Semantic Segmentation And Instance Re-Detection For Weakly Supervised Object Detection In Remote Sensing Images

Object detection is one of the main tasks in remote sensing image processing field,which aims to identify and locate the objects in remote sensing images(RSIs).The mainstream of current research are deep learning-based object detection methods.Among them,fully supervised object detection methods have higher detection accuracy,but the required instance annotations require larger time and labor cost.While weakly supervised object detection methods only require image-level category annotations to train the model.It saves lots of annotation cost and has attracted much attention from researchers in recent years.But there still exists some problems:(1)insufficient reliability and difficulty in detecting complete objects due to pure reliance on class confidence score(CCS)to mine pseudo ground truth(PGT)instances(positive samples in pseudo labels);(2)insufficient locating accuracy due to exclusive reliance on object candidate proposals for object locating.To solve the problem(1),a semantic segmentation guided pseudo label mining(SGPLM)module is proposed in this paper.Firstly,the class-specific segmentation map(CSM)is obtained by the weakly supervised semantic segmentation algorithm.Secondly,the class specific object overlap score(COOS)of each object candidate proposal is obtained by calculating the overlap degree between it and the CSM,and the class specific object confidence score(COCS)of it is obtained by combining the COOS and CCS.Finally,the PGT instance is mined to supervise the training of instance classifier refinement branches based on the COCS.To solve the problem(2),an instance re-detection(IR)module is proposed in this paper.Firstly,an enhanced PGT instances generating strategy is proposed to generate enhanced PGT instances that can more accurately cover the entire object based on the object candidate proposals.Subsequently,the training of the re-classification branch and the re-localization branch in the IR module is supervised based on the enhanced PGT instances.An experimental validation of the proposed method is provided in this paper on two authoritative remote sensing datasets,NWPU VHR-10.v2 and DIOR.The ablation experiments demonstrate the effectiveness of the SGPLM module and the IR module.Quantitative comparison experiments with popular methods demonstrate the superiority of the method in this paper.The subjective evaluation intuitively demonstrates that the method in this paper is able to detect objects more accurately and completely.