Remote sensing image detection based on feature enhancement SSD

For lightweight networks, SSD identifies dam objects in high-resolution remote sensing images, and the complex background information redundancy and similar negative sample interfere with the serious problem; this paper proposes the SSD target detection algorithm based on feature enhancement. Firstly, the appropriate pre-selected box is obtained through k-means clustering, and the RepVGG network is introduced as the backbone network. Then, the ASPP perceptual field enhancement mechanism is added to the top of the network. Finally, the classification and regression part of the network combines the idea of FPN cross-layer connection and SE attention mechanism to obtain the final detection result. To verify the validity of the improved method, the homemade dam remote sensing data sets will improve the methods, such as SSD, FSSD, and Faster R-CNN target detection methods are compared. The experimental results show that the method's precision, recall, and mAP reached 0.8690, 0.7074, and 0.7815, respectively, and the mAP is improved by 11% compared to SSD. In summary, the improved method can extract the target from the complex background of high-resolution images more accurately.