An Efficient Autoencoder-Decoder Model for Image Denoising Using Deep Convolutional Networks

Noise in any form is a crucial concern in signal and image processing tasks, including image inpainting, restoration, segmentation, classification, prediction, etc. The primary objective is to deal with such unwanted distortions to clean the signals and images and thus form the prerequisite of any image and signal processing applications. In recent years, deep convolutional networks with proper configuration and hyper-parameter tuning have proved a boon in various applications. The present article presents an Autoencoder-Decoder (AED) network model for image denoising influenced by Gaussian noise. The model successfully eliminated the noise at the bottleneck of the AED and reconstructed the images with high clarity. We evaluated our proposed AED model over PASCAL VOC 2007 dataset images due to varied image diversities and found that the model performed better for a standard deviation value equal to 10. The average peak signal-to-noise ratio (PSNR) and the structural similarity (SS) evaluated over the test images were 30.864 and 89.0%, respectively, for the images resized to 64 × 64 sizes. Over CIFAR-10 dataset images with modified AED, the values were found to be 34.6 and 99.6%, respectively.