Comparative Study of Variational Autoencoder for X-Ray Image Denoising
Asha Rani, M Akhil Raj, Bhaveshkumar Choithram Dharmani · 2024
Noise in images is the main signal distortion that obstructs the image analysis process and extraction of information. Because of environmental effect transmission channel and the retrieval process, the image gets impaired by noise during acquisition, transmission and storage that leads loss of image information. So, image denoising becomes the crucial preprocessing step in image processing. Its main aim is to recover clean images from noisy observations. Search for a potential deep learning architecture for denoising is the need of the hour. This study presents a comparative evaluation of variational autoencoder, autoencoder and median filter in the context of image denoising. In our study, variational autoencoder, which is relatively the latest deep learning algorithm for denoising, has been used and results are compared with autoencoder and median filter. First, we have implemented a standard autoencoder architecture, which is a deep learning model. It learns to map noisy images to their clean counterparts through encoder decoder framework. AEs are widely used because of its simple structure and effectiveness. Then we have used variational autoencoder, which is relatively the latest deep learning algorithm for denoising. It is a probabilistic variant of autoencoders which generates the latent space of data with a probability distribution to generate more realistic denoised image compared to conventional autoencoder. Lastly, we implemented Median filter as a baseline technique which is classical image processing technique, open-source X-ray images datasets have been used. Noise variance and signal-to-noise ratio are computed on the X-ray images. The study finds that variational autoencoder outperforms the traditional AEs and median filter. In conclusion, this study highlights the potential of deep learning-based approaches particularly VAE as an effective method for image denoising with PSNR=68.758 db. Results obtained with variational autoencoder are better than autoencoder and median filter.