Research On SAR Image Object Recognition Based On Deep Learning

With the development of radar detection technology,more and more SAR image object detection and recognition methods are known by people.Traditional detection algorithms require manual design of features,which increases the amount of computation for researchers,while the emergence of deep learning technology provides new ideas for object detection algorithms.In order to solve the problem that small objects in SAR images are difficult to be detected and SAR image object detection in complex backgrounds,this paper mainly studies the SAR image object detection and recognition method based on YOLO v3 network.On the one hand,due to its powerful self-learning ability,the convolutional neural network greatly reduces the complexity of manual feature extraction.YOLO v3 network has an integrated mechanism of detection and identification,which directly realizes end-to-end effects,reduces the complexity of the identification process,and greatly reduces the time cost.On the other hand,the proposed improved method has a good effect on the detection of small targets,and it can still maintain good performance when applied to complex background images.This article mainly carried out the following research work:1.This paper presents an improved YOLO V3 object recognition algorithm.The algorithm is mainly composed of prior boxes generation,feature extraction,multi-scale prediction,and loss function optimization.Through improvements to different modules of the network,a method suitable for SAR image object recognition is obtained,which improves the ability of the network to detect small targets in SAR images.2.This article involved a visual attention mechanism for SAR images with complex background.Based on the Gaussian pyramid model,the scale information of the SAR image target is used to select the scale to improve the saliency map generation algorithm,and it can be used for SAR image detection and recognition.The saliency map image is jointly trained with the original image,so that the background is weakened and the target object is more prominent,so that the target object is easier to detect..3.This article builds an image dataset with a complex background based on the MSTAR dataset,embeds ten different military vehicles into the background image without target information,and annotates to generate an xml file,saving energy for other future research and time.4.This paper constructs an image dataset with complex background based on the MSTAR dataset.We embed ten different types of military vehicles into the background image without target information,and mark them to generate xml files,which saves energy and time for future research.This paper completed the detection and identification experiments on the MSTAR data set and the self-built data set,and achieved good results.Comprehensive experimental results and analysis,the network proposed in this paper has high accuracy and strong feasibility.