Blind DCT-based prediction of image denoising efficiency using neural networks

Abstract

Visual quality of digital images acquired by modern mobile cameras is crucial for consumers. Noise is one of the factors that can significantly reduce visual quality of acquired data. There are many image denoising methods able to efficiently suppress noise. However, often in practice denoising does not provide sufficient enhancement of images or even demonstrates visual quality reduction compared to observed noisy data. This paper considers the problem of prediction of denoising efficiency of images in a blind manner under additive white Gaussian noise condition. The proposed technique does not require a priori knowledge of a noise variance and uses a moderate amount of image data for analysis. The denoising efficiency prediction employs neural networks (all-to-all connected multi-layer perceptron)to create a regression model. Image statistics obtained in the spectral domain are used as input data and the state-of-the-art visual quality metrics are considered as outputs of the network. As a target denoising method, block matching and 3D filtering (BM3D)technique is used. It is demonstrated that the obtained neural networks are compact and overall prediction procedure is fast and has an appropriate accuracy to confidently answer to the question: “Do we need to denoise an image?” The full dataset, executable code and demo Android application is available at https://github.com/asrubel/EUVIP2018.