Vehicle Target Detection In Very High Resolution SAR Images

Synthetic aperture radar(SAR)is an all weather,active and coherent imaging high resolution radar.The SAR images are widely used in marine research,geological exploration,and military.The application of SAR images can be mainly divided into four categories,namely classification,target recognition,segmentation and target detection.Among them,SAR image target detection is a very important type of application.It plays an enormous role in target strike task.After years of development,many SAR image detection methods have been created.However,with the improvement of SAR technology,the traditional SAR image target detection methods have become inapplicable.This is mainly caused by three reasons.First,the target in a conventional SAR image only is a few pixels with a large gray value,while the target in a very high resolution SAR image is a region with a distributed feature.Second,due to the imaging angle of the SAR system,the target in the very high resolution SAR image is usually incomplete.The last reason is that with the progress of SAR technology,the characteristic dimension of SAR images is getting higher and higher,and the amount of data is getting larger and larger.The computational complexity of extracting features from traditional methods is very large.In response to the above-mentioned problems,this article proposes corresponding solutions.First,this paper proposes a classification method for target detection of high resolution SAR images.In this method,the non-target areas such as buildings,trees,grasslands,and highways are unified into one category.The target is regarded as another category.So the target detection task is regarded as a two categories problem,and the detection of very high resolution SAR image targets is completed.Second,this paper proposes a method to compensate targets in very high resolution SAR images and construct a compensated data set.This article compensates half of the target's shadow area to the target so that the detected target is more complete.Third,this paper proposes the variable structure convolutional neural network for target detection of very high resolution SAR images.The input layer of the network inputs three images,namely the gradient magnitude image of the original image,the original image and the filtered image of the original image.The second to seventh layers of the network are three combinations of single convolution layers and a single pooled layer.The eighth and ninth layers of the network are two convolution layers.In addition,in each convolution layer of the network,the feature image undergoes a batch normalization process after performing a convolution operation.The three input images increase the edge information of the image for the network and weaken the noise effect,making the model more learned data.The network replaces the multilayer perception with multiple convolution layers,reducing the parameters that the network needs to learn,and reducing the possibility of over-fitting the network.Through experiments,the method proposed in this paper has achieved good results.