An On-Board Real-Time Target Detection and Recognition Method of High-Resolution Remote Sensing Satellite Images

Object detection algorithms for satellite remote sensing images have made some progress over the years, and this task has also had an important impact in the military and civilian fields. The existing research on remote sensing image target detection tasks mainly focuses on the optimization and improvement of the algorithm on GPU. In practical application scenarios, edge devices such as satellites and drones cannot perform target detection and recognition tasks on the collected remote sensing image data in real-time. The remote sensing image detection and recognition algorithm for edge-end deployment can better meet the timeliness requirements of remote sensing image detection and recognition tasks. Therefore, we propose an intelligent detection and recognition algorithm for satellite remote-sensing images deployed on a brain-like chip. Firstly, we optimize and improve the existing advanced object detection algorithm for the edge-end deployment task. We introduce the latest lightweight convolution GSConv into the detection network and refer to the design method of slim-neck to reconfigure the head of the detection network. Then we introduce a lightweight triple attention mechanism into the backbone of the detection network to improve the detection accuracy, which further improves the detection accuracy while maintaining lightweight. After that, we deploy the algorithm on Lynxi KA200 brain-like chip for further inference testing. We also compare and analyze several mainstream object detection algorithms and their variants, and deploy them on the brain-like chip. We conduct experiments on three datasets, and the results show that the algorithm we use outperforms other methods in both detection accuracy and inference speed on the brain-like chip.