Research On Incomplete Supervision Learning Remote Sensing Object Detection

In recent years,more and more high resolution remote sensing images are available with the development of remote sensing technology and deep learning object detection can be applied into the remote sensing field.However,object detectors require huge amount of labeled samples for training to achieve high performance,while remote sensing images have large size and numerous objects in each and corresponding annotation cost limits both scale of training dataset and object detectors’ performance.In fact,there are numerous unlabeled remote sensing images available,this project focuses incomplete supervision learning on object detection to improve detection performance and save annotation cost with designing active learning algorithms to wisely select samples and using semi-supervised learning to annotate some objects automatically.Firstly,this project researches how to define objects’ classification uncertainty and images’ uncertainty accumulation equation in the active learning algorithm to achieve higher performance on object detection and relative saving of labeled images.We bring out relative saving of labeled objects as a new evaluation index which reflects incomplete supervision learning algorithms’ performance more accurately and employ this index to evaluate other algorithms’ capacity in this paper.Secondly,this project analyzes object detectors’ regression task and designs an object regression uncertainty which estimates objects’ contribution to detector with their sizes.Therefore,we design a scale-adaptive active learning algorithm through combining classification uncertainty and regression uncertainty.Besides,we use active learning algorithms to evaluate labeled samples and assign more weights to hard samples to improve object detector’s performance.Finally,this project researches active semi-supervised learning object detection.The active learning designed before can wisely select unlabeled samples but these samples still own high confident objects because remote sensing images have huge size and may contain many different categories’ objects.This project annotates high confident objects automatically which can save annotation cost and improve object detector’ performance.Besides,this project uses image self-learning strategy to annotate high confident images automatically which clears object detector’s inference boundary with no human labor cost.The best incomplete supervision learning designed by this project allows object detector to achieve almost fully supervised performance with 58.2% relative saving of labeled objects.