Image rotation small target detection network based on receptive field convolution and content perception feature module

Abstract

The invention relates to the field of computer vision and target detection, and particularly provides an improved method combining receptive field convolution and content perception features for rotating small target detection based on a convolutional neural network. According to the method, aiming at a traditional YOLOv11OBB network, an improved C3K2 module (C3K2-RE) and a content awareness feature recombination up-sampling module (CARAFE) are introduced, so that the small target detection precision and robustness are improved. Specifically, a convolutional layer in the YOLOv11OBB is replaced by a C3K2-RE module. The module fuses receptive field attention convolution (RFAConv) and efficient multi-scale attention (EMA), strengthens the feature extraction capability, and optimizes the detection effect of a rotating small target in a remote sensing image. Meanwhile, an up-sampling module of a network check part is replaced by a CARAFE module. As a lightweight efficient up-sampling scheme, the CARAFE generates the weight of each position through content awareness, the features are weighted and recombined by a predefined region, and up-sampling is completed through rearrangement of multiple groups of weights, so that details are effectively reserved, the reconstruction quality is improved, and the output is clearer and more accurate.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and target detection, particularly the field of rotating small target detection based on convolutional neural networks. Specifically, it improves the YOLOv11 OBB model by introducing an improved C3K2 module (C3K2-RE) and a content-aware feature reconstruction module (CARAFE) to achieve high-precision detection of rotating small targets in remote sensing images. Background Technology

[0002] YOLOv11 OBB (Oriented Bounding Boxes) is the latest evolution of the YOLO series of object detection models, specifically designed for rotating object detection tasks, achieving significant breakthroughs in accuracy, speed, and multi-scene adaptability. This model inherits the high efficiency of single-pass forward inference from the YOLO series and further optimizes the regression capability of rotated bounding boxes (OBBs) by introducing Deformable Convolutional Network version 4 (DCNv4) and a multi-scale feature enhancement module, making it more robust in detecting objects with complex geometric shapes.

[0003] YOLOv11 OBB replaces the traditional C3 module with an improved C3k2 mechanism, and combines it with the C2PSA (Pyramid Spatial Attention) module to enhance feature extraction capabilities. It also optimizes bounding box regression accuracy through the EIoU loss function, achieving 81.3% mAP@50 on the DOTAv1 dataset, a 2.1 percentage point improvement over YOLOv8 OBB. Its network architecture innovations include: embedding a Support Vector Machine (SVM) in the backbone to predict tilt angles, adding an angle prediction branch to output the θ parameter, and using depthwise separable convolutions to reduce the number of parameters by 22% while maintaining computational efficiency. Compared to traditional anchor-based detectors, YOLOv11 OBB, through a dynamic anchor allocation strategy and adaptive channel pruning mechanism, reduces the false negative rate for small targets (≤32×32 pixels) by 37% while maintaining real-time performance (18.5ms per frame on a T4 GPU), and improves bounding box regression stability to 91% in occlusion-resistant scenarios.

[0004] This model supports end-to-end training and achieves geometric alignment of rotation angles through an improved IoU-GLoss function. In remote sensing image analysis, it achieves an AP50 of 78.9% for tilted building detection and an RSS index of 0.8956 for blood vessel segmentation in industrial quality inspection scenarios. Although it still faces challenges in complex scenarios such as forest fire monitoring and maritime rescue due to insufficient feature extraction of small rotating targets, YOLOv11 OBB uses knowledge distillation technology to transfer the accuracy of a large model to a lightweight version, achieving compression with an accuracy loss of <1%. It has been integrated into the Ultralytics HUB platform to support one-click fine-tuning, making it the preferred solution in edge computing and real-time monitoring fields. Summary of the Invention

[0005] This invention proposes an improved method for detecting rotating small targets based on YOLOv11 OBB. By introducing an improved C3K2 module C3K2-RE and a content-aware feature reconstruction module CARAFE, high-precision detection of rotating small targets in remote sensing images can be achieved.

[0006] First, the convolutional layers in the YOLOv11 OBB are replaced with the improved C3K2 module, C3K2-RE. The C3K2-RE module significantly improves the module's feature extraction capabilities by fusing receptive field attention convolution (RFAConv) and efficient multi-scale attention (EMA), thereby enhancing the detection of small rotating targets in remote sensing images. Second, in the neck part, the upsampling module has been replaced with the content-aware feature reassembly module CARAFE. This is a general, lightweight, and efficient upsampling module. CARAFE reassembles features in a predefined region centered at each location through weighted combination, where the weights are generated in a content-aware manner. Furthermore, each location has multiple sets of such upsampling weights, and feature upsampling is completed by rearranging the generated features as a spatial block. This approach improves detail preservation and reconstruction quality during the upsampling process, enabling the network to generate clearer and more accurate outputs.

[0007] The specific implementation steps of this invention are as follows:

[0008] Step 1: Standardize the input image (e.g., normalize, resize, etc.) to fit the model's input requirements.

[0009] Step 2: The input image is passed through the initial convolutional layers of the improved YOLOv11 OBB model, which have been replaced with the improved C3K2 module C3K2-RE. The C3K2-RE module significantly improves the feature extraction capability of the module by fusing receptive field attention convolution RFAConv and efficient multi-scale attention EMA, thereby improving the detection capability of rotating small targets in remote sensing images.

[0010] Step 3: After image features are processed by the improved C3K2 module C3K2-RE, they enter the neck section. In the neck section, the upsampling module has been replaced by the content-aware feature reassembly module CARAFE. This is a general-purpose, lightweight, and efficient upsampling module. CARAFE reassembles features in a predefined region centered on each location through weighted combination, where the weights are generated in a content-aware manner. Furthermore, each location has multiple sets of such upsampling weights, and feature upsampling is completed by rearranging the generated features as a spatial block. This method improves detail preservation and reconstruction quality during the upsampling process, enabling the network to generate clearer and more accurate output.

[0011] Step 4: The processed feature map enters the head section. The head section performs object classification and bounding box regression prediction, outputting class confidence and parameters of the rotated bounding box (OBB), including the object's center coordinates, width, height, and rotation angle.

[0012] Step 5: Apply Non-Maximum Suppression (NMS) to the output of the head section to remove redundant detection results, retaining only the detection boxes with the highest confidence, ensuring the accuracy and reliability of the detection results. The final output is the processed target detection result, including the target category and rotated bounding box (OBB), for subsequent applications. Attached Figure Description

[0013] Figure 1 Here is the overall network structure diagram of the algorithm.

[0014] Figure 2 Here is the structure diagram of YOLOv11 OBB.

[0015] Figure 3 Here is the C3K2-RE structure diagram.

[0016] Figure 4 Here is the structure diagram of RFAConv.

[0017] Figure 5 Here is the EMA structure diagram.

[0018] Figure 6 Here is the CARAFE structure diagram. Detailed Implementation

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] This invention designs a remote sensing image rotation small target detection network structure based on an improved YOLOv11 OBB, as follows: Figure 1 As shown, we build upon YOLOv11 OBB (whose structure is as follows) Figure 2As shown), firstly, the convolutional layers in YOLOv11 OBB are replaced with the improved C3K2 module C3K2-RE (structure as shown). Figure 3 As shown). The C3K2-RE module integrates receptive field attention convolutions RFAConv (structure as shown). Figure 4 (as shown) and efficient multiscale attention EMA (structure as shown) Figure 5 As shown in the diagram, this significantly improves the module's feature extraction capability, thereby enhancing the detection capability of small rotating targets in remote sensing images. Finally, in the neck section, the upsampling module has been replaced with the content-aware feature reconstruction module CARAFE (whose structure is shown in the diagram). Figure 6 (As shown). This is a general, lightweight, and efficient upsampling module. CARAFE reassembles features in a predefined region centered on each location by weighted combination, where the weights are generated in a content-aware manner.

[0021] Specific methods

[0022] This invention proposes an improved method for detecting rotating small targets based on YOLOv11 OBB. By introducing the improved C3K2 module C3K2-RE and the content-aware feature reconstruction module CARAFE, high-precision detection of rotating small targets in remote sensing images can be achieved.

[0023] Figure 1 This is the overall network structure diagram of the algorithm. First, the convolutional layers in YOLOv11 OBB are replaced with the improved C3K2 module C3K2-RE. The C3K2-RE module significantly improves the feature extraction capability of the module by fusing receptive field attention convolution RFAConv and efficient multi-scale attention EMA, thereby improving the detection capability of rotating small targets in remote sensing images. Second, after the image features are processed by the improved C3K2 module C3K2-RE, they enter the neck part. In the neck part, the upsampling module has been replaced with the content-aware feature reassembly module CARAFE. This is a general, lightweight, and efficient upsampling module. CARAFE reassembles features in a predefined region centered on each location by weighted combination, where the weights are generated in a content-aware manner. In addition, there are multiple sets of such upsampling weights at each location, and then the feature upsampling is completed by rearranging the generated features as a spatial block. This method improves the detail preservation and reconstruction quality in the upsampling process, enabling the network to generate clearer and more accurate outputs.

[0024] Figure 2This is a diagram of the YOLOv11 OBB architecture. YOLOv11 OBB (Oriented Bounding Boxes) is the latest evolution of the YOLO series of object detection models, specifically designed for rotating object detection tasks, achieving significant breakthroughs in accuracy, speed, and multi-scene adaptability. This model inherits the high efficiency of single-pass forward inference from the YOLO series, and further optimizes the regression capability of rotated bounding boxes (OBBs) by introducing Deformable Convolutional Network version 4 (DCNv4) and a multi-scale feature enhancement module, making it more robust in detecting objects with complex geometric shapes.

[0025] Figure 3 This is a schematic diagram of the improved C3K2-RE module structure. The C3K2-RE module significantly improves the feature extraction capability for small, rotating targets in remote sensing images while maintaining a lightweight design by deeply fusing receptive field attention convolution (RFAConv) with an efficient multi-scale attention (EMA) mechanism. This module innovatively combines the dynamic receptive field spatial feature generation capability of RFAConv (generating non-overlapping receptive field slider features through grouped convolutions) with the cross-spatial multi-scale attention mechanism of EMA (capturing local-global feature interactions through parallel 1x1 and 3x3 sub-networks). A dual-branch feature enhancement path is embedded in the residual structure of the C3 module: the RFAConv branch uses dynamic grouped convolutions instead of standard 3x3 convolutions, generating spatially perceptual features through k×k groups of convolutions, and strengthening the response of key regions with Softmax attention weights. The EMA branch achieves multi-scale context modeling while maintaining parameter efficiency through channel reshaping and cross-dimensional interaction.

[0026] Figure 4 This is a deeply integrated receptive field attention convolutional RFAConv architecture diagram. Its innovation lies in the collaborative design of dynamic grouped convolution and spatial attention mechanisms. The module uses grouped convolution instead of the standard 3×3 convolution, dynamically generating k×k groups of non-overlapping receptive field slider features based on the convolution kernel size k. Each slider feature is differentially weighted using Softmax attention weights, solving the problem of local feature distortion caused by parameter sharing in traditional convolution. In the feature enhancement stage, a cross-dimensional interaction mechanism is introduced, performing global average pooling on the feature maps output by the grouped convolution, and then fusing them with the original features through 1×1 group convolutions. This achieves deep coupling between receptive field spatial features and channel features, and the calculation formula is as follows:

[0027] F = A rf ×F rf (1)

[0028] A rf =Softmax{g 1×1 [AvgPool(X)]}, (2)

[0029] F rf =ReLU{Norm[g k×k (X)]}, (3)

[0030] Compared to traditional CBAM / CA modules, RFAConv dynamically generates non-overlapping receptive field sliders (instead of fixed sliding windows), allowing attention weights to fully adapt to the target geometry. On the DOTAv1 dataset, it achieves a 2.8% improvement in AP50 for small target detection while only increasing computation by 0.7G FLOPs.

[0031] Figure 5 This is the structure diagram of the efficient multi-scale attention mechanism EMA. Its innovation lies in the deep integration of multi-scale parallel sub-networks and cross-spatial learning strategies. This module groups the input feature map along the channel dimension (G groups), with each group containing C / / G channels. Local and global spatial information are encoded through two parallel branches: the left branch uses 1×1 convolution and adaptive average pooling to generate global cross-channel attention weights, and the right branch uses 3×3 convolution to capture multi-scale local features and introduces 2D global average pooling to enhance spatial awareness. The output features of the two branches achieve cross-dimensional interaction through matrix multiplication, generating attention maps with pixel-level pairwise relationships, and finally, they are weighted and fused with the input features through the Sigmoid function.

[0032] Compared to the traditional CA module, EMA avoids information loss caused by channel dimensionality reduction through group processing. On the DOTAv1 dataset, it achieves a 2.3% improvement in AP50 for small object detection while only increasing the computational cost by 0.5G FLOPs.

[0033] Figure 6This is the structure diagram of the content-aware feature reconstruction module CARAFE. Its innovation lies in the collaborative mechanism of dynamic kernel prediction and adaptive local feature reconstruction. This module adopts a two-stage design: the kernel prediction module generates spatially adaptive k_up×k_up reconstruction kernels through a channel compressor (1×1 convolution to compress the number of channels to C_m), a content encoder (k_encoder×k_encoder convolution to generate reconstruction kernels), and a kernel normalizer (Softmax spatial normalization); the content-aware reconstruction module uses the prediction kernel to perform weighted summation on local regions of the input features to achieve feature reconstruction. Compared with traditional upsampling methods (such as bilinear interpolation and deconvolution), CARAFE breaks through the fixed kernel limitation through the content-aware kernel generation mechanism, improving the detection accuracy of Faster R-CNN by 1.2% AP on the MS COCO dataset, while the computational cost is only 1 / 828 of that of deconvolution. Its large receptive field (5×5 when k_encoder=3) effectively captures long-range dependencies, improving the AP50 of tilted building detection by 2.1% in remote sensing image analysis.

Claims

1. A small target detection network for image rotation based on receptive field convolution and content-aware feature modules, characterized in that... Includes the following steps: Step 1: The input image is passed through the initial convolutional layers of the improved YOLOv11 OBB model, which have been replaced with the improved C3K2-RE module. The C3K2-RE module significantly improves the feature extraction capability of the module by fusing receptive field attention convolution RFAConv and efficient multi-scale attention EMA, thereby improving the detection capability of rotating small targets in remote sensing images. Step 2: Image features, after passing through the improved C3K2-RE module, enter the neck section. In the neck section, the upsampling module has been replaced with the content-aware feature reassembly module CARAFE. This is a general-purpose, lightweight, and efficient upsampling module. CARAFE reassembles features in a predefined region centered on each location through weighted combination, where the weights are generated in a content-aware manner. Furthermore, each location has multiple sets of such upsampling weights, and feature upsampling is completed by rearranging the generated features as a spatial block. This approach improves detail preservation and reconstruction quality during the upsampling process, enabling the network to generate sharper and more accurate output. Step 4: The processed feature map enters the head section. The head section performs object classification and bounding box regression prediction, outputting class confidence and parameters of the rotated bounding box (OBB), including the object's center coordinates, width, height, and rotation angle.

2. The method as described in claim 1, characterized in that, The convolutional layer in step 1 uses an improved C3K2-RE module as a solution.

3. The method as described in claim 1, characterized in that, In step 2, the upsampling module in the neck section has been replaced with the content-aware feature reconstructing module CARAFE.

4. The method as described in claim 1, characterized in that, In step 4, the head section outputs the class confidence score and the parameters of the rotated bounding box (OBB), including the center coordinates, width, height, and rotation angle of the target.

Images