Texture Features Description For High Spatial Resolution Remote Sensing Images Based On Sparse Represetation

In recent years, high spatial resolution satellite remote sensing technology has been developed rapidly. Compared with the low spatial resolution remote sensing images, high resolution remote sensing images has richer structure, shape and texture information, while the traditional methods of feature extraction for remote sensing texture images have not solved the problem of texture feature extraction for high spatial resolution remote sensing images. In addition, the same kind of texture images often occur at different angles of rotation in an image, which always causes great difficulties of interpretation and analysis for high spatial resolution remote sensing images. Therefore, scholars have to study a new method of texture feature extraction for high spatial resolution remote sensing images.Sparse representation has become a research focus in the field of image processing and pattern recognition, and has been widely used. Biological vision studies have shown that sparse coding in line with the encoding mammalian of visual system, and it can be characterized as being spatially localized, oriented and bandpass. For the above issues, this paper constructs a framework of texture feature extraction based sparse representation, and in this framework the method of rotation robust feature extraction for high spatial resolution remote sensing images is proposed. According to biological characteristics, this paper assumes positions of sparse coefficients'nonzero elements represent the location of the neurons, and the overcomplete dictionary learnt is not a rotation robust set of neurons. Later, the sparse codes of texture images will be learned via the extend dictionary under the constraint that the number of nonzero element must be 1. After that, according to the locations of nonzero elements, the histogram of an image will be created, pooled, and then normalized to achieve the feature of this image. Then sparse features of texture images will be robust for rotation transformation.In this paper, classification experiments in a large number of high spatial resolution remote sensing texture image, measured by classification accuracy and verify that the proposed texture description method based on sparse representation has a strong robustness for the rotation transformation of texture images, and the classification in high spatial resolution remote sensing texture images achieved satisfactory results.