A small target detection method based on refined shallow features

Abstract

The application relates to the technical field of computer vision, and provides a small target detection method based on refined shallow layer features, which comprises the following steps: extracting a multi-scale feature map of an unmanned aerial vehicle aerial image through a backbone feature extraction network; adaptively optimizing a P2 feature map through an SCM module to obtain a refined P2 feature map; constructing a neck network of a bidirectional feature aggregation path; performing multi-scale feature fusion on the refined P2 feature map, P3 layer features, P4 layer features and P5 layer features through a top-down semantic injection path and a bottom-up detail enhancement path to generate a detection feature map containing deep layer semantic information and shallow layer spatial details; and performing target detection through the detection feature map to obtain a target detection result. The application realizes effective reservation of small target spatial details and suppression of background noise, thereby significantly improving the small target detection precision and robustness in an unmanned aerial vehicle scene.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for small target detection based on refined shallow features. Background Technology

[0002] Images captured from a drone's perspective typically face severe challenges, including extremely small target scale, dense distribution, and complex and ever-changing background environments. Unlike conventional scenarios, the core challenge of drone aerial image detection lies in the inherent contradiction between deep semantic abstraction and the preservation of fine-grained spatial details. While standard detection frameworks perform well in general scenarios, when processing tiny targets occupying only a few pixels, the frequent downsampling operations of the backbone network cause irreversible loss of the discriminative features of these targets before they are passed to the deep detection head.

[0003] To address the issue of feature vanishing in small targets, researchers have proposed a series of improved schemes, but significant shortcomings remain. For image-level enhancement strategies, existing techniques often employ tiled inference or slice-assisted hyper-inference (SAHI). Even the latest SAHI v2, which optimizes the processing flow, significantly increases computational load and inference latency due to multi-stage repeated inference, making it difficult to meet the stringent real-time processing requirements of UAV embedded platforms. For multi-scale feature fusion methods, traditional Feature Pyramid Networks (FPNs) and their variants attempt to fuse features via a top-down path. However, standard architectures typically only perform detection at feature layers (P3-P5) with strides of 8, 16, and 32. For targets of 16x16 pixels or even smaller, P3-level features tend to become blurred. Even YOLOv12, proposed in 2025 with an attention mechanism, still prioritizes deep semantic enhancement, resulting in a lack of effective high-resolution spatial priors when dealing with extremely small targets. Methods that introduce shallow detection branches attempt to recover details by introducing even shallower P2 detection branches (with a stride of 4). While shallow feature layers preserve spatial details, they are also filled with a large amount of background noise, such as vegetation and road texture interference. Existing methods lack effective feature filtering mechanisms, which leads to the erroneous amplification of noise signals, resulting in a significant increase in false alarm rates.

[0004] Existing methods still face an imbalance between detection sensitivity and anti-interference capability. On the one hand, the high-resolution features introduced to capture small targets come with huge computational overhead. For example, while the latest RF-DETR has excellent accuracy, its complex Transformer structure makes it difficult to deploy in real time on edge devices. On the other hand, shallow noise limits the model's generalization ability to complex environments. In addition, existing attention mechanisms are often complex in design, making it difficult to achieve joint refinement of the spatial location and channel features of small targets while maintaining real-time performance. Summary of the Invention

[0005] This invention aims to address at least one of the technical problems existing in related technologies. To this end, this invention provides a small target detection method based on refined shallow features, which effectively preserves spatial details of small targets and suppresses background noise, thereby significantly improving the accuracy and robustness of small target detection in UAV scenarios.

[0006] This invention provides a small target detection method based on refined shallow features, comprising: S1: Acquire drone aerial images, extract feature maps of different scales from the drone aerial images through a backbone feature extraction network, and obtain multi-scale feature maps, including P2 feature map, P3 feature map, P4 feature map and P5 feature map. S2: Adaptive optimization of the P2 feature map is performed through the spatial-channel feature refinement module to obtain the refined P2 feature map; S3: Construct the neck network of the bidirectional feature aggregation path, which includes a top-down semantic injection path and a bottom-up detail enhancement path. S4: Upsample the P5 feature map through a top-down semantic injection path and concatenate it with the P4 feature map to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. S5: The P2 fusion feature map is convolved through a bottom-up detail enhancement path and then concatenated with the P3 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and then concatenated with the P5 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P5 detection feature map. S6: Target detection is performed based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain the target detection result.

[0007] Furthermore, the backbone feature extraction network includes a P2 feature layer with a downsampling step size of 4, a P3 feature layer with a downsampling step size of 8, a P4 feature layer with a downsampling step size of 16, and a P5 feature layer with a downsampling step size of 32.

[0008] Furthermore, the spatial-channel feature refinement module takes the shallow feature map as input and enhances the effective information in the shallow feature map and suppresses irrelevant background noise through parallel channel feature modeling and spatial feature modeling mechanisms.

[0009] Furthermore, in the process of channel feature modeling, the input features are subjected to global average pooling in the spatial dimension to extract channel-level statistical information, and channel attention weights are generated through a bottleneck mapping structure composed of two one-dimensional convolutions.

[0010] Furthermore, in the process of spatial feature modeling, by performing max pooling and average pooling operations on the channel dimension respectively, spatial response information from different statistical perspectives is obtained, and the information is then concatenated and input into a convolution operation with a large receptive field to generate a spatial attention weight map.

[0011] Furthermore, spatial channel enhancement features are obtained based on channel attention weights, and spatial enhancement features are obtained based on spatial attention weights. The spatial channel enhancement features, spatial enhancement features, and P2 feature maps are then superimposed using parallel residual fusion to generate refined P2 features.

[0012] Furthermore, in step S6, large-scale targets are detected using the P5 branch detection head based on the P5 detection feature map; Medium-scale targets are detected using the P4 branch detection head and based on the P4 detection feature map. The P3 branch detection head is used to detect common small targets based on the P3 detection feature map. By refining the P2 branch detection head, extremely small targets from the perspective of the UAV are detected based on the P2 fused feature map.

[0013] The present invention also provides a small target detection system based on refined shallow features, for performing the above-mentioned small target detection method based on refined shallow features, comprising: The feature extraction module acquires aerial images taken by the UAV and extracts feature maps of different scales from the UAV aerial images through a backbone feature extraction network to obtain multi-scale feature maps, which include P2 feature map, P3 feature map, P4 feature map and P5 feature map. The spatial-channel feature refinement module adaptively optimizes the P2 feature map to obtain a refined P2 feature map. The construction module constructs a neck network for a bidirectional feature aggregation path, which includes a top-down semantic injection path and a bottom-up detail enhancement path. The semantic injection module upsamples the P5 feature map and concatenates it with the P4 feature map through a top-down semantic injection path to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. The detail enhancement module performs convolution on the P2 fusion feature map through a bottom-up detail enhancement path, concatenates it with the P3 fusion feature map processed by the A2C2f module, and inputs it into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P4 detection feature map. Finally, the P4 detection feature map is convolved with the P5 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P5 detection feature map. The target detection module performs target detection based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain target detection results.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a small target detection method based on refined shallow features as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a small target detection method based on refined shallow features as described above.

[0016] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: This invention significantly improves the detection sensitivity of tiny targets in drone aerial photography scenarios. By constructing a high-resolution P2 detection branch with a step size of 4, it effectively compensates for the spatial information loss problem caused by frequent downsampling in standard YOLO architectures (such as P3-P5 layers). This invention enables tiny targets to retain more pixel definitions and alignment points on the feature map, fundamentally solving the problem of extremely small objects "disappearing" in deep networks.

[0017] This invention effectively suppresses background noise interference caused by shallow high-resolution features. Addressing the low signal-to-noise ratio (SNR) issue often associated with high-resolution feature layers, this invention designs a lightweight spatial-channel feature refinement module (SCM). Through the parallel attention fusion mechanism of SCM, the network can adaptively filter shallow features, effectively removing background clutter such as tree textures and water ripples while preserving the target's geometric structure, significantly reducing false detection and false alarm rates.

[0018] This invention achieves a deep balance between high-precision detection and the real-time requirements of edge computing. By utilizing the SCM module and an efficient feature reconstruction strategy, it improves accuracy while strictly controlling computational overhead. Compared to models such as RF-DETR that rely on complex global attention mechanisms, this invention significantly reduces inference latency and memory usage while maintaining excellent performance, making it more suitable for resource-constrained edge platforms such as unmanned aerial vehicle (UAV) terminals.

[0019] This invention enhances the robustness of multi-scale target detection in complex environments. By constructing a full-scale feature path from P5 to P2 and supplementing it with a bidirectional feature aggregation strategy, it achieves deep synergy between deep semantic information and shallow fine details. This not only improves the recall rate of small targets, but also enhances the model's generalization ability in extreme aerial photography environments such as different lighting and dense occlusion through the guidance of global context.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a small target detection method based on refined shallow features provided by the present invention.

[0023] Figure 2This is a schematic diagram of the network structure of a small target detection method based on refined shallow features provided by the present invention.

[0024] Figure 3 This is a schematic diagram of target detection results in a small target scenario according to an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram comparing the detection results of an embodiment of the present invention in a scenario with densely distributed targets.

[0026] Figure 5 This is a schematic diagram comparing the target detection performance under complex lighting conditions according to an embodiment of the present invention.

[0027] Figure 6 This is a visual comparison diagram of the detection results of different target detection methods under complex lighting conditions according to embodiments of the present invention.

[0028] Figure 7 This is a schematic diagram of the feature response on the multi-scale prediction layer in an embodiment of the present invention.

[0029] Figure 8 This is a schematic diagram of the structure of a small target detection system based on refined shallow features provided by the present invention.

[0030] Figure 9 This is a structural block diagram of the electronic device provided by the present invention.

[0031] Figure label: 101. Feature Extraction Module; 102. Spatial-Channel Feature Refinement Module; 103. Construction Module; 104. Semantic Injection Module; 105. Detail Enhancement Module; 106. Target Detection Module; 201. Processor; 202. Communication Bus; 203. Communication Interface; 204. Memory. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0033] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0034] The following is combined with Figures 1 to 9 This invention describes a small target detection method based on refined shallow features.

[0035] To address the characteristics of small-scale, densely distributed targets and complex backgrounds in UAV aerial images, this invention proposes a high-resolution feature-refined target detection model based on an improved YOLO architecture. This model introduces a high-resolution shallow detection branch on top of the existing detection network and combines a feature refinement mechanism that uses spatial and channel joint modeling to effectively preserve spatial details of small targets and suppress background noise, thereby significantly improving the detection accuracy of small targets in UAV scenarios.

[0036] like Figure 1 As shown, a small target detection method based on refined shallow features includes: S1: Acquire drone aerial images, extract feature maps of different scales from the drone aerial images through a backbone feature extraction network, and obtain multi-scale feature maps, including P2 feature map, P3 feature map, P4 feature map and P5 feature map. The overall network structure is as follows Figure 2 As shown, the input drone aerial image is first processed by a backbone feature extraction network for multi-layer feature encoding, generating feature maps with different spatial resolutions. The backbone feature extraction network includes a P2 feature layer with a downsampling stride of 4, a P3 feature layer with a downsampling stride of 8, a P4 feature layer with a downsampling stride of 16, and a P5 feature layer with a downsampling stride of 32. A spatial-channel joint attention module is introduced in the shallow stages of the backbone feature extraction network to enhance fine-grained feature representation, and a high-resolution detection branch is added to the feature pyramid structure to reduce the resolution difference between shallow features and deep semantic features.

[0037] Compared to existing YOLO-like detection models that typically only perform target detection at the P3 level and above, this invention introduces a P2 feature layer with a downsampling step size of 4 in the shallow layers of the backbone network, so that small targets can still maintain sufficient spatial resolution in the feature space.

[0038] S2: Adaptive optimization of the P2 feature map is performed through the spatial-channel feature refinement module to obtain the refined P2 feature map; A C3k2 module is configured in the P2 layer. The P2 feature map output by the C3k2 module does not directly enter the neck network. Instead, it is adaptively optimized by a Spatial-Channel Feature Refinement (SCM) module before entering the detector head. The Spatial-Channel Feature Refinement module uses the P2 feature map... For input, By using parallel channel feature modeling and spatial feature modeling mechanisms, effective information in the features is enhanced and irrelevant background noise is suppressed.

[0039] A dual-path parallel structure is used to adaptively weight the features.

[0040] In the channel feature modeling process, SCM performs global average pooling on the input features in the spatial dimension to extract channel-level statistical information, and generates channel attention weights through a bottleneck mapping structure consisting of two one-dimensional convolutions. The calculation expression is as follows: in, For channel attention weights, For P2 feature map, This is a global average pooling operation. For dimensionality reduction convolutional layers, To modify the activation function of the linear unit, For up-dimensional convolutional layers, Use the Sigmoid activation function; The network can adaptively emphasize feature channels that are relevant to the target semantics.

[0041] In the spatial feature modeling process, SCM performs max pooling and average pooling operations on the channel dimension to obtain spatial response information from different statistical perspectives. This information is then concatenated and input into a convolutional operation with a large receptive field to generate a spatial attention weight map. The calculation expression is as follows: in, Spatial attention weights, Max pooling for the channel dimension. For channel-dimensional average pooling, for Large kernel convolution operation; Using a larger convolution kernel size can effectively cover the typical spatial scale of small targets in high-resolution feature layers in UAV scenarios, thereby more accurately distinguishing target structures from random background textures.

[0042] Spatial channel enhancement features are obtained based on channel attention weights, and spatial enhancement features are obtained based on spatial attention weights. These features are then superimposed with the P2 feature map using parallel residual fusion to generate refined P2 features. The calculation expression is as follows: in, This is the refined P2 feature map. This is an element-wise multiplication operation.

[0043] SCM cleans and enhances P2 features, using spatial and channel attention mechanisms to suppress background noise and generate refined P2 feature maps.

[0044] This design achieves explicit enhancement of key target areas while ensuring feature integrity.

[0045] S3: Construct the neck network of the bidirectional feature aggregation path, which includes a top-down semantic injection path and a bottom-up detail enhancement path. This invention constructs a bidirectional feature aggregation path in the feature fusion stage, enabling the transmission of fine spatial information from the shallow layer to the higher layer, while the global semantic information from the higher layer can be fed back into the shallow features, thereby ensuring the consistency and stability of target detection at different scales.

[0046] The neck network employs a structured bidirectional feature fusion strategy that combines Feature Pyramid Network (FPN) and Path Aggregation Network (PANet) to promote full interaction between deep semantic information and shallow spatial information.

[0047] S4: Upsample the P5 feature map through a top-down semantic injection path and concatenate it with the P4 feature map to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. The refined P2 feature map is fused with semantic features from higher levels to enhance the semantic discriminative power of high-resolution features. Specifically, the P3 fused feature map is input into the A2C2f module and concatenated with the refined P2 feature map along the channel dimension. The fusion is then completed through a feature recalibration module, and the calculation expression is as follows: in, For P2 fusion feature maps, For upsampling operation, This is for channel splicing operations. For the feature fusion and recalibration module, This is a P3 fusion feature map.

[0048] This step allows high-resolution features to retain spatial details while obtaining semantic guidance from deep networks, effectively reducing the probability of false detections.

[0049] The A2C2f module utilizes its internal attention mechanism and gradient flow optimization capabilities to shuffle and recombine the spliced ​​features.

[0050] The A2C2f module utilizes its internal attention mechanism and gradient flow optimization capabilities to shuffle and recombine the spliced ​​features, generating the final P2 fused feature map.

[0051] S5: The P2 fusion feature map is convolved through a bottom-up detail enhancement path and then concatenated with the P3 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and then concatenated with the P5 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P5 detection feature map. S6: Target detection is performed based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain small target detection results.

[0052] This invention is equipped with four detection heads of different scales for predicting the category and location of targets on feature maps of different resolutions; Large-scale targets are detected using the P5 branch detection head based on the P5 detection feature map, with a step size of 32. The P4 branch detection head is used to detect medium-scale targets based on the P4 detection feature map, with a step size of 16. The P3 branch detection head is used to detect regular small targets based on the P3 detection feature map, with a step size of 8. By refining the P2 branch detection head, extremely small targets from the perspective of the UAV are detected based on the P2 fused feature map, with a step size of 4.

[0053] The refined P2 branch detection head retains extremely high spatial resolution (downsampled by only 4 times) and has undergone noise filtering and deep semantic guidance by SCM, thus it is specifically designed for detecting tiny objects from the perspective of drones.

[0054] like Figure 8 As shown, a small target detection system based on refined shallow features is used to perform the aforementioned small target detection method based on refined shallow features, including: The feature extraction module 101 acquires aerial images taken by the UAV and extracts feature maps of different scales from the UAV aerial images through the backbone feature extraction network to obtain multi-scale feature maps, which include P2 feature map, P3 feature map, P4 feature map and P5 feature map. The P2 feature map is adaptively optimized by the spatial-channel feature refinement module 102 to obtain the refined P2 feature map. Module 103 constructs a neck network for bidirectional feature aggregation paths, which includes a top-down semantic injection path and a bottom-up detail enhancement path. The semantic injection module 104 upsamples the P5 feature map and concatenates it with the P4 feature map through a top-down semantic injection path to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. The detail enhancement module 105 convolves the P2 fusion feature map through a bottom-up detail enhancement path, concatenates it with the P3 fusion feature map processed by the A2C2f module, and inputs it into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P4 detection feature map. The P4 detection feature map is then convolved with the P5 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P5 detection feature map. The target detection module 106 performs target detection based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain the target detection result.

[0055] Through the collaborative work of the aforementioned modules, this invention significantly improves the detection sensitivity of small targets in UAV aerial photography scenarios. By constructing a high-resolution P2 detection branch with a stride of 4, it effectively compensates for the spatial information loss problem caused by frequent downsampling in standard YOLO architectures (such as P3-P5 layers). This invention allows small targets to retain more pixel definitions and alignment points on the feature map, fundamentally solving the problem of extremely small objects "disappearing" in deep networks. This invention effectively suppresses background noise interference caused by shallow high-resolution features. Addressing the low signal-to-noise ratio problem often associated with high-resolution feature layers, this invention designs a lightweight Spatial-Channel Feature Refinement (SCM) module. Through the parallel attention fusion mechanism of SCM, the network can adaptively filter shallow features, effectively filtering out background clutter such as tree textures and water ripples while preserving the target's geometric structure, significantly reducing false detection and false alarm rates. This invention achieves a deep balance between high-precision detection and the real-time requirements of edge computing. By utilizing the SCM module and efficient feature reconstruction strategies, it improves accuracy while strictly controlling computational overhead. Compared to models like RF-DETR that rely on complex global attention mechanisms, this invention significantly reduces inference latency and memory usage while maintaining excellent performance, making it more suitable for resource-constrained edge platforms such as UAV-borne terminals. This invention enhances the robustness of multi-scale target detection in complex environments. By constructing a full-scale feature path from P5 to P2 and employing a bidirectional feature aggregation strategy, it achieves deep synergy between deep semantic information and shallow fine details. This not only improves the recall rate of small targets but also enhances the model's generalization ability under extreme aerial photography environments such as different lighting conditions and dense occlusion through the guidance of global context.

[0056] In practical applications, users first acquire drone aerial image data, normalize the image size, and then input it into the target detection system of this invention. The input image sequentially passes through a backbone feature extraction network, a high-resolution feature refinement module, and a multi-scale feature aggregation network, achieving step-by-step modeling from low-level texture information to high-level semantic information. By introducing a high-resolution P2 detection branch, the system can accurately locate and identify small, densely distributed targets while maintaining real-time detection capabilities, ultimately outputting the target's category information and its corresponding spatial location. The entire detection process is an automatic inference process, requiring no additional image cropping or multi-stage post-processing, making it suitable for deployment on real-time drone platforms.

[0057] To verify the effectiveness of the present invention, by means of... Figures 3 to 7The comparative experimental results clearly demonstrate that this invention exhibits significant technical advantages in complex UAV scenarios. Experimental results show that this invention maintains stable and accurate detection performance in scenarios with strong background interference, dense target distribution, and small-scale targets. By introducing high-resolution features, the model significantly improves the localization capability of small targets; through the spatial-channel joint feature refinement mechanism, the interference of background noise on the detection results is effectively suppressed, thereby achieving a simultaneous improvement in the accuracy and robustness of small target detection in UAV aerial images (GT).

[0058] like Figure 3 As shown in the diagram, YOLOv8-s, YOLOv11-s, YOLOv12-s, and the target detection results of the present invention in small target scenarios demonstrate that the present invention can effectively detect small targets in UAV aerial images containing a large number of small targets. The detection results show accurate target localization and few false detections, reflecting the stability and reliability of the method of the present invention in small target scenarios.

[0059] like Figure 4 As shown in the diagram, the detection results of YOLOv8-s, YOLOv11-s, YOLOv12-s, and the present invention in a densely distributed target scenario are compared. This demonstrates the detection performance of the present invention in a high-density target area. It can be seen that the present invention has significant advantages in target localization accuracy and the ability to distinguish adjacent targets, effectively reducing missed detections and false detections.

[0060] like Figure 5 As shown in the diagram, the target detection performance of YOLOv8-s, YOLOv11-s, YOLOv12-s, and the present invention under complex lighting conditions is compared. This demonstrates the detection results of various detection methods in scenarios with drastic changes in lighting. Among them, the present invention performs better in terms of maintaining target integrity and positioning stability.

[0061] like Figure 6 As shown in the diagram, the detection results of YOLOv8-s, YOLOv11-s, YOLOv12-s, and different target detection methods of the present invention under complex lighting conditions are visualized and compared. Through the heat map and the marked areas in the actual detection and inspection, it can be seen that the present invention can still stably detect small targets with fine structures under extreme environments such as strong light, weak light, and uneven lighting.

[0062] like Figure 7As shown, this invention visualizes the feature responses on the multi-scale prediction layer, illustrating the differences in response between the high-resolution prediction layer and the medium and low-resolution prediction layers to small targets. The results demonstrate that the high-resolution prediction layer exhibits a more concentrated and clearer response to small target regions, effectively mitigating the feature degradation problem and validating the necessity and effectiveness of introducing a high-resolution detection branch.

[0063] Figure 9 An example is a block diagram of an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 201, a communication interface 203, a memory 204, and a communication bus 202. The processor 201, communication interface 203, and memory 204 communicate with each other via the communication bus 202. The processor 201 can call logical instructions from the memory 204 to execute a small target detection method based on refined shallow features.

[0064] Furthermore, the logical instructions in the aforementioned memory 204 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to execute a small target detection method based on refined shallow features provided by the above methods.

[0066] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a small target detection method based on refined shallow features provided by the above methods.

[0067] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for small target detection based on refined shallow features, characterized in that, include: S1: Acquire drone aerial images, extract feature maps of different scales from the drone aerial images through a backbone feature extraction network, and obtain multi-scale feature maps, including P2 feature map, P3 feature map, P4 feature map and P5 feature map. S2: Adaptive optimization of the P2 feature map is performed through the spatial-channel feature refinement module to obtain the refined P2 feature map; S3: Construct the neck network of the bidirectional feature aggregation path, which includes a top-down semantic injection path and a bottom-up detail enhancement path. S4: Upsample the P5 feature map through a top-down semantic injection path and concatenate it with the P4 feature map to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. S5: The P2 fusion feature map is convolved through a bottom-up detail enhancement path and then concatenated with the P3 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and then concatenated with the P5 fusion feature map processed by the A2C2f module. This concatenation is then input into the A2C2f module to obtain the P5 detection feature map. S6: Target detection is performed based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain the target detection result.

2. The small target detection method based on refined shallow features according to claim 1, characterized in that, The backbone feature extraction network includes a P2 feature layer with a downsampling step size of 4, a P3 feature layer with a downsampling step size of 8, a P4 feature layer with a downsampling step size of 16, and a P5 feature layer with a downsampling step size of 32.

3. The small target detection method based on refined shallow features according to claim 1, characterized in that, The spatial-channel feature refinement module takes shallow feature maps as input and enhances the effective information in the shallow feature maps and suppresses irrelevant background noise through parallel channel feature modeling and spatial feature modeling mechanisms.

4. The small target detection method based on refined shallow features according to claim 3, characterized in that, In the process of channel feature modeling, the input features are subjected to global average pooling in the spatial dimension to extract channel-level statistical information, and channel attention weights are generated through a bottleneck mapping structure consisting of two one-dimensional convolutions.

5. The small target detection method based on refined shallow features according to claim 4, characterized in that, In the process of spatial feature modeling, by performing max pooling and average pooling operations on the channel dimension respectively, spatial response information from different statistical perspectives is obtained, and the information is then concatenated and input into the convolution operation with receptive field to generate a spatial attention weight map.

6. The small target detection method based on refined shallow features according to claim 5, characterized in that, Spatial channel enhancement features are obtained based on channel attention weights, and spatial enhancement features are obtained based on spatial attention weights. Spatial channel enhancement features, spatial enhancement features and P2 feature maps are superimposed through parallel residual fusion to generate refined P2 features.

7. The small target detection method based on refined shallow features according to claim 1, characterized in that, In step S6, large-scale targets are detected using the P5 branch detection head based on the P5 detection feature map. Medium-scale targets are detected using the P4 branch detection head and based on the P4 detection feature map. The P3 branch detection head is used to detect common small targets based on the P3 detection feature map. By refining the P2 branch detection head, extremely small targets from the perspective of the UAV are detected based on the P2 fused feature map.

8. A small target detection system based on refined shallow features, for performing a small target detection method based on refined shallow features as described in any one of claims 1 to 7, comprising: The feature extraction module acquires aerial images taken by the UAV and extracts feature maps of different scales from the UAV aerial images through a backbone feature extraction network to obtain multi-scale feature maps, which include P2 feature map, P3 feature map, P4 feature map and P5 feature map. The spatial-channel feature refinement module adaptively optimizes the P2 feature map to obtain a refined P2 feature map. The construction module constructs a neck network for a bidirectional feature aggregation path, which includes a top-down semantic injection path and a bottom-up detail enhancement path. The semantic injection module upsamples the P5 feature map and concatenates it with the P4 feature map through a top-down semantic injection path to obtain the P4 fused feature map. The P4 fused feature map is then input into the A2C2f module, upsampled, and concatenated with the P3 feature map to obtain the P3 fused feature map. The P3 fused feature map is then input into the A2C2f module and fused with the refined P2 feature map to obtain the P2 fused feature map. The detail enhancement module performs convolution on the P2 fusion feature map through a bottom-up detail enhancement path, concatenates it with the P3 fusion feature map processed by the A2C2f module, and inputs it into the A2C2f module to obtain the P3 detection feature map. The P3 detection feature map is then convolved with the P4 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P4 detection feature map. Finally, the P4 detection feature map is convolved with the P5 fusion feature map processed by the A2C2f module and input into the A2C2f module to obtain the P5 detection feature map. The target detection module performs target detection based on the P2 fusion feature map, P3 detection feature map, P4 detection feature map and P5 detection feature map to obtain target detection results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the small target detection method based on refined shallow features as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the small target detection method based on refined shallow features as described in any one of claims 1 to 7.

Images