Improved generalised fuzzy peer group with modified trilateral filter to remove mixed impulse and adaptive white Gaussian noise from colour images

In image processing applications, image denoising is an emerging area that recovers the original image from noisy image, which is essential in the applications like pattern analysis. The main aim of this study is to propose an effective filtering technique with peer groups for effective image restoration processes. Existing bilateral filter has lower performance in smoothing the high curvature and high gradient regions due to nearby input signals are outliers that miss the filter window. In this research paper, a new approach is proposed that includes fuzzy-based approach and similarity function for filtering the mixed noise. In a peer group, the similarity function was adaptive to edge information and local noise level, which was utilised for detecting the similarity among pixels. In addition, a new filtering method modified trilateral filter (MTF) with improved generalised fuzzy peer group (IGFPG) is proposed to remove mixed impulse and adaptive white Gaussian noise from colour images. The modified trilateral filter includes Kikuchi algorithm and loopy belief propagation to solve the inference issues on the basis of passing local message. In this research work, the images were collected from KODAK dataset and a few real time multimedia images like Lena were also used for testing the effectiveness of the proposed methodology. The collected colour images were contaminated with adaptive white Gaussian noise (AWGN) of standard deviation σε[05,20] and impulse noise of probability pε[0.05, 0.20]. The proposed MTF-IGFPG method has the advantages of consider image photometric and geometric similarities. The local structural similarity and narrow spatial window is applied for smoothing the images to effectively preserve the images. From the simulation, the proposed MTF-IGFPG approach attained better performance related to the existing approaches in light of normalised colour difference (NCD), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE). Hence, the proposed approach almost achieved PSNR of 36 dB, which shows 0.2 dB to 2 dB improvement than the existing methods.