Detection Method Based on Image Enhancement and an Improved Faster R-CNN for Failed Satellite Components

Abstract

The components detection of a failed satellite is an important work in space on-orbit service. However, the current detection methods for failed satellite components do not consider the effects of low illumination and small targets on the detection accuracy of components at the same time, and most of the datasets used are manually designed. This article proposes a detection method based on image enhancement and an improved faster region-based convolutional neural network (R-CNN) for small components of a failed satellite in low illumination. First, the dataset of failed satellite components dataset containing low-illumination scenarios is established in a simulated real space environment. Second, an image enhancement based on reflection model and principal component analysis is proposed, which further enhances images while conserving richer details. Finally, an improved faster R-CNN for small components is proposed. To improve the detection accuracy of small components, the original faster R-CNN is improved by modifying three modules: backbone, region proposal network, and region of interest (RoI) pooling layer, respectively [i.e., modified high-resolution neural network (M-HRNet), intersection over union (IoU)-balanced sampling, and RoI align]. Experimental results show that the proposed method can accurately detect all the small components on satellite capture plane, and the detection performance for low illumination and small components is improved significantly compared to the state-of-the-art methods and original faster R-CNN.