Research On Typical Target Detection In Remote Sensing Images Based On Online Hard Example Mining Network

Remote sensing image target detection has important research value in various fields such as military target detection,urban construction planning,and outdoor facility monitoring of enterprises.With the development of remote sensing technology,the spatial resolution of remote sensing images has gradually improved,and the details of the available remote sensing images have become more and more abundant,making the typical target recognition in remote sensing images highly feasible.In this paper,the existing deep learning target detection algorithm in remote sensing image target detection has made some research and improvement.In view of the low accuracy of the general target detection algorithm,the large number of negative samples in the detected target cannot be separated online,and it is necessary to manually add negative samples and iterate repeatedly.Two difficult case mining methods are proposed and a data set for typical target detection in remote sensing images is established.The detection accuracy of the existing target detection algorithm is improved,so that the existing deep learning methods can be better applied to remote sensing images.This article mainly completes the following work.1.Propose a double-threshold detection algorithm.In the algorithm research of double threshold network,we make a second decision for the detection results of the basic network.In the original single-stage target detection network,a branch network is re-extracted.For targets whose confidence falls within the threshold interval,feature extraction is performed in the branch network,and the confidence is recalculated after the pool of interest is pooled to achieve positive and negative samples.Separation,thereby improving the accuracy of remote sensing image target detection.2.Propose a point-constrained model algorithm.The feature information of key points is used to constrain the detection results from the initial network.The two pieces of detection information are fused to make difficult case mining.In the backbone network of the initial detection network,some features are extracted.This part of the feature is fed into a network,and finally the confidence of a key point is predicted,and the confidence of the key point and the position and category information of the target frame in the backbone network are synthesized to obtain more accurate difficult case results.3.In order to verify the effect of the algorithm,this paper also proposes a data set of transmission towers in remote sensing images.The data set uses the electric towers in remote sensing images as typical targets.In this task,a large number of transmission tower-like targets are prone to appear in the remote sensing images.,Such as signal towers,roofs,special triangular landforms,etc.,so the use of this data set evaluation algorithm is more convincing.Experiments show that,on this data set,the double-threshold double-threshold network has a 7% improvement in accuracy compared to the corresponding basic network.The detection network based on point constraints is difficult to mine in remote sensing images.Compared with the single-stage target detection algorithm,the accuracy rate is improved by 6%.Through experimental analysis,it is proved that the difficult case mining method proposed in this paper is effective.