Target identification method based on adaptive dynamic convolution feature extraction and detail enhancement

Abstract

The invention discloses a target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, relates to the technical field of image target recognition, and designs an adaptive dynamic convolution feature extraction structure (C3DC) in order to solve feature learning difficulty caused by small target scale in an image. The adaptability of the model to a small-scale target is enhanced by dynamically adjusting a convolution kernel; in order to solve the problem of small target feature loss caused by a complex background, the model adopts a detail enhancement shared convolution detection head (GNDE), group normalization and detail enhancement convolution technologies are combined, and features of small targets are enhanced, so that the small targets are separated from the complex background; in order to solve the problem that loss calculation is difficult due to uneven sample quality, an HWIoU loss function is designed, interference of geometric measurement on the training process is reduced through a dynamic adjustment mechanism, and the precision and generalization ability of the method are improved.

Description

Technical Field

[0001] This invention relates to the field of image target recognition technology, and in particular to a target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement. Background Technology

[0002] Drone aerial imagery is playing an increasingly crucial role in the booming low-altitude economy due to its wide coverage, multispectral information acquisition capabilities, and unique high-angle overhead views. Drone aerial imagery is not only relatively inexpensive and flexible in deployment, but it also provides observational dimensions that are difficult to achieve with traditional remote sensing or ground-based photography. This provides unprecedented high spatiotemporal resolution perspectives for numerous fields, including spatial data updates in Geographic Information Systems (GIS), current status surveys and dynamic monitoring in urban planning, rapid response and assessment after natural disasters (such as floods and earthquakes), crop growth analysis and precision fertilization in agriculture, and geomorphological identification and mineral exploration in geological exploration.

[0003] With the continuous advancement of image processing and artificial intelligence technologies, drone aerial images still face a series of severe and unique technical challenges when applied to core target recognition tasks. These challenges mainly stem from the inherent conditions of their imaging: First, due to the typically high flight altitude, targets of interest (such as vehicles, pedestrians, and small facilities) occupy a very small proportion of the entire image, often consisting of only tens or even a few pixels, resulting in blurred target features and a lack of information; second, despite advancements in sensor technology, the actual usable resolution of some images remains unsatisfactory due to limitations in platform payload and data transmission, further exacerbating the difficulty of recognizing small targets; third, the background of aerial photography scenes is usually extremely complex, encompassing dense building clusters, irregular farmland, dense forests, or complex road networks, resulting in low distinction between targets and background, easily leading to missed or false detections; in addition, the variable lighting conditions in outdoor environments, such as high-contrast shadows under strong sunlight, low contrast on rainy days, and uneven lighting during dawn and dusk, can all cause drastic changes in the appearance of targets, affecting the stability of recognition.

[0004] Faced with these challenges, traditional target recognition methods (such as Viola-Jones and HOG+SVM based on handcrafted features) exhibit significantly limited performance when processing drone aerial images due to their general workflow (image preprocessing, manually designed feature extraction, classifier design and training). The fundamental bottleneck lies in the feature extraction stage: these methods heavily rely on specific feature descriptors designed by researchers (such as SIFT and HOG), which are often designed for a particular scene or target, resulting in limited representational and discriminative power. When confronted with complex challenges such as small targets, multi-scale features, complex backgrounds, and varying lighting conditions in drone images, handcrafted features struggle to capture robust and sufficiently discriminative information, exhibiting severely insufficient generalization ability and failing to adapt to diverse new scenarios and more complex practical application needs. Summary of the Invention

[0005] To address the above technical problems, this invention provides a target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, comprising the following steps: Step S1: Acquire drone aerial images and corresponding annotation information; Step S2: Construct a target recognition network. In the backbone network of the target recognition network, introduce a C3DC module for automatically generating weights based on the current image. The C3DC module is constructed by replacing the ordinary convolutions of the Bottleneck module in the C3k2 module with DyConv convolutions. The DyConv convolutions are used to combine multiple sets of convolutional kernels according to weights. The backbone network outputs multi-scale features, namely small target scale feature P3, medium target scale feature P4, and large target scale feature P5. When the target area is smaller than... When the target area is within a certain range (in pixels), it is defined as a small target; when the target area is within a certain range (in pixels), it is defined as a small target. and When the target area is between pixels, it is defined as a medium target; when the target area is larger than a certain value... When the target size is in pixels, it is defined as a large target. Step S3: By fusing semantic and detail information across scales through the neck network of the target recognition network, the three features P3, P4 and P5 are fused to output the fused features; Step S4: Introduce the GNDE module into the detection head of the target recognition network. The GNDE module includes the Conv_GN module and the DEConv module. The Conv_GN module includes convolution and GN normalization. The three features P3, P4 and P5 are first aligned and stably normalized by the Conv_GN module, and then the edge and orientation details are enhanced by several layers of DEConv module. Then, the class probability and bounding box parameters are output by the classification branch and the regression branch, respectively. The regression branch is followed by Scale to scale and calibrate the regression amplitude of features at different scales. Step S5: Combine classification and confidence loss and optimize the regression branch using HWIoU; Step S6: Filter candidate boxes based on confidence scores and remove duplicate boxes using NMS to obtain the final detection results, including the final target category, bounding box position, and corresponding confidence score.

[0006] The technical solution further defined in this invention is: Furthermore, in step S2, a DyConv convolution is designed using MoE-style convolution combined with a parameter expansion function, as shown in the following formula: ; ; Where X represents the input feature map, Y represents the convolution output feature map, and * represents the convolution operation. This represents the adaptive convolution kernel generated by DyConv convolution, where M represents the number of expert convolution kernels. This represents the weight parameters of the i-th expert convolution kernel; This represents the i-th expert weight coefficient adaptively generated from the input, satisfying... and .

[0007] As described above, in the target recognition method based on adaptive dynamic convolutional feature extraction and detail enhancement, in step S2, the UAV aerial image enters the backbone network for feature extraction and downsampling in sequence. The image passes through the CBS module twice to complete basic convolutional feature extraction and resolution compression; then, it passes through the C3DC module for adaptive dynamic convolutional feature extraction; after downsampling by the CBS module, it enters the second C3DC module to obtain the first lateral output feature of the backbone network; then, it continues to be downsampled by the CBS module and enters the third C3DC module to obtain the second lateral output feature; after further downsampling by the CBS module, it enters the fourth C3DC module, then performs multi-scale pooling through SPPF spatial pyramid pooling, and finally passes through the C2PSA attention module to form the third output feature of the backbone network.

[0008] As described above, in the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, in step S3, the output of the backbone network enters the neck network for cross-scale feature fusion. First, starting from the third output feature of the backbone network, a top-down FPN path is executed, which upsamples the feature and concatenates it with the output of the third C3DC module in the backbone network. Then, it is fused and compressed by the fifth C3DC module. Next, it is upsampled and concatenated with the output of the second C3DC module in the backbone network. After fusion by the sixth C3DC module, the small target detection scale feature P3 is output. Then, a bottom-up PAN path is executed, which downsamples feature P3 by the CBS module and concatenates it with the output of the fifth C3DC module. Then, it is fused by the seventh C3DC module to output the medium target scale feature P4. Finally, feature P4 is downsampled by the CBS module and concatenated with the third output feature of the backbone network. After fusion by the eighth C3DC module, the large target scale feature P5 is output.

[0009] As described above, the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement includes a CBS module comprising a 3×3 convolution, a normalization layer, and an activation function. In the C3DC module, the input features are first processed through a Conv layer for channel adjustment and feature extraction. Then, the features are split into two branches according to the channels. One branch serves as a shortcut branch to directly retain detailed information, while the other branch enters the main branch for modeling. The main branch sequentially passes through several C3k2 modules to enhance the nonlinear expression. The output of the main branch and the output of the shortcut branch are concat-fused along the channel dimension. Finally, another Conv layer is used to complete feature integration and output.

[0010] As described above, in the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, the input in the C3k2 module is fused by Concat after passing through two parallel branches. In one branch, the input passes through only one layer of Conv, and in the other branch, the input passes through one layer of Conv and then several Bottleneck modules. In the Bottleneck module, the input passes through two DyConv convolutions in sequence and is then fused with the input itself by add.

[0011] As described above, in the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, the Conv_GN module in step S4 uses GN as regularization. GN divides the channel dimension into G groups, and independently calculates the mean and variance in each group. For an input feature map of size N×C×H×W, where N represents the batch size, H represents the height, and W represents the width, GN divides channel C into G groups, each containing... There are 4 channels; for the g-th group, the formulas for calculating its mean and variance are as follows: ; ; Where g represents the channel group index. This represents the set of all activation values ​​in group g that are included in the statistics. Represents a set The i-th activation value in the set, where m represents the set. The total number of active values; This represents the mean of the g-th group; GN calculates statistics only within the channel grouping of each sample. GN standardizes the activation values ​​of each group as follows: ; Finally, GN applies learnable scaling and offset operations, as shown in the following formula: ; in, Indicates to The standardized results after group normalization This represents a very small constant used to avoid a denominator of zero. This represents the final output of GN. and These represent the learnable scaling factor and offset, respectively.

[0012] As described above, in the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, in the DEConv module of step S4, the input features are processed by standard convolution, and simultaneously by center difference, angle difference, vertical difference and horizontal difference convolution in parallel. The outputs of each convolution are fused at the end in the channel dimension to obtain the final features.

[0013] As described above, the target recognition method based on adaptive dynamic convolutional feature extraction and detail enhancement utilizes reparameterization technology to merge the outputs of multiple parallel convolutional layers into the output of a single convolutional layer; assuming a convolutional layer has a corresponding weight kernel. , where i represents the sequence number of the convolutional layer; during the training phase, the weight kernel... They are learned and updated independently; during forward propagation, the DEConv module adds these weight kernels at the corresponding positions to form an equivalent weight kernel. As shown in the following formula: ; in, This represents the input feature map of the DEConv module. This represents the output feature map of the DEConv module; This represents the convolution kernel of the i-th parallel branch, corresponding to standard convolution, central difference, angular difference, horizontal difference, and vertical difference convolution; 5 indicates that the number of parallel branches is 5; This represents the equivalent convolution kernel obtained by adding the convolution kernels of each branch at their corresponding positions during the inference phase.

[0014] As described above, in the target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, step S5 constructs HWIoU based on the dynamic adjustment of WloU, as shown in the following formula: ; in, This represents the regression loss of HWIoU; This represents a dynamic weighting term, used to adjust the weights of samples of different quality and at different stages. This represents the basic IoU regression loss term; x and y represent the coordinates of the center point of the prediction box; Represents the coordinates of the center point of the true bounding box; Represents the width and height of the actual bounding box; v and v represent the adaptive weighting function and the metric that describes the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box, respectively.

[0015] The beneficial effects of this invention are: (1) Traditional static convolution faces the contradiction of "simultaneous processing of multi-scale targets by the same weight kernel" in UAV aerial photography. That is, small targets occupy only a few pixels on the feature map, have weak response and are easily submerged by background texture. This invention introduces DyConv convolution based on the MoE idea in the backbone and neck key modules. Multiple sets of convolution kernels are weighted and combined through attention weights, so that the convolution kernels adapt to the input content. Essentially, it is equivalent to dynamically selecting more suitable effective receptive fields and filtering modes for different image regions. In this way, when the target is very small or the edge texture is extremely weak, the network tends to activate expert kernels that are beneficial to "high-frequency details / edge structures". In large targets or background-dominated areas, it can favor more robust smoothing and context modeling kernels. This mechanism enhances the "visibility" of learnable features of small targets without significantly increasing inference overhead, reduces the probability of missed detection and improves scale adaptation. (2) Drone aerial photography data often has high resolution and high memory pressure. During training, the batch size is relatively small, and the unstable statistics of BN will cause the output of the detector head to jitter and the details to be "averaged", which is especially unfavorable for weak signal categories such as small targets; This invention will use Group Normalization is introduced into the detection head, enabling normalization statistics to be calculated "by channel within the sample," independent of batch size. This allows for stable feature distribution and gradient propagation even with small batches or even batch=1. Simultaneously, the DEConv module integrates standard convolutions with various difference branches (central difference, directional difference, horizontal / vertical difference, etc.), explicitly injecting gradient priors into convolution calculations, making the network more sensitive to edges, corners, and directional textures. Combined with shared convolutions and reparameterization, stronger representations are obtained during training, while equivalent convolutions can be merged during inference to control computational load. Overall, this combination of "steady-state normalization + detailed gradient enhancement" significantly improves the separability of small targets in complex backgrounds (roads, building textures, shadows, night scene noise), reduces false detections of fragmented backgrounds as targets, and improves the quality of small target boundary localization. (3) The overlap between the predicted bounding box and the ground truth box of small targets is usually very low in the early stage. If the loss function penalizes the geometric terms too strongly or the weight allocation is unreasonable, the training will be pulled by a large number of low-quality matching samples, resulting in problems such as "large gradient noise, slow convergence, and difficulty in fine-tuning localization in the later stage". This invention proposes HWIoU to introduce a dynamic adjustment mechanism in the regression loss: in the early stage of training, the effective gradient of low IoU samples is appropriately amplified to prompt the network to quickly improve the overlap and establish basic alignment. When the IoU gradually converges, the weight of this part is reduced, and the optimization focus is shifted to the center offset and aspect ratio and other fine-tuning terms to achieve more refined localization convergence. At the same time, a more "mild" penalty strategy is adopted for geometric measures such as distance, scale, and shape (avoiding excessive dominance of geometric terms) to reduce their interference on the training process and make the model more robust to "difficult samples" such as occlusion, blur, and low contrast. For small targets of UAVs, this is equivalent to letting the model first "pull the box to the vicinity of the target" and then gradually "fit the edge and corner" of the box, thereby improving the final mAP and cross-scene generalization ability. (4) In this invention, the C3DC module solves the problem of "not being able to see clearly" (weak feature enhancement and scale adaptation), the GNDE module solves the problem of "not being able to distinguish" (detail separation and stable training in complex backgrounds), and the HWIoU solves the problem of "not being able to learn stably / not being able to learn accurately" (dynamic steady-state optimization of the regression process). The three work together to improve the model's mAP, and more importantly, the Precision and Recall can grow synchronously, indicating that false negatives and false positives are suppressed at the same time. In addition, the design emphasizes heavy parameterization and lightweight structure, so that the new capabilities come more from "more effective features and loss mechanisms" rather than simply stacking up the amount of computation. Therefore, it is more suitable for deployment on the UAV end side or near-end equipment with limited computing power, so as to realize the application of real-time or near-real-time high-precision inspection / traffic monitoring / security and other applications. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the target recognition network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the MoE hybrid expert model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the C3DC module in an embodiment of the present invention; Figure 5 These are heatmap comparison images of the single image detection effect in the embodiments of the present invention, wherein (a) is the heatmap of the single image effect of YOLOv11n; and (b) is the heatmap of the single image effect of YOLOv11n-C3DC. Figure 6 This is a schematic diagram of the DEConv module in an embodiment of the present invention; Figure 7 This is a schematic diagram of the GNDE detection head in an embodiment of the present invention; Figure 8 This is a schematic diagram of CIoU in an embodiment of the present invention; Figure 9 This is a confusion matrix diagram of YOLOv11 in an embodiment of the present invention; Figure 10 This is a confusion matrix diagram of the method of the present invention; Figure 11 A schematic diagram of the F1-confidence curve of YOLOv11 in an embodiment of the present invention; Figure 12 This is a schematic diagram of the F1-confidence curve of the method of the present invention. Detailed Implementation

[0017] This embodiment provides a target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, such as... Figure 1As shown, it includes the following steps: Step 1: Acquire drone aerial images and corresponding annotation information.

[0018] Step 2: Data Augmentation and Preprocessing Size normalization: Scale the input aerial image to the network input size H×W and record the scaling ratio and padding amount.

[0019] Synchronous transformation of annotations: The true bounding boxes are scaled, cropped, and stitched together, and updated synchronously to the new coordinate system. Parts that exceed the image boundaries are cropped, and oversized and invalid boxes are removed.

[0020] Pixel normalization: Normalize pixel values ​​to [0,1] or standardize them according to mean and variance to ensure stable training values.

[0021] Enhancement operators: random flip (often horizontal flip) to improve viewpoint robustness; random crop, scaling and translation to simulate changes in scale and shooting position; color rotation (brightness, contrast and saturation, etc.) to adapt to changes in lighting.

[0022] Step 3, construct as follows Figure 2 The target recognition network shown introduces a C3DC module in its backbone network to automatically generate weights based on the current image. The C3DC module is constructed by replacing the ordinary convolutions of the Bottleneck module in the C3k2 module with DyConv convolutions. DyConv convolutions are used to combine multiple sets of convolutional kernels according to their weights. The backbone network outputs multi-scale features: small target scale feature P3, medium target scale feature P4, and large target scale feature P5. When the target region area is smaller than... When the target area is within a certain range (in pixels), it is defined as a small target; when the target area is within a certain range (in pixels), it is defined as a small target. and When the target area is between pixels, it is defined as a medium target; when the target area is larger than a certain value... When the pixel value is specified, it is defined as a large target.

[0023] Dynamic convolution is a novel convolution method that adaptively changes the attention kernel based on the input, rather than using the same weight matrix for each inference as in static, fixed convolution kernels in CNNs. Dynamic convolution allows the model to use different convolution kernels for each inference based on the input data, enabling the model to more flexibly adapt to different input features and acquire more feature information.

[0024] The traditional convolution formula is as follows: ; Where X represents the input feature map (or input image), W represents the convolution kernel weight parameters, the symbol "*" represents the convolution operation, and Y represents the convolution output feature map.

[0025] FLOPs reflect the computational complexity of a model; the lower the value, the faster the computation. However, increasing the number of parameters leads to more computation. The goal of this embodiment is to maximize the number of learnable parameters while maintaining low FLOPs. To this end, MoE (Mixture of Experts) type convolution is used (model structure as follows...). Figure 3 As shown in the figure, combined with the parameter expansion function, we design the adaptive dynamic convolution DyConv, as shown in the following formula: ; ; in, This represents the adaptive convolutional kernel generated by DyConv convolution, where M represents the number of expert convolutional kernels. This represents the weight parameters of the i-th expert convolution kernel; This represents the i-th expert weight coefficient adaptively generated from the input, which typically satisfies... and .

[0026] In drone aerial photography, small targets occupy a small percentage of pixels and are often sparse and inconspicuous on feature maps. Dynamic convolution automatically generates weights based on the current image, making it easier to capture the edges and textures of small targets than fixed convolution. When large and small targets coexist in the same image with significant scale variations, a fixed receptive field struggles to accommodate them all. DyConv convolution uses multiple sets of convolutional kernels combined with weights, essentially automatically selecting the appropriate kernel size and effective receptive field, thus better adapting to multi-scale scenes.

[0027] Based on the above considerations, an adaptive dynamic convolution DyConv is introduced based on the C3k2 module. By replacing the ordinary convolution of the Bottleneck module with the DyConv convolution of the method in this embodiment, a C3DC adaptive dynamic convolution feature extraction module is proposed.

[0028] like Figure 4As shown, in the C3DC module, the input features first undergo channel adjustment and preliminary feature extraction through a Conv layer. Then, a split layer divides the features into two branches by channel: one branch acts as a shortcut branch, directly preserving detailed information, while the other branch enters the main branch for deep modeling. The main branch first stacks several Bottlenecks using a C3k2 structure to enhance non-linear representation. At key Bottlenecks, ordinary convolutions are replaced with DyConv, allowing the convolution kernel weights to adaptively change with the input, thus dynamically enhancing the edge texture and weak features of small objects. The output of the main branch and the shortcut branch are concatenated along the channel dimension, and finally, another Conv layer completes feature integration and output. This structure maintains the lightweight bottleneck advantage of C3 classes while introducing dynamic convolution to improve adaptability to different scales and complex backgrounds.

[0029] Heatmap comparison of single image testing results, as shown below Figure 5 As shown, by comparing the heatmap results of the two methods on a single image, it can be observed that the original method has many omissions in small target recognition. However, using the improved model, the heatmap shows more comprehensive and accurate coverage of small targets. This demonstrates that the improved model is more adaptable to small target recognition tasks and effectively compensates for the shortcomings of the original method in this area.

[0030] Step 4: By fusing semantic and detail information across scales through the neck network of the target recognition network, the three features P3, P4 and P5 are fused to output the fused features.

[0031] Step 5: Introduce the GNDE module into the detection head of the target recognition network. The GNDE module includes the Conv_GN module and the DEConv module. The Conv_GN module includes convolution and GN normalization. The three features P3, P4 and P5 are first processed by the Conv_GN module for channel alignment and stable normalization. Then, they are enhanced with edge and orientation details through several layers of DEConv module. Subsequently, the classification branch and the regression branch output the class probability and bounding box parameters respectively. The regression branch is followed by Scale to scale and calibrate the regression amplitude of features at different scales.

[0032] In drone aerial images, small targets are small in size and have weak features, making it difficult for the detection head to converge stably. To improve performance, Group Normalization (GN) is used as a regularization method. GN is independent of batch size, making it suitable for small-batch training. It can mitigate the performance degradation of Batch Normalization (BN) with small batches and avoid detail blurring caused by global statistics. Even with a batch size of 1, GN can maintain a stable normalization effect.

[0033] GN divides the channel dimension into G groups, independently calculating the mean and variance in each group; for an input feature map (with dimensions N×C×H×W, where N represents the batch size, H represents the height, and W represents the width), GN divides channel C into G groups, each containing There are 4 channels; for the g-th group, the formulas for calculating its mean and variance are as follows: ; ; Where g represents the channel group index. This represents the set of all activation values ​​in group g that are included in the statistics. Represents a set The i-th activation value in the set, where m represents the set. The total number of active values; This represents the mean of group g.

[0034] GN calculates statistics only within the channel grouping of each sample, without depending on other samples in the batch; GN standardizes the activation values ​​for each group as follows: ; Finally, GN applies learnable scaling and offset operations, as shown in the following formula: ; in, Indicates to The standardized results after group normalization This represents a very small constant, used to avoid zero denominators and improve numerical stability; This represents the final output of GN. and These represent the learnable scaling factor and offset, respectively.

[0035] DEConv (Detail-enhanced convolution) aims to incorporate differential information into standard convolutions, thereby enhancing feature representation and generalization capabilities. For example... Figure 6 As shown, DEConv consists of five parallel branches: standard convolutions provide basic representations at the intensity level; central difference convolutions highlight edges and details through the difference between the neighborhood and the center pixel; angular difference convolutions enhance sensitivity to directional structures, facilitating the capture of rotationally related patterns; horizontal and vertical difference convolutions encode gradient priors on the two axes respectively, refining directional features. Finally, the outputs of each branch are fused along the channel dimension to obtain a more detailed feature representation with a more stable directional response. Through the collaboration of different types of convolutional layers, rich and discriminative feature information is extracted.

[0036] In the DEConv module, the input features are processed through the standard convolutional branch to obtain basic semantics, and then enter the center difference, angle difference, vertical difference, and horizontal difference convolutional branches in parallel to explicitly enhance edge, orientation, and gradient details. The outputs of each branch are fused at the end to obtain the final features, thereby improving the ability to represent fine-grained textures and small object contours with almost no increase in inference overhead.

[0037] Reparameterization is used to merge the outputs of multiple parallel convolutional layers into the output of a single convolutional layer, thereby reducing additional non-learning parameters and computational cost. Specifically, it assumes that a convolutional layer has the following corresponding weight kernel. (i represents the sequence number of the convolutional layer). During the training phase, this weight kernel is learned and updated independently; however, during forward propagation, DEConv adds these weight kernels together at the corresponding positions to form an equivalent weight kernel. As shown in the following formula: ; in, This represents the input feature map of DEConv; This represents the output feature map of DEConv; This represents the convolution kernel of the i-th parallel branch (corresponding to standard convolution, central difference, angular difference, horizontal difference, vertical difference, etc.); "5" indicates that the number of parallel branches is 5; This represents the equivalent convolution kernel obtained by adding the convolution kernels of each branch at their corresponding positions during the inference phase.

[0038] Based on the characteristics of GN, it is incorporated into existing ordinary convolutions, and DEConv is used as a shared convolution. This is then integrated into existing target detection heads to design a GNDE detection head, enabling efficient and accurate small target detection.

[0039] like Figure 7 As shown, in the GNDE detection head, features from different scales are first processed by Conv_GN (convolution + group normalization) for channel alignment and stable normalization, and then several layers of DEConv are used to enhance edge and orientation details. Subsequently, the features are divided into a classification branch (Conv_Cls) and a regression branch (Conv_Reg) to output the class probability and bounding box parameters respectively. The regression branch is followed by Scale to scale and calibrate the regression amplitude of features at different scales, thereby improving the stability and accuracy of small target localization.

[0040] Step 6: Combine classification and confidence losses and optimize the regression branch using HWIoU. Calculate the classification loss to constrain the target category prediction results; calculate the confidence loss to constrain the target existence probability prediction results; calculate the HWIoU regression loss to optimize the prediction box position and scale.

[0041] The bounding box loss function is a key component of the target recognition loss system, and its proper design can significantly improve the model's recognition performance. YOLOv11 uses CIoU loss in bounding box regression, which can effectively overcome the limitations of traditional metrics: when the center points of the ground truth box and the predicted box coincide, even if the intersection-union ratio is the same, accurate evaluation can be achieved through the difference in aspect ratio.

[0042] CIoU diagram as follows Figure 8 As shown, the CIoU penalty formula is defined as follows: ; Among them, b and These represent the center points of the predicted bounding box and the ground truth bounding box, respectively. represents the Euclidean distance between the two center points; c represents the diagonal length of the smallest closed region that simultaneously encloses the predicted bounding box and the ground truth bounding box. v and v represent the adaptive weighting function and the metric that describes the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box, respectively.

[0043] The loss function CIoU is defined as follows: .

[0044] In drone aerial photography, small targets have low overlap with the anchor frame, and the gradient contribution of CIoU is relatively weak. Wise-IoU, by weighting different regions of IoU, amplifies the loss weight of small targets and hard-to-detect regions, strengthens localization and regression, and is therefore better than CIoU in small target detection.

[0045] Based on the dynamic adjustment concept of WloU, a novel IoU loss function calculation rule, called HWIoU, is designed, and its expression is defined as follows: ; in, This represents the regression loss of HWIoU; This represents a dynamic weighting term introduced by the Wise-IoU concept (used to adjust the weights of samples of different quality / different stages); This represents the basic IoU regression loss term; x and y represent the coordinates of the center point of the prediction box; Represents the coordinates of the center point of the true bounding box; This represents the width and height of the actual bounding box.

[0046] Compared to CIoU and Wise-IoU, the HWIoU proposed in this embodiment introduces dynamic adjustment in the regression loss: in the early stage of training, the weights of low IoU are increased to encourage the model to improve overlap; as the IoU converges, the weights are gradually reduced to optimize the center of gravity by shifting the center offset and aspect ratio to refine the localization. At the same time, a gentle "half-subtraction" penalty is applied to geometric terms such as distance, scale, and shape to avoid excessive interference with training and improve accuracy and generalization.

[0047] Step Seven: Reasoning Stage Confidence threshold filtering: Calculate the final score for each candidate box (usually "objectivity × class probability" or direct class confidence), and eliminate those with scores below the threshold. The candidate box.

[0048] NMS removes duplicate bounding boxes (threshold) Sort by score, retain high-scoring bounding boxes and suppress those with IoU greater than 1. Overlapping boxes (can be performed separately by category) reduce duplicate detection.

[0049] Output results: The final target category, bounding box location, and corresponding confidence score are obtained.

[0050] like Figure 2 As shown, in the overall network structure, the input drone aerial image (640×640×3) first enters the Backbone part for feature extraction and downsampling: the image passes through the CBS module twice to complete basic convolutional feature extraction and resolution compression. The CBS module consists of 3×3 convolutions, normalization layers (such as BN), and activation functions, transforming the feature map size from 640×640×3 to 320×320×16 and then to 160×160×32, respectively. Subsequently, the feature map enters the C3DC module for further processing. To adapt to dynamic convolutional feature extraction, C3DC introduces the DyConv / MoE concept into the C3-class bottleneck structure. It generates adaptive convolutional kernels by weighting multiple expert convolutional kernels through attention weights, thereby adaptively enhancing the edge and texture response of small targets according to the input content. The output is still 160×160×32. Then, the feature map is downsampled to 80×80×64 by the CBS module and enters the second C3DC module to further enhance the detailed representation, resulting in the first lateral output feature of the Backbone (80×80×64).

[0051] The feature map is then downsampled to 40×40×128 using CBS and enters the C3DC module to enhance mid-scale semantic and localization features, resulting in the second lateral output (40×40×128). It is then downsampled to 20×20×256 using CBS and enters the C3DC module to enhance high-level semantic expression. After that, multi-scale pooling is performed using SPPF spatial pyramid pooling to expand the receptive field and integrate contextual information. Finally, the C2PSA attention module is used to weight and emphasize key regions in the channel and spatial dimensions and suppress complex background interference, forming the third output feature of Backbone (20×20×256). Thus, the sizes of the three scale feature maps output by Backbone are 80×80, 40×40, and 20×20, respectively.

[0052] The Neck section then performs cross-scale feature fusion: First, a top-down FPN path is executed starting from 20×20×256, upsampling to 40×40×256 and concatenating with the Backbone's 40×40×128 to form 40×40×384, which is then fused and compressed by C3DC to obtain 40×40×128; this 40×40×128 is then upsampled to 80×80×128 and concatenated with the Backbone's 80×80×64 to obtain 80×80×192, which is then fused by C3DC and output as 80×80×64 as the small target detection scale feature (P3); subsequently, a bottom-up PAN path is executed, combining P3 (80... The 40×80×64 feature is downsampled to 40×40×64 by CBS and fused with the aforementioned 40×40×128 feature to obtain 40×40×192. The fusion output of 40×40×128 by C3DC is then used as the medium target scale feature (P4). P4 (40×40×128) is then downsampled to 20×20×128 by CBS and fused with the 20×20×256 feature at the backbone end to obtain 20×20×384. The fusion output of 20×20×256 by C3DC is then used as the large target scale feature (P5). Thus, the three scale fusion features sent to the detection head are 80×80×64, 40×40×128 and 20×20×256, respectively.

[0053] Finally, in the Head section, the three features are input into the GNDE detection head for prediction. GNDE adopts a shared convolutional structure and introduces Group Normalization to adapt to small batch training and stabilize normalization. At the same time, DEConv detail enhancement convolution is used to fuse standard convolution with branches such as center difference, angular difference, horizontal difference, and vertical difference to strengthen gradient and orientation texture priors, thereby improving the separability of small target contours and details in complex backgrounds. The detection head finally outputs the class probability (p_{cls}) and bounding box parameters (b_{pred}) of candidate targets at each scale. During the training phase, classification / confidence loss is combined and HWIoU is used to optimize the regression branch to improve the effective learning of low-overlap samples of small targets. During the inference phase, the output results are filtered by confidence threshold and duplicate boxes are removed by NMS to obtain the final detection results.

[0054] The VisDrone2021 dataset was chosen for experimental validation. VisDrone2021 is a large-scale visual dataset built by the AISKYEYE team at the Machine Learning and Data Mining Laboratory of Tianjin University. It focuses on processing complex computer vision tasks in drone images and videos, including target categories such as pedestrians, vehicles, and bicycles. It covers various tasks in target detection, tracking, counting, and attribute analysis, involving different types of targets and analysis tasks in various environments. This dataset contains high-resolution images and video footage with detailed annotations on object location, category, and motion. It emphasizes the difficulty and challenges in variable weather conditions, the presence of targets of significantly different scales, and high-density clustering, promoting the development and improvement of visual analysis capabilities and related technologies in complex environments. The objects in the images within this dataset possess scene object characteristics consistent with the present invention; therefore, this dataset was chosen for experimental validation.

[0055] The experiment used Python 3.10.15 as the programming language and Ubuntu 18.04 LTS as the operating system. The experimental hardware was an Intel(R) Xeon(R) Gold 5117 CPU@2.00GHz. The algorithm was implemented based on the PyTorch 1.12.1 framework. The deep learning environment was CUDA 11.4, and GPU acceleration was performed using an NVIDIA RTX 4090 graphics card (24GB VRAM).

[0056] A confusion matrix is ​​used to visually display a model's performance in classification tasks. In small object recognition, the confusion matrix can analyze the model's ability to identify different object categories. In small object recognition tasks, the objects are small and their features are not obvious, making it easy for the model to miss or falsely detect targets. The confusion matrix can overcome these problems to some extent.

[0057] To more clearly compare the classification accuracy of YOLOv11 and the method of this embodiment, the confusion matrix was standardized. On the VisDrone2021 dataset, the confusion matrices generated by YOLOv11 and the method of this invention are as follows: Figure 9 and Figure 10 As shown.

[0058] contrast Figure 9 and Figure 10 The two confusion matrix diagrams show that the improved model in this embodiment is more accurate in classifying multiple target types: people improved from 16% to 18%; bicycles from 6% to 8%; vehicles from 72% to 73% (with fewer cross-class misclassifications); trucks from 21% to 24%; buses from 34% to 44%; and motorcycles from 28% to 30%. Overall, the model in this embodiment better grasps the characteristics of people and vehicles in small targets and complex backgrounds. At the same time, the background misclassification rate decreased: pedestrians decreased from 68% to 65%, bicycles from 83% to 80%, trucks from 49% to 48%, and the background misclassification rate for buses decreased to 41%. This indicates that the model in this embodiment is clearer in background differentiation and has more stable and accurate overall recognition.

[0059] The F1 score is an important metric for evaluating model performance in classification tasks. It is the harmonic mean of precision and recall, and is calculated using the following formula: ; The F1 score ranges from 0 to 1, with a higher value indicating better classification performance for that category. The F1 score strikes a balance between precision and recall. Therefore, in situations with uneven data distribution or high misclassification costs, the F1-confidence curve is used to observe the model's performance at different confidence thresholds, allowing for the selection of an optimal confidence threshold to improve overall recognition capabilities.

[0060] like Figure 11 and Figure 12 As shown, comparing the two F1-confidence curves, YOLOv11 peaks at approximately 0.36 when the confidence level is approximately 0.146; the improved model in this embodiment increases the peak value to approximately 0.39 when the confidence level is approximately 0.173. Overall, the method in this embodiment outperforms the baseline in terms of overall F1 score and reaches its optimum at a higher threshold; its confidence level matches the prediction quality better, reducing misjudgments caused by excessively high or low confidence levels, resulting in more robust and reliable decision-making.

[0061] To further validate the improved algorithm, it was compared with mainstream one-stage / two-stage detectors. Evaluation metrics included recognition accuracy, number of parameters, and computational cost. The participating models were: SSD, RetinaNet, Faster R-CNN, CenterNet, Cascade, RT-DETR, ASF-YOLO, and YOLOv3 / 5 / 8 / 11. The YOLO series employed multi-scale variants ("n / s / m / l / x") to balance model complexity, parameters, FLOPs, accuracy, and inference speed. YOLO-"n" was used as the baseline, with additional "s" and "x" scales selected for cross-sectional evaluation; the comparison results are shown in Table 1.

[0062] Table 1 Comparison Experiments of Classic Algorithms

[0063] As can be seen from Table 1, the improved models in this embodiment all show improvement compared to other models: Ours-x has mAP@50 and mAP@50:95 of 48.5% and 30.2% respectively, which is significantly better than the other methods, indicating that it has more advantages in multi-scale and more difficult targets; Ours-n has mAP@50 and mAP@50:95 of 35.1% and 25.0% respectively, which is the best performance among YOLO variants of the same size.

[0064] In terms of parameters and computational overhead: Ours-n has only 2.2M parameters (less than YOLOv5n's 7.2M) and 23.3G FLOPs, making it suitable for resource-constrained devices; Ours-x has 65.3M parameters and 223.7G FLOPs (less than YOLOv5x's 249.3G), controlling the computational load while maintaining high accuracy.

[0065] Overall, the improved model in this embodiment has high accuracy and efficiency in small target identification in UAV aerial photography: Ours-x is more computationally efficient while maintaining accuracy, and Ours-n is lightweight and efficient. The two achieve a good balance between performance and overhead and have strong application potential.

[0066] To verify the advantages of HWIoU in bounding box regression, using CIoU as the baseline, we compared small target recognition on YOLOv11 by replacing CIoU with GIoU, SIoU, and WIoU v3, as shown in Table 2. The results show that HWIoU performs best in Precision, Recall, and mAP50, at 42.1%, 34.3%, and 34.0%, respectively.

[0067] Table 2. Comparison Experiment of Bounding Box Loss

[0068] To verify the effectiveness of each module in the method of this embodiment and the advantages of their combination, a systematic ablation experiment was conducted. A YOLOv11 model of size n was used to ensure that the model has sufficient learning capacity without excessively consuming computational resources. The ablation experiment results are shown in Table 3, which illustrates the performance changes after introducing the C3DC, GNDE, and HWIoU modules, intuitively reflecting the contribution of each module to the model performance.

[0069] Table 3 Analysis of Ablation Experiment Results

[0070] In Table 3, “√” indicates that the method was used in the YOLOv11n model, C3DC represents the dynamic feature extraction structure, GNDE represents the detail enhancement shared convolution, and HWIoU represents the optimized loss function.

[0071] As shown in Table 3, when using only the YOLOv11n model, the precision is 42.1%, recall is 33.4%, mAP@50 is 32.4%, and mAP@50:95 is 18.5%. After adding the Dynamic Feature Extraction Structure (C3DC) to the YOLOv11n model, the model's precision improved by 1.0%, recall decreased by 0.2%, mAP@50 improved by 0.5%, and mAP@50:95 improved by 0.6%. After introducing the Enhanced Details Shared Convolutional Module (GNDE), the recall increased by 1.0%, mAP@50 increased by 1.3%, and mAP@50:95 increased by 1.2%.

[0072] After simultaneously adding C3DC and GNDE modules to the model, the precision and recall reached 45.3% and 35.1%, respectively, with mAP@50 increasing to 34.4% and mAP@50:95 increasing to 20.0%. Finally, an optimized loss function, HWIoU, was introduced to form the final improved model M5. Compared with the baseline YOLOv11n, precision improved by 4.1%, recall by 1.9%, mAP@50 by 2.7%, and mAP@50:95 by 2.2%. mAP improved at different IoU thresholds, with precision and recall simultaneously enhanced.

[0073] To address the limitation in target recognition performance in UAV aerial images due to their small size, low resolution, and complex backgrounds, this embodiment proposes a target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement. While maintaining low computational complexity, improvements are made in three aspects: feature extraction improvement, detector head reconstruction, and loss function optimization. To more fully capture fine-grained information and enhance the representation of small targets, adaptive dynamic convolution (DynamicConv) is introduced into the C3k2 network structure, the Bottleneck module is reconstructed, and a C3DC dynamic convolution feature extraction module is constructed. To enhance the boundaries and textures of small targets, Group Normalization is combined with convolution, and DEConv is used as a shared convolution, designing a detail-enhancing shared convolutional detector head (GNDE). To address the issues of low-quality samples weakening generalization and geometric metrics interfering with training, a novel IoU loss calculation rule, HWIoU, is proposed, enhancing the generalization ability of the method in this embodiment.

[0074] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement, characterized in that: Includes the following steps: Step S1: Acquire drone aerial images and corresponding annotation information; Step S2: Construct a target recognition network. In the backbone network of the target recognition network, introduce the C3DC module for automatically generating weights based on the current image. The C3DC module is constructed by replacing the ordinary convolution of the Bottleneck module in the C3k2 module with DyConv convolution. The DyConv convolution is used to combine multiple sets of convolution kernels according to weights. The backbone network outputs multi-scale features, namely small target scale feature P3, medium target scale feature P4, and large target scale feature P5. When the target area is smaller than When the target area is in pixels, it is defined as a small target; when the target area is within a certain range... and When the target area is between pixels, it is defined as a medium target; when the target area is larger than a certain value... When the target size is in pixels, it is defined as a large target. Step S3: By fusing semantic and detail information across scales through the neck network of the target recognition network, the three features P3, P4 and P5 are fused to output the fused features; Step S4: Introduce the GNDE module into the detection head of the target recognition network. The GNDE module includes the Conv_GN module and the DEConv module. The Conv_GN module includes convolution and GN normalization. The three features P3, P4 and P5 are first aligned and stabilized by the Conv_GN module, and then the edge and orientation details are enhanced by several layers of DEConv module. Subsequently, the classification branch and the regression branch output the class probability and bounding box parameters respectively. The regression branch is followed by Scale to scale and calibrate the regression amplitude of features at different scales. Step S5: Combine classification and confidence loss and optimize the regression branch using HWIoU; Step S6: Filter candidate boxes based on confidence scores and remove duplicate boxes using NMS to obtain the final detection results, including the final target category, bounding box position, and corresponding confidence score.

2. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 1, characterized in that: In step S2, MoE-style convolution combined with a parameter expansion function is used to design DyConv convolution, as shown in the following formula: ； ； Where X represents the input feature map, Y represents the convolution output feature map, and * represents the convolution operation. This represents the adaptive convolution kernel generated by DyConv convolution, where M represents the number of expert convolution kernels. This represents the weight parameters of the i-th expert convolution kernel; This represents the i-th expert weight coefficient adaptively generated from the input, satisfying... and .

3. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 1, characterized in that: In step S2, the drone aerial image enters the backbone network for feature extraction and downsampling in sequence. The image passes through the CBS module twice to complete basic convolutional feature extraction and resolution compression; then through the C3DC module for adaptive dynamic convolutional feature extraction; and then through the CBS module for downsampling before entering the second C3DC module to obtain the first lateral output feature of the backbone network. After further downsampling by the CBS module, the signal enters the third C3DC module to obtain the second lateral output feature. After further downsampling by the CBS module, the signal enters the fourth C3DC module. Then, multi-scale pooling is performed through SPPF spatial pyramid pooling, and finally, the signal enters the C2PSA attention module to form the third output feature of the backbone network.

4. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 3, characterized in that: In step S3, the output of the backbone network enters the neck network for cross-scale feature fusion. First, starting from the third output feature of the backbone network, a top-down FPN path is executed, which upsamples the feature and concatenates it with the output of the third C3DC module in the backbone network. Then, it is fused and compressed by the fifth C3DC module. Next, it is upsampled and concatenated with the output of the second C3DC module in the backbone network. After fusion by the sixth C3DC module, the small target detection scale feature P3 is output. Then, a bottom-up PAN path is executed. Feature P3 is downsampled by the CBS module and concatenated with the output of the fifth C3DC module. Then, it is fused by the seventh C3DC module to output the medium target scale feature P4. Feature P4 is downsampled by the CBS module and concatenated with the third output feature of the backbone network. After fusion by the eighth C3DC module, the large target scale feature P5 is output.

5. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 4, characterized in that: The CBS module includes a 3×3 convolution, a normalization layer, and an activation function. In the C3DC module, the input features first pass through a Conv layer for channel adjustment and feature extraction. Then, the features are divided into two branches by channel using a split method. One branch acts as a shortcut branch to directly retain detailed information, while the other branch enters the main branch for modeling. The main branch sequentially passes through several C3k2 modules to enhance the nonlinear expression. The output of the main branch and the output of the shortcut branch are concat-fused along the channel dimension. Finally, the features are integrated and output after passing through another Conv layer.

6. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 5, characterized in that: In the C3k2 module, the input is fused by Concat after passing through two parallel branches. In one branch, the input passes through only one Conv layer, while in the other branch, the input passes through one Conv layer and then several Bottleneck modules. In the Bottleneck module, the input passes through two DyConv convolutions in sequence and then is fused with the input itself by add.

7. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 1, characterized in that: In the Conv_GN module of step S4, GN is used as regularization. GN divides the channel dimension into G groups, and independently calculates the mean and variance in each group. For an input feature map of size N×C×H×W, where N represents the batch size, H represents the height, and W represents the width, GN divides channel C into G groups, each containing... There are 4 channels; for the g-th group, the formulas for calculating its mean and variance are as follows: ； ； Where g represents the channel group index. This represents the set of all activation values ​​in group g that are included in the statistics. Represents a set The i-th activation value in the set, where m represents the set. The total number of active values; This represents the mean of the g-th group; GN calculates statistics only within the channel grouping of each sample. GN standardizes the activation values ​​of each group as follows: ； Finally, GN applies learnable scaling and offset operations, as shown in the following formula: ； in, Indicates to The standardized results after group normalization This represents a very small constant used to avoid the denominator being zero. This represents the final output of GN. and These represent the learnable scaling factor and offset, respectively.

8. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 1, characterized in that: In the DEConv module of step S4, the input features are processed by standard convolution, and simultaneously by center difference, angular difference, vertical difference and horizontal difference convolution in parallel. The outputs of each convolution are fused at the end in the channel dimension to obtain the final features.

9. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 8, characterized in that: The DEConv module uses reparameterization to merge the outputs of multiple parallel convolutional layers into the output of a single convolutional layer; assuming a convolutional layer has a corresponding weight kernel. , where i represents the sequence number of the convolutional layer; exist During the training phase, weight kernels They are learned and updated independently; during forward propagation, the DEConv module adds these weight kernels at the corresponding positions to form an equivalent weight kernel. As shown in the following formula: ； in, This represents the input feature map of the DEConv module. This represents the output feature map of the DEConv module; This represents the convolution kernel of the i-th parallel branch, corresponding to standard convolution, central difference, angular difference, horizontal difference, and vertical difference convolution; 5 indicates that the number of parallel branches is 5; This represents the equivalent convolution kernel obtained by adding the convolution kernels of each branch at their corresponding positions during the inference phase.

10. The target recognition method based on adaptive dynamic convolution feature extraction and detail enhancement according to claim 1, characterized in that: In step S5, HWIoU is constructed based on the dynamic adjustment method of WloU, as shown in the following formula: ； in, This represents the regression loss of HWIoU; This represents a dynamic weighting term, used to adjust the weights of samples of different quality and at different stages. This represents the basic IoU regression loss term; x and y represent the coordinates of the center point of the prediction box; Represents the coordinates of the center point of the true bounding box; Represents the width and height of the actual bounding box; v and v represent the adaptive weighting function and the metric that describes the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box, respectively.

Images