| Abstract |
Object detection technology for images generated by optical sensors is of great significance in areas such as national defense security, disaster prediction, and smart city construction. Aiming at the problems of small target size, arbitrary target direction, and complex background in aerial image object detection using optical sensors. By using the images captured by a wild range of optical sensors such as CCD（Charge Coupled Device） and SAR(Synthetic Aperture Radar), we propose a more efficient rotating frame object detection algorithm that introduces the MCAB (Multi-branch Convolutional Attention Block) that can process these same types of images taken by the aerial platform using light sensors. First, we construct the module by adding identity, residual branch, and CBAM (Convolutional Block Attention Module) structure to the traditional convolution layer and the activation function to substitute the simple conv module in the network. Secondly, we introduce the Transformer layer at the end of the backbone network to enhance the global perception of the model with low overhead, realize the relationship is modeling between the target and the scene content, and reduce the amount of calculation. We improve the structure of the neck layer by BiFPN (Bidirectional Feature Pyramid Network) to speed up network operation, and CBAM is added after C3 to highlight important characteristic information. Our improved YOLOv5 algorithm is tested on the self-made rotating small target aerial data set. Compared with the original algorithm, our model's mean Average Precision (mAP) is improved by 1.8 percentage points, and the accuracy and recall are also enhanced by 0.46 and 0.34 percentage points. © 2022 Elsevier B.V. |
