Improving Small and Cluttered Object Detection by Incorporating Instance Level Denoising Into Single-Shot Alignment Network for Remote Sensing Imagery

Abstract

Object detection on satellite and aerial images has gained the attraction of the scientific community of computer vision due to its immense value, high difficulty, and the development of large-scale datasets enough to train deep learning models. The progress is tremendous with the increase in the precision of the fast single-stage object detection models which used to sacrifice precision for speed. Aerial and satellite images are of large sizes which makes slow models infeasible for production. Single-shot Alignment Network (S2A-Net) is a fast and competitive single-stage model in terms of precision. However, there is a potential to increase its precision in detecting small and cluttered objects in a complex background. In this paper, a new hybrid approach by incorporating Instance Level Denoising (ILD) module from Small, Cluttered, and Rotated Object Detector ++ (SCRDet++) into S2A-Net is proposed. The model was trained and tested on Dota V1.0. The proposed model achieves a higher mean average precision (mAP) than S2A-Net to be 79.73% and when it was trained using the Kullback-Leibler divergence as a regression loss function, the proposed model can reach as high as 80.39% mAP.