Research On Automatic Detection Of Aircraft Targets In High Resolution Remote Sensing Images

Deep learning is the hotspot and focus of current research,and classical methods need artificial design features.Unlike classical methods,deep learning automatically extracts features in images,thus the extracted features are more representative and robust than those extracted from traditional features.There are many excellent networks for natural images and they have achieved superior results in classification,target detection,and semantic segmentation.Some mature methods in deep learning are gradually applied to remote sensing image processing,and many new methods for remote sensing images have been proposed especially in the fields of land use classification,change detection,target detection,etc.However,the research on remote sensing image semantic segmentation is relatively less than other fields,mainly due to the difficulty of constructing semantic segment dataset,it takes a long time to train the segmentation network,and the segmentation effect is difficult to improve.Based on the existing deep learning network,this paper proposes some useful improvements for the detection and segmentation of airplane targets in remote sensing images.There are two steps to identify airplanes,the first step is airplane detection and the second step is airplane segmentation.During this process,we have proposed some specific improvement methods.In the target detection step,considering that there are some tiny targets included in the aircraft target,and the distance between these planes are very small,causing the higher level of difficulty to locate.This paper adds FPN to Faster R-CNN,so that multi-scale information is combined to improve detection accuracy.In semantic segmentation,firstly,in order to combine more high-resolution information,we reduce the shrink rate of feature map in net.Secondly,due to the memory limitation,we cannot use deeper network ResNet-101 as backbone when we change the shrink rate of net as 1,so we use the deconvolution layer to restore the feature map to the input image size,and using skip link to combine high-resolution information.Finally,when using the cross entropy function as a loss function,the error caused by the simple samples will cover the error caused by the more complicated samples,which will result in poor segmentation result of the more complicated samples.In order to solve this problem,we replace the cross entropy loss as the focal loss,so that a small weight is assigned to the simple samples.In addition,this paper also built the airplane detection dataset and semantic segmentation dataset.Based on these datasets,we did some experiments,the experimental results show that the improved method proposed in this paper can significantly improve the recognition accuracy of airplane targets.