Feature Extraction of SAR Image Based on Local Important Sampling Binary Encoding
He Ch · Acta Automatica Sinica · 2014
Synthetic aperture radar(SAR) image presents properties of texture because of speckle and the spatial variability of target response. Methods of feature extraction based on local binary encoding have achieved good results and are popular with texture description in recent studies. In this paper, we make use of the framework of texture feature and propose an approach of feature extraction based on local important sampling binary encoding representation(LISBF)for SAR image classification. Firstly, use some images to sample local key points randomly and adaptively, and output the recursive learning position which is based on the important sampling method. Then, we use the learning position to form binary code. Finally, we develop the feature descriptor by mapping and statistics. This feature an provide a wider range of information than fixed location sampling and avoid the feature dimension increasing sharply by sampling. Compared with random sampling, this feature is easier to capture the texture information via adaptive learning and more fits the very low SNR and speckle of SAR image. This paper applies this feature to the classification of real SAR image and standard texture library. The effectiveness of this feature is tested and verified by the experiment results.