Image Noise Reduction with Residual-based Edge Reconstruction

Abstract

Image noise is often generated during images acquisition, processing, and transmission. The image noise degrades performance of various processing applications such as image segmentation, video coding, and medical imaging. Image denoising is therefore a very important process for improving the image quality. A variety of the image denoising methods have been proposed. They include anisotropic diffusion (AD) [1], bilateral filter [2], and non-local means filter [8]. Among these methods, the AD model is widely used for the image denoising because of its good denoising performance with simple operations. Based on the AD model, diverse denoising methods have been proposed [3-5]. However, the existing denoising methods do not effectively distinguish between weak image details and image noise, thereby they cannot preserve image details effectively during the noise reduction. This thesis proposes an AD-based denoising method that utilizes the residual images in wavelet domains to improve the quality of image detail preservation. The proposed method consists of three steps: a noise reduction channel, a detail extraction channel, and a combination channel. In the noise reduction channel, the typical AD is applied to eliminate the image noise. However, because the typical AD do not effectively distinguish between weak image details and the image noise, it induces the loss of image details during the noise reduction process. To compensate for the loss of image details, we extract image details from the residual image in the detail extraction channel. At first, we generate the residual image. Using the residual image, we extract the image details in the wavelet domain. After extracting image details in the sub-channel, in the procedure of combination channel, we add wavelet planes of the noise reduction channel and the detail extraction channel except the HH plane. Because the image noise is the diagonal component, HH plane contains the greatest number of image noise among the wavelet planes. Hence, HH plane is excluded for compensation process. Finally, we perform the inverse wavelet transform to produce the final denoised image. The main contribution of the proposed method is the compensation of detail loss during the denoising by restoring image details from the residual images in wavelet planes. We evaluated the image quality of the proposed approach using peak signal to noise ratio (PSNR) [6] and the structural similarity index (SSIM) [7]. For the test image, we used the Kodak lossless true color image set, the Misc1 image set, the Cannon image set, and the IEC image set that was randomly captured from an IEC62087 video. The images were degraded using Additive White Gaussian Noise (AWGN) with 5% and 10% standard deviations (σns) to generate noisy images. For the benchmark methods, we used the typical AD (TAD) [1], Weickert's AD (WAD) [3], Chao's AD (CAD) [4], and Li's AD (LAD) method [5]. All of them are AD-based noise reduction method. In the comparisons, the parameters of each benchmark methods were adjusted to obtain the best PSNR and SSIM for each noise level. The proposed method showed higher PSNR and SSIM values for each noise level, when compared to the benchmark methods. The proposed method suppressed the image noise effectively and preserved image details compared to the benchmark methods.