Research On Local Invariant Feature Based On Gradient Difference And Nonorthogonal Mappings

With the rapid development of computer technology, the image data is significantly growing, image feature extraction has been the key to correctly identify images. Global feature extraction algorithm is simple and has strong commonality. To some extent, it can get some invariance. However, influenced by practical geometric distortion and illumination variance, the global feature effect is not ideal. In recent years, the local invariant features have become one of research focuses in the field of computer vision and successfully used in image retrieval recognition, robot navigation, scene matching etc. Local feature extraction algorithm mainly solves the problem how to extract stable enough features which have good robustness to the transform from scale, rotation, illumination. As a milestone in the research process of local invariant features, SIFT algorithm uses pixel gradient to describe the feature points and has obtained robustness for a variety of image transformations. However, its shortcomings of high computational complexity and big description dimension, led to poor real-time performance. PCA-SIFT which used PCA to low dimension of SIFT descriptor vector from128to36, descends cost reduction in the process of key points matching. However, extral training process and dimension reduction process adds formation time of feature vector and brings the loss of feature information, which didn’t make the whole algorithm computational speed increase, but decreased algorithm matching performance, resulting in the reduction of algorithm matching performance. SIFT descriptor and PCA-SIFT descriptor are both based on gradient in nature, which can simply describe smaller geometric distortion and is very sensitive to big image affine transformation. In addition, the loss of pixels location information around the feature point in the two algorithms, affects the stability of the descriptor to a certain extent. Facing high-dimensional vector and insufficient stability of SIFT, this paper proposes a improved SIFT algorithm (ISIFT), which is based on gradient difference and non-orthogonal mapping moment. Compared with SIFT algorithm, the algorithm not only improves invariance for blurring transformation and illumination transformation, but also its dimension is only60. Firstly computing gradient difference in the block of neighborhood pixels of feature points, non-orthogonal mapping of gradient difference of subblock according to position and direction is done. Then the low level center moment of the non-orthogonal mapping moment in all subblocks is calculated and arranged according to the direction to the descriptor of the feature points. The improved algorithm and SIFT algorithm, PCA-SIFT algorithm, HSIFT algorithm are experimented on Mikolajczyk data sets, so as to identify performance of different algorithms. This paper uses Euclidean distance to compare matching performance of different methods and recall vs.1-precision curve to assess performance of different algorithms. The experimental results show that not only the calculation cost of ISIFT algorithm is low, but the algorithm has also higher matching accuracy rate in blurring transformation and illumination transformation.