Embedded Deep Learning to Improve the Performance of Approaches for Extinct Heritage Images Denoising

Abstract

Many advanced deep convolutional neural network (DCNN) methods have proven their efficacy in reconstructing the texture of super-resolution images (SR) from low-resolution images (LR).   Nevertheless, the objective of achieving super-resolution (SR) reconstruction using Deep Convolutional Neural Networks (DCNN) becomes difficult when the input image is distorted by noise.  Photographs captured at the inception of the camera are presently regarded as a cultural heritage that chronicles an important period in human history; however, they are marred by low resolution and noise as a result of obsolescence and the primitive nature of the technology that captured them, in contrast to the technological advances that cameras benefit from today. We proposed embedded deep learning to improve the performance of approaches for extinct heritage images denoising to denoise and reconstruct Baghdad heritage images. First, stage super-resolution (SR) noisy image generation from low-resolution (LR) heritage noisy image aims to enable to extraction of noise features for the target images. Second, remove visible noise features on images and restore their surface texture, giving them a more modern and clearer scene while preserving their original identity. PSNR, SNR, and SSIM quantitative metrics and the visual comparison analysis between the proposed method and state-of-art methods: Total variation denoising (TV), Bilateral filter denoising (BF), Median filter Denoising (MF) Gaussian filter image denoise (GF), and Non-Local Bayes (NLB) denoising demonstrated a better performance in reducing noise from the target images and a high-frequency flow with more information. Our approach restores heritage in a way that mimics modern photographic scenes using deep learning algorithms.