Research on Satellite Image Object Detection Based on Neural Network Algorithm

Currently, in the field of object detection and recognition, data-driven deep learning technologies have achieved groundbreaking progress, significantly outperforming traditional methods in detecting and recognizing near-field images. To address the deficiencies of the Faster R-CNN algorithm, which has a low detection rate for small and weak targets and a high false alarm rate, this paper proposes a SE (Squeeze and Excitation) Faster R-CNN. By integrating feature maps of different scales, it generates a multi-scale joint feature map, thereby enhancing the detection capability for small and weak targets. By constructing a Squeeze and Excitation module, we learn a serialized weight vector for each region of interest (RoI) corresponding to the extracted feature submap during the discrimination process. This vector is used to recalibrate the weights of each channel in the feature submap, suppressing redundant dimensions in the multi-scale joint feature submap and consequently enhancing the algorithm's performance. Furthermore, by incorporating a sorting operation, we propose a SER (Squeeze and Excitation Rank) Faster R-CNN, which completely suppresses redundant feature dimensions in the multi-scale joint feature submap, further improving the overall detection performance. Experimental results based on the Sentinel-1 satellite dataset demonstrate that the improvements made to the algorithm can effectively enhance the performance of ship target detection in SAR images.