Object Detection In High-Resolution Remote Sensing Images Based On Multi-Level Deep Features Fusion And IGIoU Loss

Object detection in high-resolution remote sensing images is one of the most important tasks in the field of optical remote sensing image processing.It aims to locate and classify the objects with high value in high-resolution remote sensing images.At present,the anchor based two-stage object detection methods have been widely used in the field of object detection in high-resolution remote sensing images due to its high detection accuracy.However,there are still three problems exist in the anchor based object detection in high-resolution remote sensing images:(1)the metric could not measure the distance between the two bounding boxes when the predicted bounding box and the ground truth box are non-overlapping;(2)there is no direct relationship between the existing bounding box regression loss and the metric,i.e.,the existing bounding box regression loss could not optimize the metric directly in the training process;(3)the existing strategy of feature extraction of region proposal based on hierarchical deep network only utilize single-level feature,and fail to make full use of the advantage of the multi-level features.To solve the problem(1),this paper adopts a new metric,i.e.,generalized intersection over union(GIoU),which can comprehensively consider the overlapping degree of the predicted bounding box and ground truth box,measure the proximity degree of two bounding box,and thus effectively alleviate the problem that the existing metric could not measure the distance of the two bounding box without any overlapping.To overcome the problem(2),a novel bounding box regression loss,i.e.,improved generalized intersection over union loss(IGIoU Loss),is proposed in this paper.The bounding box regression loss can not only optimize the metric directly but also change the gradient adaptively in the training process,and thus improve the training efficiency and training effect of the model.To resolve the problem(3),a multi-level deep features fusion(MLFF)method is proposed in this paper,incorporating the MLFF into the hierarchical deep network and can make full use of multi-level feature information,and can obtain more discriminative features,and thus further improve the effect of object detection in high-resolution remote sensing image.In order to validate the effectiveness of the proposed methods,experiments are conducted on the NWPU VHR-10 and DIOR dataset,respectively.Compared with the baseline model,the MLFF method is used to extract the features of region proposal,the IGIoU loss is adopted as the bounding box regression loss,and the combination of MLFF and IGIoU loss have an improvement on mean average precision(mAP)for NWPU VHR-10 dataset,which are respectively 1.0%,2.0% and 2.7%.Similarly,the improvement for mAP are 0.7%,1.4% and 2.2% respectively on DIOR dataset.Moreover,the multi-threshold curves consistently show that the detection results under high threshold are obviously superior to other methods on both public available dataset,which indicates that the proposed methods have an obvious advantage on the location accuracy when the threshold is strict.