Noise Detection And Image De-noising Based On Multi-feature Combination And Support Vector Machine Ensemble

Image de-noising is a key pretreatment link before image is processed. For noisy image, most noise and image details are hard to be distinguished because they distribute in high frequency area, which leads to the loss of image detail information in different degree while noise points are removed. Therefore, how to retain image details to the utmost extent while noise points are removed is a research focus.Aiming at the problem that the approach for removing noise from images based on SVM fails to gain better de-noising effect and leads to low performance of classifier and noise recognition rate when only rely on a single image feature, an image de-noising method based on multi-feature combination and SVM is proposed on the basis of analyzing and summarizing the algorithms related to image de-noising. According to the adjacent pixels correlation in the image and the characteristics of salt-pepper noise, the method comprehensively describes the pixels by using the way of multi-feature combination. Thus noise points and non-noise points can be differentiated accurately. Experimental results show that the image de-noising method based on multi-feature combination and SVM gains better de-noising effect. In addition, the support vector classifier based on multi-feature combination has higher noise recognition rate and classification performance compared with support vector classifier based on single image feature.Considering that SVM ensemble has better classification performance, stability and generalization capability, SVM ensemble is applied to image de-noising. A new approach for removing salt-pepper noise based on multi-feature combination and SVM ensemble is presented in this thesis. First, according to the correlation of adjacent pixels in the image and the characteristics of salt-pepper noises, multiple features are extracted from noisy image and constituted sample set. Second, the sample set is normalized. Then the dual disturbance mechanism and majority voting method are applied to construct SVM ensemble classifier. The dual disturbance mechanism disturbs training set and classifier model parameter. Third, the SVM ensemble classifier is used to recognize noise points and support vector regression is applied to predicting the original gray value of noise points. Finally, image is reconstructed so as to achieve the purpose of de-noising. Experimental results show that this approach can further improve de-noising effect, and can retain detail information of image better while noise points are removed. The method is especially effective in lower noise ratio. Moreover, the SVM ensemble classifier based on multi-feature combination has better classification, stability and generalization capability compared with the SVM classifier based on multi-feature combination.