Research On Texture Classification Under Varying Illumination

Texture, capturing the main feature of natural object, is a vital part of human visualinformation. And visual perception of texture is an import way for human to understandthe world. As an effective description of the image pattern, texture description andclassification has been a lasting concern and leading subject of researches in the area ofcomputer vision and pattern recognition. Along with the rapid development of computervision technology, various systems based on texture classification and recognition havebeen widely used in the field of industry manufacture, medical diagnosis, weatherforecasts, geological exploitation and etc.In practical applications, ambient light changes usually are uncontrollable. Both thegray value distribution and the textures of the image change significantly with respect toillumination changes. Similarly, the performance of texture classification andrecognition algorithms could vary significantly under varying illumination condition.Thus, removing the effect that changing illumination impacts on the textureclassification and recognition system, has been one of most challenging task in the areaof computer vision and pattern recognition. Based on the multi-resolution analysistheory and study of the effect that illumination impacts on texture image, weconcentrated ourselves on the effort to extract illumination invariant of texture image inthis paper. And then we employed the illumination invariant to perform our textureclassification algorithm. The main work is as follows:①.Based on reflection model, we engaged ourselves on the research involving inthe perturbations that introduced by changing illumination conditions to the gray leveldistribution and texture details of color images. By modeling the perturbations inmulti-scale space, we proposed our illumination invariant extracting model in this paper.In our model, the perturbations were each considered as the effect that illuminationimpacting on approximate component and detail image. According to the differentextent of two kinds of effect, different strategies were employed to eliminate the twotype of influence.②. Based on the idea of denoising and our illumination invariant extracting model,we proposed our new schema founded on wavelet transform and BayesShrink denosing.In our method, ideal image was regarded as noise signal. Firstly, the input image wastransformed to log-domain. The wavelet transform was then performed on the logarithmic image. In multi-scale space, we employed a low-pass filter to perform onthe approximate component which was considered to be an approximation of the grayscale distribution of original image. Meanwhile, BayesShrink denoising was thenadopted to deal with to the high frequency wavelet coefficient components, whichmainly consisted of texture details. Thus, the illumination distribution image was gainedby inverse wavelet transform. And we obtain the illumination invariant with respect toreflection model by subtracting illumination distribution image from the log-observedimage. In our texture classification experiment, primary component is utilized to cutdown the illumination invariant feature dimensions. And a K-Nearest Feature Lineclassifier was used to classify the testing image. Experimental results showed that ourmethod gained better result than existing method such as traditional illuminationinvariant extracting method based on wavelet transform, LBP, methods based onpreprocessing by color constancy and etc, with classifying accuracy higher from5.56%to22.1%...