Research On Image Feature Descriptors Based On The SIFT Algorithm

SIFT is a feature matching algorithm based on the scale space, which can keep invariant in the image transformation of scale, rotation and affine distortion. Classical SIFT feature descriptors only use image intensity data and ignore the color information, resulting in the poor performance of color image matching. At present, there are a variety of SIFT algorithms for color images, such as HSV-SIFT, Opponent-SIFT, W-SIFT and RGB-SIFT. The algorithms above calculate an128-dimensional feature vector in each channel, which form a3×128-dimensional feature vector, and realize the task of color targets matching. However, the matching efficiency of these algorithms is generally poor in case of light changes, and the dimensions of their feature vectors are much higher than those of classical SIFT algorithms which are only128-dimensional. Obviously the higher the vector dimension, the more the matching time.In order to avoid the disadvantages of heavy computational complexity and poor matching performance in existing SIFT descriptors for color images, this thesis presents a novel SIFT descriptor for color images based on color quantization matrix, which is called CQM-SIFT descriptor. The color quantization matrix is obtained by triple-color components of the HSV color space. And then it is applied to generate128-dimensional SIFT descriptors to match color objects. The experiment results show that, compared with existing SIFT descriptors for color images, this proposed method has the advantages of better matching performance and less matching time in the situation of illumination change, scale change, rotation change and overall change.Furthermore, in order to decrease the matching time and to improve the robustness of CQM-SIFT descriptor, the PCA is applied in the CQM-SIFT descriptor to generate a new descriptor, which is called CQM-PCASIFT. Firstly PCA transform is applied in the CQM-SIFT descriptor of each feature point to reduce128dimensions of the feature vector to36dimensions. The experiment results show that this proposed method has the advantages of less matching time and better matching performance in the situation of illumination change, scale change, rotation change and overall change.