A Space Non-Cooperative Object Active Recognition Method Based on Zoom Area Extract

Space non-cooperative object detection is a key technology for on-orbit mission. At present, the space non-cooperative object detection method based on machine learning improves the detection accuracy of the target from different perspectives. However, non-cooperative object lacks prior information, and space tasks have a certain urgency, which underscores the need for efficient implementation of detection methods. Active filtering of image information is a challenge to improve the performance of space non-cooperative object detection. To address the challenges in space applications, we designed a space non-cooperative object detection method based on active zoom area extraction. This method consists of two modules: an active zoom area extraction module, which combines reinforcement learning and feature weight heatmaps, and an object detection module utilizing ResNet. The active zoom area extraction module down-samples the original image and employs the strategy network to actively select the zoom area with the highest probability of containing the object. It then optimizes this zoom area using feature weighting to obtain the final zoom area. The object detection module detects the zoom area and the downsampled image. According to the experimental results of two public datasets, this method can actively extract the zoom area without reducing the detection accuracy as much as possible, effectively reduce the processed image information, and improve the computational efficiency.