SAR Image Objection Detection Based On Adaptive Weight Loss Function And Situation Recognition

With the development of computer software and hardware,machine learning and deep learning have become the current research concern.based on them,great progress has been made in the field of computer vision.SAR images have significant applications in geological exploration,military warfare and other fields,thanks to the advantages that they can be got in any time and any weather.However,available SAR images are too scare,so the accuracy of SAR image object detection is limited,and the situation recognition accuracy based on object detection results is also limited.In this thesis,we mainly propose corresponding improvement methods for the problems of lack of images,low accuracy of object detection and low accuracy of situation recognition of SAR images.The main contributions can be divided into three parts:(1)Aiming at the problem that there are scare available SAR images,we propose a simple and efficient method of data augmentation based on CNN.The available objects and background images are combined automatically,which can double the number of training sets.Experiments have been conducted on object detection tasks.The experimental results show that the data augmentation method can increase the m AP by 7.57%,which significantly improves the accuracy of SAR image object detection.(2)In the task of SAR image object detection,in view of the problem that the detection effect of some categories of samples is poor,an adaptive weight loss function(1/Io U Loss)is proposed,which is applied to the regression loss part of the region proposal network.The loss function can adaptively increase the contribution of the samples with poor detection effect in the loss function by adjusting the factor,so as to enhance the detection effect of those hard samples,thereby improving the overall detection accuracy.The experimental results indicate that 1/Io U Loss method works better especially when the original task is challenging.Among them,the data enhanced data set is used to compare the method before and after the improvement,and the m AP obtained is increased by 4.61%.(3)In the task of SAR image situation recognition,the detection result in the second work is regarded as one dimension in the feature vectors,and is input into the SVM together with other designed features that can reflect the situation information to multi-classify the enemy’s motivation,including attacking,defending or retreating,which can assist experts to predict the battlefield situation.Among them,since the Gaussian kernel of SVM is sensitive to two parameters,a step-by-step grid search algorithm is proposed to adaptively search for the optimal parameter combination.Manual search can hardly reach the height of the efficiency and the classification accuracy of the proposed method.The classification accuracy of SVM can reach 100% in experiments.At the same time,the work of this thesis is integrated into a visual system that can display the situation and hotspots of the battlefield.