A Comprehensive Analysis on Regularization Techniques and Activation Functions on Image Denoising with Autoencoder

Image denoising through an autoencoder has emerged as a powerful and versatile tool in the domain of image denoising. However, overfitting and capturing unnecessary noise-related features remain significant hurdles. Regularization techniques are essential for guiding the learning process and boosting desirable qualities in the denoised output to overcome these challenges. This comprehensive study explores the various regularization techniques used in image denoising autoencoders and looks at how they affect the model’s resilience, performance, and overall denoising abilities. We evaluate the efficacy of different autoencoder architectures, such as convolutional autoencoders, on the MNIST dataset. Specifically, we assess their performance when subject to various types of noise, including Gaussian, Salt and pepper, Poisson, and Speckle noise. The evaluation metrics we utilize are Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). We examine the impacts of conventional regularization techniques, such as \(L_1\) and \(L_2\) regularization, on these architectures and evaluate the influence of various activation functions on the denoising process. Also, we examine the impact of noise level on denoising performance. Our findings reveal that a convolutional autoencoder, combined with \(L_2\) regularization, yields superior results in terms of denoising performance. The analysis provides insightful information on how autoencoder architectures, activation functions, and regularization strategies work together to enhance image denoising.