Target Detection And Classification Methods Based On Convolutional Neural Network For SAR Image

In recent years,higher requirements of the target detection and recognition algorithm are put forward due to the rapid development of SAR imaging technology.Conventional identification and detection algorithm needs to design appropriate feature extractor for the special data set.However,designing a feature extractor requires the professional knowledge and rich experience,which is difficult to obtain in practice.Moreover,they are often unable to meet the demand of the complex and changeable SAR image datasets.Compared with the traditional detection and classification algorithm,convolution neural network can automatically find the complex structure in the input datasets,and integrate the simple feature extracted from the shallow layers to an abstractive feature which is more suitable for the task.More importantly,it's easy to control their capacity by varying their depth and breadth.This paper studies the SAR image object classification algorithm and SAR image target detection algorithm based on the convolution neural network.The main contents of this paper can be summarized as the following two points:1?Convolution neural network contains a large number of free parameters,leading to the fact that it is prone to over-fitting in the case of less training data.Compared with the optical datasets,it is difficult to obtain large scale of SAR datasets with high quality.Therefore,the SAR image recognition algorithm based on convolution neural network are confronted with two major problems: 1)effectively augment the current datasets to obtain a complete training dataset for training the convolution neural network;2)design a model with acceptable size to prevent over-fitting,while keeping its capacity and reducing the training time.In chapter 3,we argument the current datasets with label-preserving transformations,for examples,target area translation,adding the noise,filtering,rotation,mirror and so on.We also design a model with 4 convolutional layers and 4 max pooling layers,2 fully connected layers.This depth seems to be important: 4 pooling layers can reduce the number of parameters in the convolution neural network,further decreasing the probability of over-fitting.The experimental results on the MSTAR dataset demonstrate that the proposed method can improve the classification performance.2?A large scale SAR image is often complex,including a variety of clutters.the traditional CFAR detection algorithm is based on the pixel level,so the algorithm is sensitive to noise,which results in a large number of false alarm targets.The Convolution neural network not only uses the pixel intensity information in the image,but also the target structure information,which would reduce the number of false alarm targets and strengthen the ability of location.In chapter 4,we propose an strategy which use the augmented samples to assist training CNN model for the detection of the complex scene.On the one hand,we pre-train a CNN model using the samples in simple scene to accelerate the convergence rate of the detection network.On the other hand,the samples from simple scene are scaled and then used to enlarge the training datasets for the detection network.Finally,the region proposal model and detection model are obtained through the augmented training dataset.The experimental results on the Mini SAR dataset demonstrate that the proposed method can improve the detection performance.