Small target feature processing method and system based on multi-scale convolution

Abstract

The present application belongs to the technical field of image analysis, and particularly relates to a small target feature processing method and system based on multi-scale convolution, which utilizes multi-scale shared convolution processing, spatial-to-depth convolution processing and cross-stage multi-shape convolution processing, and combines multi-scale feature fusion processing and a double attention level detection head to realize accurate perception of tiny and dense defect targets in a complex scene, and has comprehensive advantages in terms of detection accuracy and model compactness.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image analysis, and particularly relates to a small target feature processing method and system based on multi-scale convolution. BACKGROUND

[0002] With the continuous expansion of the scale of high-voltage power transmission networks, the safety inspection of transmission lines is crucial to ensure the stable operation of the power grid. A large number of lines pass through complex terrains such as mountains, plateaus and deserts, and are exposed to harsh environments such as strong winds, lightning and snow for a long time, resulting in structural damage such as cracks, fractures and shedding of key components such as insulators, anti-vibration hammers and spacer bars. At the same time, there may be bird nests, plastic films and other foreign objects attached to the lines, which may cause flashover, short circuit and other serious faults if not discovered in time, threatening the safety of the power grid.

[0003] In recent years, target detection technology based on unmanned aerial vehicles and deep learning has been widely applied to power inspection, significantly improving the efficiency and coverage of inspection. However, in actual aerial images, most defects and foreign objects are small in size, low in contrast and fuzzy in texture, and are often affected by changes in lighting, background interference and occlusion, belonging to typical weakly visible targets. Existing detection models have deficiencies in feature extraction and multi-scale fusion, and have weak response to small-scale and low signal-to-noise ratio targets, resulting in high miss rate and inaccurate positioning, making it difficult to meet the high reliability detection requirements.

[0004] To further improve detection performance, some methods introduce deep network structures or complex attention mechanisms to enhance feature expression capability, but often result in increased model parameters and redundant structures, which is not conducive to deployment and application in resource-constrained scenarios. Therefore, how to effectively enhance the feature perception capability of small and low-contrast targets while controlling the complexity of the model, and improve the accuracy and robustness of detection, is still a technical problem to be solved. SUMMARY

[0005] To solve the problems in the background art, the application provides a small target feature processing method and system based on multi-scale convolution.

[0006] The technical scheme of the application is as follows:

[0007] The application provides a small target feature processing method based on multi-scale convolution, comprising:

[0008] S1: Obtain a transmission line image, and after multiple downsampling, obtain corresponding first, second and third extracted feature maps;

[0009] The third extracted feature map is processed by multi-scale shared convolution to obtain a first feature map;

[0010] S2: The first feature map is spliced with the third extracted feature map, and after multi-scale feature fusion processing, a first fused feature map is obtained;

[0011] The first extracted feature map is subjected to spatial-to-depth convolution processing to obtain a second fused feature map, specifically:

[0012] The first extracted feature map is divided into four non-overlapping sub-blocks in the spatial dimension according to different row and column indexes, and the four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map. After that, local context features are extracted through convolution with grouping, and weighted fusion is performed combined with a lightweight spatial attention mechanism to obtain a second fused feature map.

[0013] The second fused feature map, the first fused feature map, and the second extracted feature map are spliced, and after cross-stage multi-convolution processing, multi-scale feature fusion processing is performed to obtain a third fused feature map. After upsampling, it is spliced with the first extracted feature map to obtain a second feature map.

[0014] S3: The second feature map is subjected to multi-scale feature fusion cascade processing, specifically:

[0015] After the second feature map is subjected to multi-scale feature fusion processing and SimAM attention processing in turn, a first scale feature map is obtained.

[0016] After the first scale feature map is spliced with the third fused feature map, multi-scale feature fusion processing and SimAM attention processing are performed in turn to obtain a second scale feature map.

[0017] After the second scale feature map is spliced with the first fused feature map, multi-scale feature fusion processing and SimAM attention processing are performed in turn to obtain a third scale feature map.

[0018] The first scale feature map, the second scale feature map, and the third scale feature map constitute a third feature map group.

[0019] S4: Each feature map in the third feature map group is processed by a dual attention layer detection head to obtain a fourth feature map group, which is processed by a regression head and a classification head for detection.

[0020] The multi-scale shared convolution processing of S1 is specifically:

[0021] After the third extracted feature map is subjected to convolution for channel compression to obtain an intermediate feature map, multi-scale context features are extracted through three dilated convolution branches with different dilation rates and one ghost convolution branch, while the third extracted feature map is reserved as a residual branch. After splicing the dilated convolution branch, the ghost convolution branch, and the residual branch features in the channel dimension, feature fusion and compression are performed through convolution to obtain the first feature map.

[0022] The multi-scale feature fusion processing described in S2 is as follows:

[0023] After the features to be processed are extracted by reparameterized convolution, they are sequentially processed by spatial attention and depthwise separable convolution to obtain the main branch enhanced features.

[0024] The main branch enhancement features and residual branch features are concatenated through channels and then fused using convolution.

[0025] The cross-stage polymorphic convolution processing described in S2 is specifically as follows:

[0026] The concatenated features are subjected to four different shapes of depth convolution operations in parallel to obtain multi-shape convolution features;

[0027] After the spliced ​​features are processed by frequency domain channel attention, spatial attention processing is performed to obtain spatially weighted frequency domain features.

[0028] Spatially weighted frequency domain features are fused with multi-shape convolutional features. The fused features are then concatenated with the residual branch features of the spliced ​​features, and channel attention processing is performed after the number of channels is compressed by convolution.

[0029] Each feature map in the third feature map group described in S4 is processed by a dual-attention layer detection head, specifically as follows:

[0030] Each scale feature in the third feature map group is first compressed through convolution, then scale attention and spatial attention are applied sequentially, and feature extraction is completed in a shared convolution.

[0031] Furthermore, it also includes: calculating the loss, which consists of classification loss and regression loss.

[0032] It also includes: employing a double distillation mechanism, including a block-related knowledge distillation strategy loss that calculates Kullback–Leibler divergence for each category, and a feature alignment loss that uses channel weight distance.

[0033] It also includes: building a graphical user interface that supports image input, model loading, visualization of detection results, defect statistical analysis, and result export.

[0034] This invention also provides a small target feature processing system based on multi-scale convolution, comprising:

[0035] Feature extraction module: used to acquire images of transmission lines, and after multiple downsampling, obtain the corresponding first, second and third extracted feature maps;

[0036] The third extracted feature map is processed by multi-scale shared convolution to obtain the first feature map;

[0037] The feature fusion module: after the first feature map and the third extracted feature map are spliced, multi-scale feature fusion processing is performed to obtain a first fused feature map;

[0038] The first extracted feature map is subjected to spatial-to-depth convolution processing to obtain a second fused feature map, specifically:

[0039] The first extracted feature map is divided into four non-overlapping sub-blocks in the spatial dimension according to different row and column indexes, the four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map, local context features are extracted through convolution with grouping, and weighted fusion is performed in combination with a lightweight spatial attention mechanism to obtain a second fused feature map;

[0040] The second fused feature map, the first fused feature map and the second extracted feature map are spliced, subjected to cross-stage multi-shape convolution processing, and then subjected to multi-scale feature fusion processing to obtain a third fused feature map, which is up-sampled and spliced with the first extracted feature map to obtain a second feature map;

[0041] The cascade processing module: the second feature map is subjected to multi-scale feature fusion cascade processing, specifically:

[0042] The second feature map is sequentially subjected to multi-scale feature fusion processing and SimAM attention processing to obtain a first scale feature map;

[0043] The first scale feature map and the third fused feature map are spliced and sequentially subjected to multi-scale feature fusion processing and SimAM attention processing to obtain a second scale feature map;

[0044] The second scale feature map and the first fused feature map are spliced and sequentially subjected to multi-scale feature fusion processing and SimAM attention processing to obtain a third scale feature map;

[0045] The first scale feature map, the second scale feature map and the third scale feature map constitute a third feature map group;

[0046] The detection module: each feature map in the third feature map group is processed by a dual attention level detection head to obtain a fourth feature map group, which is processed by a regression head and a classification head for detection.

[0047] Advantages

[0048] The present application utilizes multi-scale shared convolution processing, spatial-to-depth convolution processing and cross-stage multi-shape convolution processing, in combination with multi-scale feature fusion processing and a dual attention level detection head, to realize accurate perception of small and dense defect targets in complex scenes, and has comprehensive advantages in detection accuracy and model compactness.

[0049] The application embeds SimAM attention processing when fusing features, and designs a composite loss function that fuses SlideLoss, IoU and normalized Gaussian Wasserstein distance, to optimize model performance from two dimensions of 'feature focusing' and 'loss modeling'. On the one hand, SimAM is a lightweight attention module that can significantly improve saliency expression without introducing trainable parameters; on the other hand, the loss function comprehensively considers the smoothness of the classification boundary, the accuracy of target positioning and the shape matching ability, effectively accelerating model convergence, improving precision stability, and suppressing overfitting in multi-scale scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The detection results of the model using the method of the application and the model not using the method of the application are compared. DETAILED DESCRIPTION

[0051] The following examples are intended to illustrate the application, but not to further limit the application.

[0052] The application provides a small target feature processing method based on multi-scale convolution, comprising:

[0053] S1: Obtain a power transmission line image, and obtain corresponding first, second and third extraction feature maps after multiple downsampling (the number of downsampling is not less than 3);

[0054] The third extraction feature map is processed by multi-scale shared convolution to obtain a first feature map.

[0055] First, the obtained power transmission line image is preprocessed, including image enhancement, the preprocessed image is labeled, and a training set, a validation set and a test set are divided. For example, the preprocessed image is labeled by Roboflow for target detection, and finally the labeled image is divided into a training set (4141 images), a validation set (591 images) and a test set (1184 images) according to a ratio of 7:1:2.

[0056] Among them, the labeled data analysis shows that the number of small targets in the image is significant, and the class distribution is uneven. Most of the targets have a normalized size less than 0.2, and the small targets are mainly distributed in the upper region of the image, showing complex multi-scale spatial distribution characteristics.

[0057] In the downsampling stage, the first, second and third extraction feature maps can be 4, 8 and 16 times feature maps, respectively, and are denoted as P2 feature map, P3 feature map and P4 feature map.

[0058] Preferably, the multi-scale shared convolution processing (denoted as MDSC) is specifically:

[0059] The third extracted feature map is subjected to channel compression through convolution (such as 1x1) to obtain an intermediate feature map ; Multi-scale context features are extracted through three hollow convolution branches with different expansion rates and one Ghost Convolution (Ghost-Conv) branch respectively, while the third extracted feature map is reserved as a residual branch; the features of the hollow convolution branch, the Ghost Convolution branch and the residual branch are spliced in the channel dimension, and then subjected to feature fusion and compression through convolution (such as 1x1) to obtain the first feature map . The following expression is satisfied: ;

[0060] In the formula, 、 、 , and respectively represent the features of the hollow convolution branches with expansion rates of 1, 3 and 5; is the feature of the 5x5 Ghost Convolution branch; is the third extracted feature map connected by residual connection.

[0061] S2: After the first feature map and the third extracted feature map are spliced, multi-scale feature fusion processing is performed to obtain a first fused feature map;

[0062] The first extracted feature map is subjected to spatial-to-depth convolution processing to obtain a second fused feature map, specifically:

[0063] The first extracted feature map is divided into four non-overlapping sub-blocks according to different row and column indexes in the spatial dimension, and the four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map. Then, local context features are extracted through convolution with grouping, and weighted fusion is performed in combination with a lightweight spatial attention mechanism to obtain the second fused feature map;

[0064] The second fused feature map, the first fused feature map and the second extracted feature map are spliced, subjected to cross-stage multi-convolution processing, and then subjected to multi-scale feature fusion processing to obtain a third fused feature map. After upsampling, the third fused feature map is spliced with the first extracted feature map to obtain a second feature map.

[0065] First, the first feature map is subjected to nearest neighbor upsampling and spliced with the downsampled 16 times feature map (P4 feature map, i.e. the third extracted feature map) in the feature extraction stage, and subjected to multi-scale feature fusion processing (denoted as MSFA-Block) for preliminary semantic integration.

[0066] The multi-scale feature fusion processing adopts a lightweight multi-path convolution structure and an attention fusion mechanism, including a main branch and a residual branch. Specifically:

[0067] The feature to be processed The features are extracted by a reparameterized convolution (denoted as RepConv), then a lightweight spatial attention is applied to enhance the response of salient regions, and then a depth separable convolution (DSConv) is used for spatial feature extraction to obtain the enhanced feature of the main branch ; which can be expressed as:

[0068] ;

[0069] wherein, represents a spatial attention processing, is a depth separable convolution operation, is an activation function.

[0070] The enhanced feature of the main branch is concatenated with the residual branch feature , and fused by a 1x1 convolution.

[0071] Next, regarding the spatial-to-depth convolution processing (denoted as SPDConv), the down-sampled 4 times feature map (P2 feature map, i.e. the first extracted feature map) in the feature extraction stage is divided into four non-overlapping sub-blocks according to different row and column indexes in the spatial dimension by a 2x2 grid, specifically: the first sub-block is composed of all pixels with even row and column indexes, the second sub-block is composed of all pixels with odd row index and even column index, the third sub-block is composed of all pixels with even row index and odd column index, and the fourth sub-block is composed of all pixels with odd row and column indexes, then the four sub-blocks are concatenated along the channel dimension to generate the enhanced feature map , the expression is as follows:

[0072] ;

[0073] wherein, represents a concatenation operation along the channel dimension, represents four spatial sub-blocks obtained by dividing according to different row and column indexes.

[0074] The enhanced feature map is extracted by a 3x3 convolution with grouping to obtain local context features, and further combined with a lightweight spatial attention mechanism for weighted fusion, and the obtained second fusion feature map is denoted as . The expression is as follows:

[0075] ;

[0076] wherein, represents a convolution operation with grouping, is a global average pooling, is a spatial attention weight kernel, is an activation function, Represents element-wise multiplication. This structure improves spatial sensitivity while enhancing local receptive fields, making it suitable for small target feature modeling.

[0077] Then, the second fused feature map is spliced with the first fused feature map after nearest neighbor upsampling and the shallow layer feature down-sampled 8 times feature map (P3 feature map, i.e. the second extracted feature map), and cross-stage multi-shape convolution processing (denoted as CSPOKM) is performed. The CSPOKM adopts a typical CSP (Cross Stage Partial) structure to divide the features into a main branch and a residual branch. Specifically:

[0078] The main branch first performs four different shape depth convolution operations on the spliced features in parallel, with convolution kernel sizes of 1x31, 31x1, 31x31, and 1x1, to extract multi-shape direction perception features and enhance spatial receptive fields and direction discrimination ability. To further enhance frequency domain feature expression, a frequency domain channel attention mechanism (FCA) is introduced. Specifically, the spliced features are subjected to Fast Fourier Transform (FFT), and a 1x1 convolution and a Sigmoid activation function are used to generate channel attention weights, achieving adaptive enhancement of frequency domain amplitude information. Subsequently, spatial attention processing (SCA) is used to enhance the saliency of spatial positions.

[0079] The spatially weighted frequency domain features are fused with the multi-shape convolution features. The fused features are concatenated with the residual branch of the spliced features, and the channel number is compressed through convolution (such as 1x1). Finally, a lightweight channel attention (CA) is applied to further enhance important features while maintaining the integrity of semantic information. Thus, the CSPOKM processed features are obtained.

[0080] Finally, the CSPOKM processed features are subjected to multi-scale feature fusion processing for multi-path convolution and residual fusion, and are up-sampled and spliced with the down-sampled 4 times feature map (P2 feature map, i.e. the first extracted feature map) from the feature extraction stage to obtain the second feature map .

[0081] S3: The second feature map is subjected to multi-scale feature fusion cascade processing, specifically:

[0082] After the second feature map is subjected to multi-scale feature fusion processing and SimAM attention processing in sequence, the first scale feature map is obtained.

[0083] After the first scale feature map is spliced with the third fused feature map and subjected to multi-scale feature fusion processing and SimAM attention processing in sequence, the second scale feature map is obtained.

[0084] The second scale feature map is spliced with the first fusion feature map, and then sequentially subjected to multi-scale feature fusion processing and SimAM attention processing to obtain a third scale feature map.

[0085] The first scale feature map, the second scale feature map and the third scale feature map constitute a third feature map group.

[0086] The SimAM module is a parameter-free attention mechanism, and the core idea is to weight each pixel point based on its similarity with the whole image to enhance the response of the salient region. The weighted attention feature map is the third feature map group .

[0087] S4: Each feature map in the third feature map group is processed by a double attention hierarchical detection head to obtain a fourth feature map group, which is processed by a regression head and a classification head for detection.

[0088] Preferably, the third feature map group Each feature map is processed by a double attention hierarchical detection head (DAHD), specifically:

[0089] Each scale feature of the third feature map group is first subjected to channel compression by convolution (such as 1x1), and then scale attention and spatial attention are sequentially applied, and feature extraction is completed in shared convolution.

[0090] ;

[0091] In the formula, represents global average pooling; represents spatial attention processing; represents element-wise multiplication; represents convolution in shared convolution extraction; represents different scale feature maps of the third feature map group; each scale fusion feature map processed by DAHD is denoted as , and the fourth feature map group obtained finally is used for final detection prediction.

[0092] The DAHD branch generates frame offset and class confidence through the regression head and the classification head respectively, and is spliced in the channel dimension to form scale prediction . In the inference stage, the three scale results are combined to obtain the final detection , and the specific operation is as follows:

[0093] ;

[0094] ;

[0095] In the formula, denotes channel dimension concatenation; denotes a scaling factor; for the bounding box regression branch, for the classification branch; denotes a function that converts the predicted distribution into actual bounding box coordinates.

[0096] The application utilizes multi-scale shared convolution processing, spatial-to-depth convolution processing, cross-stage multi-convolution processing, and combines multi-scale feature fusion processing and dual attention hierarchical detection heads to realize accurate perception of small and dense defect targets in complex scenes, and has comprehensive advantages in detection accuracy and model compactness.

[0097] The application also includes using a combined loss function in the training stage to improve target detection accuracy and robustness, which is composed of classification loss and regression loss.

[0098] Specifically, the classification loss uses SlideLoss, which adjusts the adaptive modulation weight by predicting the confidence of the model to realize smooth processing of difficult samples, and the calculation method is:

[0099] ;

[0100] In the formula, is the predicted confidence of the model; is the real label; denotes the standard cross-entropy loss function; is the IoU adaptive threshold (generally 0.5); is an indicator function, which takes the value 1 when the condition is met, otherwise 0. This mechanism improves the sensitivity of the model to difficult classification samples by dynamically adjusting the loss weight.

[0101] The regression loss uses a weighted combination of the IoU loss and the normalized Gaussian Wasserstein distance (NWD) loss to improve the accuracy and stability of the bounding box prediction. This combination enhances the model's perception of scale changes and fuzzy boundaries while maintaining positioning accuracy.

[0102] The application embeds the SimAM parameter-free attention mechanism in the feature fusion stage and designs a composite loss function that combines SlideLoss, IoU and NWD to optimize model performance from two dimensions of "feature focusing" and "loss modeling". On the one hand, SimAM as a lightweight attention module can significantly improve the saliency expression without introducing trainable parameters; on the other hand, the loss function considers the classification boundary smoothness, target positioning accuracy and shape matching ability, effectively accelerating model convergence, improving precision stability, and suppressing overfitting in multi-scale scenarios.

[0103] Further, the application also introduces a knowledge distillation strategy to alleviate the performance degradation problem of small models in the compression process. Specifically, a double distillation mechanism is adopted, including a block correlation knowledge distillation (BCKD) loss based on the calculation of Kullback-Leibler divergence for each category, and a feature alignment loss using channel weight distance (CWD).

[0104] The application introduces the BCKD distillation strategy in the training process, so that the student model maintains the reasoning ability while obtaining detection performance close to the teacher model. The detection performance is close to the teacher model, which significantly alleviates the problem of insufficient precision of small models in edge deployment scenarios.

[0105] The application also includes: constructing a graphical user interface to support image input, model loading, detection result visualization, defect statistical analysis and result export. For example, a graphical visualization interface is constructed based on PyQt5, which integrates image input, model loading, detection result display, defect category and quantity statistics, result saving and other functions. The operation process is simple and the interaction is friendly, which provides a good foundation for subsequent embedding into existing inspection business systems or accessing edge terminal platforms, and significantly improves the detection efficiency and on-site decision support capability of power operation personnel.

[0106] The application also provides a small target feature processing system based on multi-scale convolution, comprising:

[0107] The feature extraction module is used to obtain the power transmission line image, and after multiple downsampling, the corresponding first, second and third extraction feature maps are obtained;

[0108] The third extraction feature map is processed by multi-scale shared convolution to obtain the first feature map;

[0109] The feature fusion module: after the first feature map and the third extraction feature map are spliced, they are processed by multi-scale feature fusion to obtain the first fusion feature map;

[0110] The first extraction feature map is processed by spatial-to-depth convolution to obtain the second fusion feature map, specifically:

[0111] The first extraction feature map is divided into four non-overlapping sub-blocks in the spatial dimension according to different row and column indexes, and the four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map. Then, the local context features are extracted by convolution with grouping, and the weighted fusion is performed by combining the lightweight spatial attention mechanism to obtain the second fusion feature map;

[0112] The second fusion feature map, the first fusion feature map and the second extraction feature map are spliced, and after cross-stage multi-shape convolution processing, multi-scale feature fusion processing is performed to obtain a third fusion feature map, which is up-sampled and spliced with the first extraction feature map to obtain a second feature map;

[0113] The cascade processing module: the second feature map is subjected to multi-scale feature fusion cascade processing, specifically:

[0114] The second feature map is subjected to multi-scale feature fusion processing and SimAM attention processing in sequence to obtain a first scale feature map;

[0115] After the first scale feature map is spliced with the third fusion feature map, multi-scale feature fusion processing and SimAM attention processing are performed in sequence to obtain a second scale feature map;

[0116] After the second scale feature map is spliced with the first fusion feature map, multi-scale feature fusion processing and SimAM attention processing are performed in sequence to obtain a third scale feature map;

[0117] The first scale feature map, the second scale feature map and the third scale feature map constitute a third feature map group;

[0118] The detection module: each feature map in the third feature map group is processed by a double attention level detection head to obtain a fourth feature map group, which is processed by a regression head and a classification head for detection.

[0119] Experimental results

[0120] In the experimental stage, the teacher model adopts a YOLOv8s backbone network, and the weights are fixed in the training stage; the student model is an improved version of the network structure (the method of the application), which has a smaller parameter size and stronger small target feature processing capability. In order to strengthen semantic alignment, multi-layer supervision mechanism is adopted for feature distillation, three groups of middle layers (such as the 18th, 21st and 24th layers corresponding to the 21st, 25th and 29th layers) of the teacher and student networks are selected for channel dimension alignment, the channel weight distance (CWD) loss is used to measure the channel distribution difference, and the student network is guided to learn multi-level semantic features more effectively.

[0121] Table 1. Comparison of processing results of the application and the original model on the self-built data set

[0122]

[0123] Table 1 shows the performance comparison of the method of the application and YOLOv8n on the self-built power line defect data set. The evaluation indexes include small target average precision (AP S), mAP@0.5, mAP@0.5:0.95 and model parameter amount (Params). Among them, mAP is the arithmetic mean of the average precision (AP) of each category, reflecting the overall detection performance of the model; mAP@0.5 represents the mAP calculated when the Intersection over Union (IoU) threshold is not less than 0.5, which is used to evaluate the detection ability of the model under the normal precision; mAP@0.5:0.95 represents the average value of mAP calculated at 10 levels with an interval of 0.05 from 0.5 to 0.95, which comprehensively reflects the positioning robustness of the model; AP S For measuring the detection performance of small targets with an area less than 32x32 pixels. IoU is the ratio of the overlapping area of the predicted box and the real box to the union area, which is used to judge whether the detection is correct.

[0124] From Table 1, it can be seen that the method of the application has AP S improved by 15.4 percentage points, mAP@0.5 reached 81.9%, improved by 9.9 percentage points compared with YOLOv8n, mAP@0.5:0.95 improved by 5.4 percentage points, and the model parameter amount reduced by 33.3%. The experimental results show that the application significantly improves the detection ability of small defects and low visibility targets while maintaining a small model size.

[0125] Figure 1 The detection results of the method of the application and the original YOLOv8 model (without using the method of the application) on the self-built power transmission line defect detection dataset are compared. The figure contains multiple typical scenes, and focuses on the improvement effect of the method of the application in identifying insulators and anti-vibration hammers. The false detection problem is improved: in Figure 1 , the original YOLOv8 model incorrectly identifies the background area as an insulator, resulting in false detection. In contrast, the method of the application can effectively distinguish between targets and backgrounds, successfully avoiding this false detection situation and improving the reliability of the detection results.

[0126] These results show that the method of the application has stronger robustness in handling small targets and complex backgrounds, and can significantly improve the detection accuracy and reliability.

[0127] Table 2. Comparison of processing results of the application and the original model on VisDrone2019

[0128]

[0129] Table 2 shows the performance comparison of the method of the application and the YOLOv8 model on the VisDrone2019 dataset. The dataset is built by the Machine Learning and Data Mining Lab of A University in cooperation with multiple institutions, and is a benchmark dataset for target detection in unmanned aerial vehicle aerial scene, covering complex backgrounds such as urban roads, traffic hubs, and dense crowds, with small target size, severe occlusion, and variable view angles, which can effectively evaluate the detection ability of the model in complex environments.

[0130] The experimental results show that the method of the application improves the small target detection precision (AP S ) by 1.7 percentage points, mAP@0.5 by 3.0 percentage points, and mAP@0.5:0.95 by 1.8 percentage points, while the model parameter amount is reduced from 3.0M to 2.0M, a reduction of 33.3%. The above results verify the effectiveness of the application in improving detection precision and enhancing small target recognition ability.

Claims

1. A method for processing small target features based on multi-scale convolution, characterized in that, include: S1: Acquire images of transmission lines, perform multiple downsampling operations, and obtain the corresponding first, second, and third extracted feature maps; The third extracted feature map is processed by multi-scale shared convolution to obtain the first feature map; S2: After the first feature map and the third extracted feature map are concatenated, the first fused feature map is obtained through multi-scale feature fusion processing; The first extracted feature map is subjected to spatial-to-depth convolution to obtain the second fused feature map, specifically: The first extracted feature map is divided into four non-overlapping sub-blocks in the spatial dimension according to different row and column indices. The four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map. After that, local context features are extracted by grouped convolution and then weighted and fused in combination with a lightweight spatial attention mechanism to obtain the second fused feature map. The second fused feature map, the first fused feature map, and the second extracted feature map are concatenated, and after cross-stage polymorphic convolution processing, multi-scale feature fusion processing is performed to obtain the third fused feature map. After upsampling, it is concatenated with the first extracted feature map to obtain the second feature map. S3: The second feature map undergoes multi-scale feature fusion cascade processing, specifically as follows: After the second feature map is processed by multi-scale feature fusion and SimAM attention, the first-scale feature map is obtained. After the first-scale feature map and the third-scale fused feature map are concatenated, multi-scale feature fusion processing and SimAM attention processing are performed sequentially to obtain the second-scale feature map. After the second-scale feature map is concatenated with the first fused feature map, multi-scale feature fusion processing and SimAM attention processing are performed sequentially to obtain the third-scale feature map. The first-scale feature map, the second-scale feature map, and the third-scale feature map constitute the third feature map group; S4: Each feature map in the third feature map group is processed by a dual attention layer detection head to obtain the fourth feature map group, which is then processed by a regression head and a classification head for detection.

2. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, The multi-scale shared convolution processing described in S1 is specifically as follows: The third extracted feature map is compressed through convolution to obtain an intermediate feature map. Then, multi-scale contextual features are extracted through three dilated convolution branches with different dilation rates and one ghost convolution branch. At the same time, the third extracted feature map is retained as a residual branch. The features of the dilated convolution branch, the ghost convolution branch and the residual branch are concatenated in the channel dimension and then fused and compressed through convolution to obtain the first feature map.

3. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, The multi-scale feature fusion processing described in S2 is as follows: After the features to be processed are extracted by reparameterized convolution, they are sequentially processed by spatial attention and depthwise separable convolution to obtain the main branch enhanced features. The main branch enhancement features and residual branch features are concatenated through channels and then fused using convolution.

4. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, The cross-stage polymorphic convolution processing described in S2 is specifically as follows: The concatenated features are subjected to four different shapes of depth convolution operations in parallel to obtain multi-shape convolution features; After the spliced ​​features are processed by frequency domain channel attention, spatial attention processing is performed to obtain spatially weighted frequency domain features. Spatially weighted frequency domain features are fused with multi-shape convolutional features. The fused features are then concatenated with the residual branch features of the spliced ​​features, and channel attention processing is performed after the number of channels is compressed by convolution.

5. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, Each feature map in the third feature map group described in S4 is processed by a dual-attention layer detection head, specifically as follows: Each scale feature in the third feature map group is first compressed through convolution, then scale attention and spatial attention are applied sequentially, and feature extraction is completed in a shared convolution.

6. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, Also includes: Calculate the loss, which consists of classification loss and regression loss.

7. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, Also includes: A double distillation mechanism is employed, including a block-related knowledge distillation strategy loss that calculates the Kullback–Leibler divergence for each category, and a feature alignment loss that uses channel weight distance.

8. The small target feature processing method based on multi-scale convolution according to claim 1, characterized in that, Also includes: Build a graphical user interface that supports image input, model loading, visualization of detection results, defect statistical analysis, and result export.

9. A small target feature processing system based on multi-scale convolution, characterized in that, include: Feature extraction module: used to acquire images of transmission lines, and after multiple downsampling, obtain the corresponding first, second and third extracted feature maps; The third extracted feature map is processed by multi-scale shared convolution to obtain the first feature map; Feature fusion module: After the first feature map and the third extracted feature map are concatenated, the first fused feature map is obtained through multi-scale feature fusion processing; The first extracted feature map is subjected to spatial-to-depth convolution to obtain the second fused feature map, specifically: The first extracted feature map is divided into four non-overlapping sub-blocks in the spatial dimension according to different row and column indices. The four non-overlapping sub-blocks are concatenated along the channel dimension to generate an enhanced feature map. After that, local context features are extracted by grouped convolution and then weighted and fused in combination with a lightweight spatial attention mechanism to obtain the second fused feature map. The second fused feature map, the first fused feature map, and the second extracted feature map are concatenated, and after cross-stage polymorphic convolution processing, multi-scale feature fusion processing is performed to obtain the third fused feature map. After upsampling, it is concatenated with the first extracted feature map to obtain the second feature map. Cascaded processing module: The second feature map undergoes multi-scale feature fusion cascaded processing, specifically as follows: After the second feature map is processed by multi-scale feature fusion and SimAM attention, the first-scale feature map is obtained. After the first-scale feature map and the third-scale fused feature map are concatenated, multi-scale feature fusion processing and SimAM attention processing are performed sequentially to obtain the second-scale feature map. After the second-scale feature map is concatenated with the first fused feature map, multi-scale feature fusion processing and SimAM attention processing are performed sequentially to obtain the third-scale feature map. The first-scale feature map, the second-scale feature map, and the third-scale feature map constitute the third feature map group; Detection module: Each feature map in the third feature map group is processed by a dual attention-level detection head to obtain the fourth feature map group, which is then processed by a regression head and a classification head for detection.

Images