the identifiability of small target features; the feature extraction module is used for automatically learning and extracting image features by adopting PP-YOLOv8; the association rule generation module is used for generating an association rule by analyzing a target co-occurrence mode in the training data set; and the target detection module combines the extracted features with association rules, and performs target identification and positioning by using a machine learning algorithm. According to the method, the association rule and the deep learning network are integrated, so that the detection capability of a small target is effectively improved, and particularly, the application performance in large-scale scenes such as remote sensing images and unmanned aerial vehicle aerial photography is outstanding; in addition, the method has wide application potential in the fields of border monitoring, anti-terrorism reconnaissance, judicial assistance and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a small target detection method based on association rules. Background Art

[0002] Object detection has made significant progress in the past few years. It is the research foundation of computer vision research fields such as pose estimation, pedestrian detection, face detection, object tracking, and autonomous driving. Small object detection has long been a research pain point in object detection, and there are countless scenes in our lives that require small object detection. However, due to the poor visual features and high noise of small objects, in addition, large-scale benchmark datasets for small-size object detection are still not comprehensive enough, small object detection is still one of the most challenging tasks in computer vision.

[0003] After searching, the Chinese patent application number CN202111214050.6 discloses a target detection method, system and electronic device based on digital images, which relates to the field of pattern recognition technology. The method first obtains the image to be detected; then the image to be detected is input into the trained target detection model to determine the probability value of each pixel in the image to be detected containing the target; wherein the target detection model is used to downsample the image to be detected to feature maps of 3 different scales, and obtain the probability value of the target corresponding to each pixel in the image to be detected according to the output tensor of the feature map; finally, the target area of ​​the image to be detected is determined according to the probability value of each pixel in the image to be detected containing the target, and the detection result of the target is determined according to the target area. The target detection method in the above patent has the following shortcomings: it is mainly used for text detection, and it is still insufficient for other small target detection functions. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a small target detection method based on association rules.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A small target detection method based on association rules, comprising:

[0007] Data preprocessing module: responsible for the preprocessing of input images, including image denoising, scale normalization and enhancement processing to improve the recognizability of small target features;

[0008] Feature extraction module: PP-YOLOv8 is used to automatically learn and extract image features;

[0009] Association rule generation module: Generates association rules by analyzing the target co-occurrence patterns in the training data set;

[0010] Target detection module: combines the extracted features with association rules and uses machine learning algorithms to identify and locate targets; distinguishes targets of different sizes and provides specific detection strategies;

[0011] Post-processing module: optimizes the detection results, including non-maximum suppression and bounding box correction, to obtain more accurate target boundaries and confidence assessment;

[0012] Result output module: presents the final detected small targets in a visual way and provides an interface for further analysis and processing;

[0013] Iterative optimization module: collects feedback information based on actual detection results, continuously adjusts association rules and detection model parameters, and realizes continuous iteration and optimization of the system.

[0014] Preferably: the data preprocessing module: performs image denoising and normalization processing, and adopts advanced image enhancement technology, including adaptive histogram equalization, local contrast enhancement and enhancement strategy for specific small target features, specifically:

[0015] Retinex theory is used to process color constancy to enhance the features of small targets with low contrast.

[0016] Use Gaussian-Laplacian pyramid for edge detection to highlight the edge information of small objects.

[0017] Preferably: the feature extraction module: based on the PP-YOLOv8 model, adopts the attention mechanism and feature fusion technology, including SENet and FastDCN, to focus on the features of small targets and suppress background noise, specifically:

[0018] Introducing a deformable convolutional network to adapt to small objects with irregular shapes by learning the offset of the convolution kernel;

[0019] Integrate a bidirectional feature pyramid network structure to achieve more effective fusion of features at different levels;

[0020] The feature extraction module improves the adaptability of the model to small targets with irregular shapes by adding a learnable offset based on the standard convolution. The specific formula is:

[0021]

[0022] Among them, x(i) represents the input feature map, w k is the convolution kernel weight, Δp k is the offset obtained from the learnable offset field and y(i) is the output feature map.

[0023] Preferably, the association rule generation module is not only based on the target co-occurrence pattern, but also considers the time sequence information and contextual environmental factors, and adopts an association analysis model, specifically:

[0024] Use time series analysis methods to mine dynamic association rules of target movement;

[0025] Use graph convolutional networks to analyze scene graph structures and capture spatial correlations;

[0026] The association rule generation module predicts future values ​​in time series data based on time series analysis, as follows:

[0027] Y t =c+φ1Y t-1 +…+φ p Y t-p +θ1∈ t-1 +…+θ q ∈ t-q +∈ t

[0028] Among them, Y t is the current value, φ and θ are model parameters, and ∈ is the error term, which helps capture the dynamic pattern of target motion in association rule generation.

[0029] Preferably, the target detection module adopts multi-scale detection and anchor point optimization strategies to improve the detection accuracy of small targets, specifically:

[0030] Apply multi-scale training and testing strategies to make the network better adapt to input images of different sizes;

[0031] The K-means++ algorithm is used to optimize the selection of anchor points to make it more consistent with the actual data distribution.

[0032] Preferably: the post-processing module: adopts a post-processing algorithm to improve the detection accuracy, specifically:

[0033] Use Bayesian optimization algorithm to adjust the threshold of NMS to adapt to different detection scenarios;

[0034] The conditional random field model is introduced to refine the detection results.

[0035] Preferably, the iterative optimization module: adopts online learning mechanism and model distillation technology to continuously improve system performance, specifically:

[0036] Use online learning to gradually update association rules and model parameters to ensure that the system adapts to the changing environment;

[0037] Transfer the knowledge of complex models to lightweight models through knowledge distillation to maintain the real-time performance and deployment convenience of the system;

[0038] The iterative optimization module transfers the knowledge of the complex model to the lightweight model to improve the operation efficiency and deployment convenience of the model, as follows:

[0039] L distill (W T ,S)=aL ce (W T ; x,y)+(1-a)L kd (S,W T ; x)

[0040] Among them, W T is the teacher model, S is the student model, L ce is the cross entropy loss, L kd is the knowledge distillation loss.

[0041] Preferably, the method comprises the following steps:

[0042] S1: Data preprocessing, inputting raw image data; performing image denoising; implementing scale normalization to match model input requirements; applying color constancy and local contrast enhancement strategies to improve the recognizability of small target features;

[0043] S2: Feature extraction, using PP-YOLOv8 to automatically learn image features; introducing attention mechanism and feature fusion technology, and enhancing the feature representation of small targets through SENet, FastDCN and other methods; applying deformable convolutional networks to adapt to the shape of small targets;

[0044] S3: Application of association rules: Based on the mined co-occurrence patterns and temporal information, association rules are applied to assist in identifying potential small target areas; graph convolutional networks are used to analyze the scene graph structure and obtain spatial correlation;

[0045] S4: Object detection and recognition, under the multi-scale object detection strategy, adjust the network parameters to adapt to input images of different sizes; use the optimized anchor points for object positioning; combine the detection results of multiple models and use ensemble learning to improve the final detection accuracy;

[0046] S5: Post-processing optimization, performing non-maximum suppression to remove redundant candidate boxes; using the bounding box correction algorithm to fine-tune the detection results; applying the conditional random field model to fine-tune the detection results;

[0047] S6: Result output.

[0048] Preferably, the result output method of S6 is:

[0049] The detection results are superimposed on the original image for visual presentation; time series analysis results are displayed to track target motion trends; user feedback is received through an interactive user interface for subsequent model optimization.

[0050] Preferably, the feature extraction module is based on a deformable convolutional network, which allows the convolution kernel to deform freely to adapt to targets of irregular shapes, which is particularly effective for capturing the fine structure of small targets. The specific formula is:

[0051]

[0052] Among them, y(m,n) is the output feature map, x is the input feature map, k is the deformable convolution kernel, and b is the bias term.

[0053] Preferably: the detection method also includes: feedback and model iteration, collecting user feedback and new data samples for online learning, updating association rules and model parameters; implementing knowledge distillation to transfer complex model knowledge to a lightweight model to ensure the real-time performance of the system; continuously evaluating the efficiency and accuracy of the system through a performance monitor, and making necessary adjustments.

[0054] The beneficial effects of the present invention are:

[0055] 1. The present invention effectively improves the detection capability of small targets by integrating association rules and deep learning networks, especially in large-scale scenarios such as remote sensing images and drone aerial photography. In addition, it has broad application potential in border monitoring, anti-terrorism reconnaissance, judicial assistance and other fields.

[0056] 2. The present invention continuously collects feedback and optimizes the model, so that the system can adapt to various complex environments and improve the accuracy and robustness of small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of a small target detection method based on association rules proposed by the present invention. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is further described in detail below in conjunction with specific implementation methods.

[0059] Embodiment 1:

[0060] A small target detection method based on association rules is implemented based on a small target detection system, including:

[0061] Data preprocessing module: responsible for the preprocessing of input images, including image denoising, scale normalization and enhancement processing to improve the recognizability of small target features;

[0062] Feature extraction module: PP-YOLOv8 is used to automatically learn and extract image features;

[0063] Association rule generation module: Generates association rules by analyzing the target co-occurrence patterns in the training data set;

[0064] Target detection module: combines the extracted features with association rules and uses machine learning algorithms to identify and locate targets; distinguishes targets of different sizes and provides specific detection strategies;

[0065] Post-processing module: optimizes the detection results, including non-maximum suppression and bounding box correction, to obtain more accurate target boundaries and confidence assessment;

[0066] Result output module: presents the final detected small targets in a visual way and provides an interface for further analysis and processing;

[0067] Iterative optimization module: collects feedback information based on actual detection results, continuously adjusts association rules and detection model parameters, and realizes continuous iteration and optimization of the system.

[0068] The data preprocessing module performs image denoising and normalization, and uses advanced image enhancement techniques, including adaptive histogram equalization, local contrast enhancement, and enhancement strategies for specific small target features, specifically:

[0069] Retinex theory is used to process color constancy to enhance the features of small targets with low contrast.

[0070] Use Gaussian-Laplacian pyramid for edge detection to highlight the edge information of small objects.

[0071] Among them, the feature extraction module: based on the PP-YOLOv8 model, adopts the attention mechanism and feature fusion technology, including SENet and FastDCN, to focus on the features of small targets and suppress background noise, specifically:

[0072] Introducing a deformable convolutional network to adapt to small objects with irregular shapes by learning the offset of the convolution kernel;

[0073] The Bidirectional Feature Pyramid Network (BiFPN) structure is integrated to achieve more effective fusion of features at different levels.

[0074] The feature extraction module improves the adaptability of the model to small targets with irregular shapes by adding a learnable offset based on the standard convolution. The specific formula is:

[0075]

[0076] Among them, x(i) represents the input feature map, w k is the convolution kernel weight, Δp k is the offset obtained from the learnable offset field and y(i) is the output feature map.

[0077] The association rule generation module is not only based on the target co-occurrence pattern, but also considers the time sequence information and contextual factors, and adopts the association analysis model, specifically:

[0078] Use time series analysis methods to mine dynamic association rules of target movement;

[0079] Graph convolutional networks are used to analyze the scene graph structure and capture spatial correlations.

[0080] The association rule generation module predicts future values ​​in time series data based on time series analysis, as follows:

[0081] Y t =c+φ1Y t-1 +…+φ p Y t-p +θ1∈ t-1 +…+θ q ∈ t-q +∈ t

[0082] Among them, Y t is the current value, φ and θ are model parameters, and ∈ is the error term, which helps capture the dynamic pattern of target motion in association rule generation.

[0083] The target detection module adopts multi-scale detection and anchor point optimization strategies to improve the detection accuracy of small targets, specifically:

[0084] Apply multi-scale training and testing strategies to make the network better adapt to input images of different sizes;

[0085] The K-means++ algorithm is used to optimize the selection of anchor points to make it more consistent with the actual data distribution.

[0086] Among them, the post-processing module: adopts a post-processing algorithm to improve the detection accuracy, specifically:

[0087] Use Bayesian optimization algorithm to adjust the threshold of NMS to adapt to different detection scenarios;

[0088] The conditional random field model is introduced to refine the detection results.

[0089] The iterative optimization module uses online learning mechanism and model distillation technology to continuously improve system performance, specifically:

[0090] Use online learning to gradually update association rules and model parameters to ensure that the system adapts to the changing environment;

[0091] Through knowledge distillation, the knowledge of complex models is transferred to lightweight models to maintain the real-time and deployment convenience of the system.

[0092] The iterative optimization module transfers the knowledge of the complex model to the lightweight model to improve the operation efficiency and deployment convenience of the model, as follows:

[0093] L distill (W T ,S)=aL ce (W T ; x,y)+(1-a)L kd (S,W T ; x)

[0094] Among them, W T is the teacher model, S is the student model, L ce is the cross entropy loss, L kd is the knowledge distillation loss.

[0095] Embodiment 2:

[0096] A small target detection method based on association rules includes the following steps:

[0097] S1: Data preprocessing, inputting raw image data; performing image denoising; implementing scale normalization to match model input requirements; applying color constancy and local contrast enhancement strategies to improve the recognizability of small target features;

[0098] S2: Feature extraction, using PP-YOLOv8 to automatically learn image features; introducing attention mechanism and feature fusion technology, and enhancing the feature representation of small targets through SENet, FastDCN and other methods; applying deformable convolutional networks to adapt to the shape of small targets;

[0099] S3: Application of association rules: Based on the mined co-occurrence patterns and temporal information, association rules are applied to assist in identifying potential small target areas; graph convolutional networks are used to analyze the scene graph structure and obtain spatial correlation;

[0100] S4: Object detection and recognition, under the multi-scale object detection strategy, adjust the network parameters to adapt to input images of different sizes; use the optimized anchor points for object positioning; combine the detection results of multiple models and use ensemble learning to improve the final detection accuracy;

[0101] S5: Post-processing optimization, performing non-maximum suppression to remove redundant candidate boxes; using the bounding box correction algorithm to fine-tune the detection results; applying the conditional random field model to fine-tune the detection results;

[0102] S6: Result output.

[0103] The feature extraction module is based on a deformable convolutional network, which allows the convolution kernel to deform freely to adapt to irregularly shaped targets. This is particularly effective for capturing the fine structure of small targets. The specific formula is:

[0104]

[0105] Among them, y(m,n) is the output feature map, x is the input feature map, k is the deformable convolution kernel, and b is the bias term.

[0106] The result output mode of S6 is as follows:

[0107] The detection results are superimposed on the original image for visual presentation; time series analysis results are displayed to track target motion trends; user feedback is received through an interactive user interface for subsequent model optimization.

[0108] The method also includes: feedback and model iteration, collecting user feedback and new data samples for online learning, updating association rules and model parameters; implementing knowledge distillation to transfer complex model knowledge to lightweight models to ensure the real-time performance of the system; continuously evaluating the efficiency and accuracy of the system through performance monitors, and making necessary adjustments.

[0109] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A small target detection method based on association rules, It is characterized in that Based on the small target detection system, it includes: Data preprocessing module: responsible for the preprocessing of input images, including image denoising, scale normalization and enhancement processing to improve the recognizability of small target features; Feature extraction module: PP-YOLOv8 is used to automatically learn and extract image features; Association rule generation module: Generates association rules by analyzing the target co-occurrence patterns in the training data set; Target detection module: combines the extracted features with association rules and uses machine learning algorithms to identify and locate targets; distinguishes targets of different sizes and provides specific detection strategies; Post-processing module: optimizes the detection results, including non-maximum suppression and bounding box correction, to obtain more accurate target boundaries and confidence assessment; Result output module: presents the final detected small targets in a visual way and provides an interface for further analysis and processing; Iterative optimization module: collects feedback information based on actual detection results, continuously adjusts association rules and detection model parameters, and realizes continuous iteration and optimization of the system.

2. According to the small target detection method based on association rules in claim 1, It is characterized in that The data preprocessing module performs image denoising and normalization processing, and adopts advanced image enhancement technology, including adaptive histogram equalization, local contrast enhancement, and enhancement strategies for specific small target features, specifically: Retinex theory is used to process color constancy to enhance the features of small targets with low contrast. Use Gaussian-Laplacian pyramid for edge detection to highlight the edge information of small objects.

3. According to the small target detection method based on association rules in claim 1, It is characterized in that The feature extraction module: Based on the PP-YOLOv8 model, the attention mechanism and feature fusion technology, including SENet and FastDCN, are used to focus on the features of small targets and suppress background noise. Specifically: Introducing a deformable convolutional network to adapt to small objects with irregular shapes by learning the offset of the convolution kernel; Integrate a bidirectional feature pyramid network structure to achieve more effective fusion of features at different levels; The feature extraction module improves the adaptability of the model to small targets with irregular shapes by adding a learnable offset based on the standard convolution. The specific formula is: Among them, x(i) represents the input feature map, w k is the convolution kernel weight, Δp k is the offset obtained from the learnable offset field and y(i) is the output feature map.

4. According to the small target detection method based on association rules in claim 1, It is characterized in that The association rule generation module is not only based on the target co-occurrence pattern, but also considers the temporal information and contextual factors, and adopts an association analysis model, specifically: Use time series analysis methods to mine dynamic association rules of target movement; Use graph convolutional networks to analyze scene graph structures and capture spatial correlations; The association rule generation module predicts future values ​​in time series data based on time series analysis, as follows: Y t =c+φ 1 Y t-1 +…+φ p Y t-p +θ 1 ∈ t-1 +…+θ q ∈ t-q +∈ t Among them, Y t is the current value, φ and θ are model parameters, and ∈ is the error term, which helps capture the dynamic pattern of target motion in association rule generation.

5. According to the small target detection method based on association rules in claim 1, It is characterized in that The target detection module: adopts multi-scale detection and anchor point optimization strategy to improve the detection accuracy of small targets, specifically: Apply multi-scale training and testing strategies to make the network better adapt to input images of different sizes; The K-means++ algorithm is used to optimize the selection of anchor points to make it more consistent with the actual data distribution.

6. The small target detection method based on association rules according to claim 1, It is characterized in that The post-processing module: uses a post-processing algorithm to improve the detection accuracy, specifically: Use Bayesian optimization algorithm to adjust the threshold of NMS to adapt to different detection scenarios; The conditional random field model is introduced to refine the detection results.

7. The small target detection method based on association rules according to claim 1, It is characterized in that The iterative optimization module: uses online learning mechanism and model distillation technology to continuously improve system performance, specifically: Use online learning to gradually update association rules and model parameters to ensure that the system adapts to the changing environment; Transfer the knowledge of complex models to lightweight models through knowledge distillation to maintain the real-time performance and deployment convenience of the system; The iterative optimization module transfers the knowledge of the complex model to the lightweight model to improve the operation efficiency and deployment convenience of the model, as follows: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> distill <h2 style=";text-align:left;direction:ltr"> (W<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> ,S)=aL<h2 style=";text-align:left;direction:ltr"> ce <h2 style=";text-align:left;direction:ltr"> (W<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (x,y)+(1-a)L<h2 style=";text-align:left;direction:ltr"> kd <h2 style=";text-align:left;direction:ltr"> (S,W<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (x) Among them, W T is the teacher model, S is the student model, L ce is the cross entropy loss, L kd is the knowledge distillation loss.

8. A small target detection method based on association rules according to any one of claims 1 to 7, It is characterized in that The steps include: S1: Data preprocessing, inputting raw image data; performing image denoising; implementing scale normalization to match model input requirements; applying color constancy and local contrast enhancement strategies to improve the recognizability of small target features; S2: Feature extraction, using PP-YOLOv8 to automatically learn image features; Introducing attention mechanism and feature fusion technology, and enhancing the feature representation of small targets through SENet, FastDCN and other methods; Apply deformable convolutional networks to adapt to small object shapes; S3: Application of association rules: Based on the mined co-occurrence patterns and temporal information, association rules are applied to assist in identifying potential small target areas; graph convolutional networks are used to analyze the scene graph structure and obtain spatial correlation; S4: Object detection and recognition, under the multi-scale object detection strategy, adjust the network parameters to adapt to input images of different sizes; use the optimized anchor points for object positioning; combine the detection results of multiple models and use ensemble learning to improve the final detection accuracy; S5: Post-processing optimization, performing non-maximum suppression to remove redundant candidate boxes; using the bounding box correction algorithm to fine-tune the detection results; applying the conditional random field model to fine-tune the detection results; S6: Result output; The result output mode of S6 is: Overlay the detection results on the original image for visualization; provide time series analysis results display and track the target movement trend; Receive user feedback through an interactive user interface for subsequent model optimization.

9. The small target detection method based on association rules according to claim 1, It is characterized in that The feature extraction module is based on a deformable convolutional network, which allows the convolution kernel to deform freely to adapt to targets of irregular shapes. This is particularly effective for capturing the fine structure of small targets. The specific formula is: Among them, y(m,n) is the output feature map, x is the input feature map, k is the deformable convolution kernel, and b is the bias term.

10. The small target detection method based on association rules according to claim 8, It is characterized in that The detection method also includes: feedback and model iteration, collecting user feedback and new data samples for online learning, updating association rules and model parameters; implementing knowledge distillation to transfer complex model knowledge to a lightweight model to ensure the real-time performance of the system; continuously evaluating the efficiency and accuracy of the system through a performance monitor, and making necessary adjustments.

Images