Research On Image Denoising Method Based On Integrated Support Vector Machine

Image is the basis of human vision,the objective reflection of natural things,and the important source of human understanding of the world and human itself.However,in the process of acquisition and transmission,image is often afected by many inevitable noises.Image denoising is to find a way to effectively remove noise,at the same time,it can protect the original details of the image from being changed.The noise criterion of traditional switch image denoising method is too single,which is very easy to miss and misjudge.Support vector machine(SVM),as a machine learning method for small samples,is widely used in image classification,segmentation and recognition.The combination of SVM and ensemble learning is more suitable for image denoising.Ensemble learning can comprehensively consider the differences among multiple sub-classifiers and improve the detection rate of image noise.In this paper,two methods,error diversity ensemble SVM(SVM-WCEC)and iterative update weight ensemble SVM(SVM-IUSW),are designed for image denoising.The specific work is as follows:SVM-WCEC trains multiple sub-SVMs to form ensemble SVM noise classification model based on the error diversity among sub classifiers.Training samples and test samples are formed by local binary features and weighted difference features of each pixel in the image extracted by 3 × 3 moving window.Training samples train multiple sub-SVM classifiers,calculate the weighted count of errors and correct results(WCEC)according to the classification results of different sub-SVM combinations,and vote to get the ensemble SVM noise classification model.After filtering only for the noise pixels,the denoised image is reconstructed with the value of the signal pixels.Experimental results show that the average PSNR and SSIM of BSD68 image set restored by SVM-WCEC algorithm are 1.8080dB/0.1504 higher than that of DAMF algorithm.SVM-IUSW adopts the ensemble idea of iterative updating weight to build the ensemble SVM noise classification model.The first SVM classifier is trained after the initial weight is set for the training sample,and the weight of the training sample input to the next sub-SVM classifier is updated according to the previous classification result,and the weight of the classification result in the decision-making of the ensemble noise classification model is determined according to the classification error.Experiments show that the average G-mean,F-1 and Accuracy values of SVM-IUSW classifier model are 1.56,0.62 and 0.33 higher than those of SVM-WCEC classifier model,and the average PSNR and SSIM values of restored BSD68 image set are 1.9627dB/0.1560 higher than those of DAMF algorithm.Figure[33]table[10]reference[65]...