Authors: Research Scholar Arpit Urmaliya, Professor Jitendra Kumar Maitra
Abstract: This paper is intended to create and analyze a novel Color-Aware Adaptive Wavelet Transform (CA-AWT) method for natural color-image denoising with a special focus on the robustness to speckle noise which is realized by adaptive blending and image-based noise estimation. All 68 CBSD68 color images were used in the experiments performed in MATLAB. The images have been converted from RGB to YCbCr and corrupted with six different types of noise (Gaussian, salt-and-pepper, speckle and three mixed-noise types) at three different noise levels (0.05, 0.10 and 0.15), with the result that 1,224 controlled noisy-image experiments were generated. MSE, RMSE, PSNR, SSIM and MAE were used to evaluate CA-AWT baseline V1 and progressive variants V2-V4. The severity dependent blending factors identified from speckle specific ablation analysis encouraged the implementation of V5 continuous noise-estimated adaptation and V6 calibrated nonlinear noise adaptation. Paired t-tests, Wilcoxon signed-rank tests and Cohen's d were used for statistical validation. The results indicated that CA-AWT was effective in most Gaussian and salt-and-pepper noise conditions, but was not effective at low levels of speckle noise, thus necessitating adaptive processing. In the presence of speckle noise, V4 consistently outperformed V1 by 1.277–4.848 dB with large to medium effect sizes. In general, V4 obtained the highest mean PSNR (29.7301 dB) and V6 obtained the highest mean SSIM (0.901595) and the PSNR (29.5587 dB) was also competitive. V6 is thus a potential method of adaptation based on images, but V4 is the present reference method until it is validated in this data set and in the future.
