The Study Of Local Invariant Feature Description Method Based On Colour Image

In the field of computer vision and pattern recognition or the other related fields, the extraction and quantitative description of image features can simply the high dimension image information in order to convenient for blending into practical application of image processing technology. Image feature description method plays a crucial role for describing nature characteristics of the image, similarity or otherness between images and image semantics. And also, it provides the application values for object recognition, image scene categorization and image retrieval etc. fields.Generally, the good distinctiveness and invariance should be shown in feature description method. Distinctiveness refers to that the method of feature description should easy to distinguish different categories of image. And invariance means that the performance of feature description method should not be affected when image appears geometry, illumination or blurring changes. However, in the study of practical application, these two characters can not be both satisfied in most case. Therefore, a good feature descriptor should improve its invariance under keeping its distinctiveness.SURF(Speeded-Up Robust Features) feature description method describes the attribution of image feature points through counting up gradient information of image pixel points quantitatively. It can keep the invariance of descriptors in a certain extent when image appears scale, rotation, illumination, brightness, or contrast changes. In this paper, the quaternion method and color moment feature are separately combined with SURF feature description algorithm for researching color image feature description method. The following works are mainly completed in this paper:(1)Special to feature description for color image, a novel feature description algorithm for color image based SURF feature is proposed by using quaternion in this paper. Firstly, in order to representing the color information of color image effectively, the color information of each pixel point among three channels is expressed by three imaginary parts of pure quaternion respectively. This can simplify the presentation process of color information because a whole color image is transformed into a pure quaternion matrix. Secondly, the SURF features in three channels of color image are extracted simultaneously through the above pure quaternion matrix, which can improve the computational efficiency of our algorithm. Then, taking advantage of the characteristic of rotation invariance of quaternion norm, the final feature description vectors are described by pure quaternion norm. Finally, the simulation experiments show that the correct matching ratios and time complexity of our method are superior to SURF method, FAIR-SURF(Fully Affine Invariant SURF) algorithm and other feature description algorithms based on color image in recent years, which verifies the effectiveness of our method in this paper.(2)To improve invariance of color image feature descriptor, a new feature descriptor combining with color moment feature in HSV color space is proposed in this paper. Firstly, color image is transformed from RGB color space to HSV color space. And then, the color information of three components H, S and V is expressed with the method of quaternion. Meanwhile, the SURF features of each feature points are calculated among the three channels by using quaternion matrix. Secondly, the color moment feature is extracted in the corresponding feature description region in the same color space. The final feature description vectors are consisted of SURF features and color moment features. In simulation experiments, comparing with SURF algorithm and feature description methods of color image in recent years, our algorithm shows higher correct matching ratios of feature points. And the time complexity of our algorithm is lower than other feature description algorithms of color image.