Research on Rotated Object Detection Based on YOLOv5s

To address the issues of low detection accuracy and high miss rates in remote sensing object detection, this study proposes an improved YOLOv5s_CSL method integrated with attention mechanisms. Firstly, by employing phase-shift encoding for angle precision, this method uses cosine functions of varying phases and frequencies to replace traditional CSL discontinuous encoding, thus enabling automatic derivation and continuous optimization of angles, and effectively resolving the class square problem inherent in CSL encoding. Secondly, an innovative dynamic receptive field feature fusion attention mechanism is introduced, which enhances the model's response to targets of varying scales by automatically adjusting the receptive field, while also facilitating effective cross-space information transfer, significantly improving detection capabilities in complex environments. Lastly, a new angle regression loss function was designed, combining mean squared error loss with the geometric characteristics of the bounding boxes, significantly enhancing the flexibility of angle detection for objects of various shapes and sizes, and effectively reducing the miss rate for objects with large aspect ratios. Experimental results demonstrate that this algorithm performed exceptionally well on the DOTA-v1.0 test set, achieving a detection accuracy of 77.47%, a significant improvement over the YOLOv5s_CSL method. These innovations not only enhance the accuracy of remote sensing object detection but also provide new effective tools for image recognition in complex scenarios, offering high application value and prospects for widespread use.