Study On Image Denoising And Its Effect Evaluation

From the rock paintings in stone age to the digital images in modern society, images have carried 80% of the amount of human information. As an important visual information carrier, the image is an essential source and means for human to obtain and utilize information. However, the image will be inevitably contaminated by noises in the process of image generation and transmission. Noises bring many difficulties to image processing, which have direct influence on image segmentation, feature extraction and image recognition. Therefore, image denoising process has been a vital pre-processing step of these image processing. In this paper, image denoising methods have been studied deeply, and a new effect evaluation metric for image denoising methods is proposed. The main research work of this paper can be summarized as:1. Based on the feature of images with rich texture, a novel image denoising algorithm based on morphological component analysis(MCA) is proposed. First of all, the MCA method is used to decompose the source image into two parts: low frequency part and high frequency part. Secondly, the bilateral filter is used to smooth the latter one. As single pixel’s similarity is affected by noises in traditional bilateral filter, pixel smooth weights are brought into our algorithm. It can effectively preserve the texture and edge information. Finally, low frequency part and the filtered high frequency part are reconstructed to accomplish the denoising process. The denoised images contain explicit and integrated textures and higher PSNR.2. A new image denoising algorithm based on empirical mode decomposition is proposed. The advantages of the EMD involve no need on pre-setting decomposition level and decomposition bases, and it can decomposed any image into an smoothed image and several empirical modes adaptively. For the noises that are mostly contained in empirical modes, neighboring information relation matching algorithm is proposed to smooth them. Finally, the inverse empirical mode is used to reconstruct the denoised empirical mode function and the trend image to obtain the result. Experimental results show that the proposed algorithm is superior to other algorithms in both subjective evaluation and objective evaluation.3. A novel decomposition method with directional information based on the empirical mode decomposition is proposed. First, adaptive 2D empirical mode decomposition is used; then, 2D empirical mode decomposition with directional information is proposed, and it can generate several directional sub-bands. Refined linear regressive model is used to smooth the directional sub-bands. Finally, combine the smoothed sub band image with the residual of original image to get the final image. Experimental results show that the proposed can remove noises without loss of images’ directional information.4. A full reference image quality measure is proposed. The evaluation covers the denoising ability and the one of preserving vital visual information of original images. For the first aspect, a new noise detection method is proposed. First of all, the suspected noises are selected according to the relationship between the pixel and its neighbors. Then, the true noise points are determined by comparing the reference image. Finally, judge the percentage of noise removal. In the aspect of visual similarity, the structural similarity measure(SSIM) is used to measure the similarity between denoised image and reference image. Experimental results show that the proposed evaluation measure has obvious advantages in evaluating the ability of denoising, and it is consistent with the subjective evaluation.5. The influence of noises on image segmentation is evaluated. Noises with different strength are added into images before object extraction. In this paper, two kinds of images are used: the first one is lip image, we adopt a RGB color space based regional growing algorithm to find the contour of lip. The other one is fluorescent images, we adopt a Watershed algorithm based on HSV color space to separate objects and background.