SAR Image Classification Based On Deep Feature Learning And Sparse Representation

Because of the special performance of SAR image,Now SAR image has become an important source of information in the field of remote sensing application such as target detection,urban planning and so on. In the process of the application of SAR images for various remote sensing,the key problems are mainly concentrated on the automatic and reliable image classification.Because the imaging mechanism between the SAR images and nature images is completely different,so parameters and the noise model of the SAR images are totally different.These difference brought great difficulty to SAR image understanding and interpretation.This paper put forward the SAR image classification technology based on deep feature learning and sparse representation,in the two sides of appropriate feature extraction method and effective classification decision criterion.The three mainly innovation points are as follows:(1)The convolution feature learning method for SAR image classification was proposed. In the application of bag of visual words for single SAR image classification,it must be pre segmented. The pre segmentation method directly affect the performance of the classification,and the construction of visual dictionary is easily affected by initialization.The convolution between low-level features and the visual dictionary avoid the initial pre segmentation method which reduces the noise, enhances the original data structure and the classification performance.(2)The SAR image classification technology based on the deep feature learning and watershed was proposed. As the method of appropriate feature extraction for the classification of SAR image requires a lot of professional skills and practical experience,here introduce the sparse encoder model in deep leaning.the application of self-learning ability for the features extract the complex feature which method is very difficult for human to design,liberate human and resources,mine more advanced feature,make the process of feature extraction more intelligent and automatic and improve the result of classification.(3)The improved K_SVD sparse representation method for SAR image classification was proposed. Because the theoretical basis of the traditional sparse representation method is the additive gaussian white noise of natural image.Therefore here update the K_SVD objective optimization function based on the local statistical distribution of SAR images.And the improved sparse representation method is used to reconstruct the pixels and choosing the minimum reconstruction error form the different dictionaries as category labels of pixels.Furthermore the fusion of majority voting decision criterion using the preliminary segmentation of watershed algorithm decreases the influence of noise,improves the classification accuracy.