Image Denoising for an Efficient Fake Image Identification
B. Judy Flavia, G Sharnish, Prashanth Mishra, Aashirvad Kohli · 2022 International Conference on Edge Computing and Applications (ICECAA) · 2022
Noise and blur may be removed from a photograph during picture restoration. Erasing camera shaking, radar imaging, and the influence of an image system’s reaction to blurry images is challenging in many instances, including photography. Unwanted signals such as thermal or electrical noise or precipitation or snow in a picture are examples of image noise. The deterioration of the picture might be caused by coding, resolution restriction, transmission noise, object motion, camera shaking, or a combination. A sensor, such as a thermal or electrical signal, or an ambient, such as snow or rain, might cause an image to be blurred. There are several probable causes of picture degradation, including coded pictures, resolution restrictions, transmission noise, object motion, camera shaking, and these factors. Distinguishing between high- and low-frequency components may be accomplished by a process known as "picture decomposition," which entails separating a distorted image into two distinct layers, one for texture and the other for structure, from the LF. The present approach is based on a deep CNN architecture that is very customizable and takes advantage of the frequency characteristics of different sorts of artifacts. The same technique may be used for a wide range of image restoration projects by just altering the architecture. Using a quality improvement network based on residual and recursive learning is recommended to reduce noises with similar frequency characteristics. In order to prevent the network from dying, the authors used residual learning to speed up the training process. The researchers also developed new auxiliary classifiers. The surtax was applied to the outputs of two inception modules, and an auxiliary loss over the same labels was calculated, much as in the prior experiment. Weighted averages are used to calculate the total loss function. Recursive learning may drastically decrease the number of training parameters while maintaining or increasing performance, as seen below. It is claimed that the proposed system is built on a pure version of Inception that does not include any leftover connections. Memory optimization or backpropagation may train it without splitting the copies.