Learning point and contextual processing networks for low-light image enhancement

Abstract

Point processing techniques, such as the gamma correction, are classical methods for low-light image enhancement. These methods are efficient, explicit, and interpretable. However, to tackle images of various light conditions, these methods need to be set meticulously and accordingly. To tackle this issue, I propose to combine these traditional methods with the powerful learning ability of convolutional neural networks (CNNs). Specifically, I propose a point and contextual processing network (PCPNet) consisting of two parallel branches: In the point processing branch, given an input image, a set of hidden intensities images (HIIs) are predicted, as well as a set of parameters governing the gamma corrections performing on the HIIs. Due to the nonlinearity of the gamma corrections, we can obtain diverse and complex enhancement effects merely with a shallow network (of 3 layers). This further guarantees the efficiency of the point processing branch, and allows the HIIs to be generated with the full resolution and preserve the details. While in the contextual processing branch, an encoder-decoder structure is adopted to explore contextual information, which helps to alleviate the effects of noise. The outputs of these two branches are combined as the final enhancement result. Extensive experiments on multiple datasets, including LOL, MIT, and ExDark, validate the effectiveness of the proposed method.