Research On Vehicle Detection Algorithm In High-resolution Remote Sensing Images Based On Deep Learning

In recent years,vehicle detection in remote sensing images has become an active research field based on the development of high-resolution remote sensing technologies.The vehicle extraction in remote sensing images plays an essential role in traffic information collection and military objective detection.The existing vehicle detection methods mainly rely on feature extraction operators,and the performance of the detection algorithm depends on the expression ability of the operator.Due to the limitation of operator,there are many challenges in vehicle detection for the reasons of background interference and object occlusion and the targets are far from the field of view for vehicle detection in remote sensing.According to the development of deep Convolutional Neural Networks(CNN),researchers propose to detect objects with deep learning to extract features that are most useful automatically from images rather than select features manually in traditional ways.Based on key technologies such as object detection by CNN,superpixel segmentation,NGD optimization etc.,this paper presents an automatic recognition algorithm for vehicle targets in remote sensing images with high detection accuracy.The main work of this paper includes the following aspects.In order to solve the background interference,this paper proposes a region merging algorithm based on superpixel segmentation to extract the building mask.The method firstly segments remote sensing images and corresponding elevation images by superpixel algorithm to obtain two sets of superpixels.Then a building mask based on a set of node clusters of a double segmented region set is generated.The building masks effectively eliminate redundant information in images.However,the targets in remote sensing images are far from the field of view and the detection of small targets is still difficult.Aiming at solving this problem,this paper proposes a new network structure inspired by NGD optimization.Experiments in COWC dataset show that the new network model proposed in this paper performs better in vehicle detection tasks compared with other methods.For the problems of obscured car detection,this paper proposes a multibranch convolutional neural network method by combining the information of optical images and elevation data.Then the conditional random field is used to refine the coarse edges obtained from the output of the multibranch convolutional neural network.Extensive experiments show that the performance on ISPRS datasets is improved to overcome the shortcomings in optical data.