Training of Perceptual Image Denoising Network with Weighted Sum of IQA Methods
Takamichi Miyata · 2023
Deep learning-based image denoising methods using mean squared error (MSE) as the loss function produce results of an excessively smoothed image with low perceptual quality. Although advances in image quality assessment (IQA) methods, including those using deep learning, allow us to estimate the perceptual quality of images, an existing study reveals that using such IQA alone as a loss function of the denoising methods does not improve the perceptual quality. This may be because each IQA has a special image (adversarial example) that causes the IQA to malfunction. To avoid this problem and improve the perceptual quality of denoised images, we propose a method for learning denoising methods using a loss function that combines IQA with other IQA or MSE. The experimental results show that the proposed method can improve the quantitative and qualitative perceptual quality of the image by suppressing excessive smoothing of the image.