Incorporating The Completeness And Difficulty Of Object Candidates Into Weakly Supervised Object Detection In Remote Sensing Images

Object detection in remote sensing images is one of the key tasks in remote sensing image interpretation,which aims to identify and locate the objects with high-value in remote sensing images.The fully supervised object detection method relies on fully labeled samples,which requires high labor and time-consuming.The weakly supervised object detection method based on deep learning has attracted more and more attention because it only needs image-level class labeling,but there are still two problems to be solved:(1)the detection results tend to locate the most significant part of the object rather than the whole object,especially for remote sensing images with complex background;(2)the number of hard and easy samples is imbalanced in the training process.The cumulative loss of a large number of easy samples will dominate the total loss,which limits the detection ability of object detection model to complex background objects.To solve the problem(1),this paper proposes a new quantitative evaluation metric of object candidates: object completeness prior score(OCPS),which is used to quantify the object candidates covered completeness of the object.In this paper,OCPS is fused with the traditional class confidence score(CCS)to obtain the objectness score(OS).The object candidate that covering the whole object and with high CCS will have a higher OS.To solve the problem(2),this paper proposes another quantitative evaluation metric of object candidates: detection difficulty evaluation score(DDES),which be used to quantify the detection difficulty of object candidate.Specifically,object candidate is not a hard negative sample,this paper proposes DDES calculation method based on entropy,which quantitatively evaluates the detection difficulty through the overall confidence score distribution of the object candidate in each category.For the hard negative sample,the DDES method based on entropy is no longer applicable.By this time,the DDES is calculated according to the background confidence score of object candidate.Then,this paper proposes a training loss function integrating OS and DDES into the weakly supervised object detection model.Optimizing the OS,the model will focus on high-quality object candidate.Optimizing the DDES,the model will focus on difficult object candidate.Ablation experiments show that the performance of the benchmark model can be improved by integrating OS and DDES,respectively,and combining the OS and DDES can achieve the best performance.The experimental comparison with 7 popular methods on two public datasets shows that the comprehensive performance of this method is the best,which further verifies the effectiveness of this method.