Illumination Invariant Feature Extraction Using Multi-Scale Relative Gradient Difference

Feature extraction is a very active research hot spot in some fields, such as computer vision, artificial intelligence and other professional. It is a very typical image understanding, image treatment and image classification problem. It is a combination of the computer graphics, image analysis, human-computer interaction, artificial intelligence, computer vision, physiology, and so on. And a very important part of pattern recognition is the feature extraction. It is a most basic problem, and also a most important basic image prccessing question. At present, the technology of the feature extraction is still faced with the challenge of uncertain factors changes such as posture, expression changes and light changes. The difference of the same object under different light conditions in the image is even larger than the difference between different object under same light condition. The light will lead to changes in the situation of severe degradation in an image. Changes in the light direction and the intensity will lead to the change of the shape and the location of the shadow image, in which the gradient changes occur in the case of glare. These changes make the sample image look completely different. In addition, and changes in lighting are always accompanied by other problems such as the change of the posture and time, which increase the complex of pattern recognition problems. It makes the current image recognition algorithm cannot obtain a satisfactory result, thus illumination change became one of the major bottlenecks which restricts the development of image processing technology.Image recognition is a difficult pattern classification problem due to the differences in class (the change between the images of the same object) even greater than the differences between classes (image from the differences between different objects). A key challenge of current image prcessing and recognition technology is how to eliminate illumination change between training and testing images. About the application of image feature extraction, the ability and robustness of light insensitive feature in a large database is limited and insufficient. Changes in various light environment, these approaches could realize the performance of the underlying. With the light changing, when training or test images completely exposed to different light conditions,2d surface may face serious degradation. These changes of the size of the inherent characteristics, could recognition different objects, usually less than their amplitude changes caused by illumination change images.This paper presents an approach for weakening or eliminating illumination variations using multi-scale relative gradient difference information. In this paper, the shadow and edges of the image will make the derivation of image not reasonable, and we use the Gaussian kernel function for filtering the input image to get a smooth image. In the suggested method, we smooth the illumination picture by using Gauss kernel function. The relative gradient of the image is obtained by calculating the image matrix. The relative gradient difference information is formulated as illumination invariant feature. Effective texture features of picture are getted by multi-scale analysis using Gabor wavelet. Then the illumination invariant feature can be used for pattern recognition. The main advantage of the method is that it can effectively eliminate the influence factors of various illuminations. This paper uses L1 as a minimum distance classifier classification. These results shows the robuness about light of the proposed method.