Local Backlight Dimming for Liquid Crystal Displays via Convolutional Neural Network

This paper presents a new local backlight dimming (LBD) for liquid crystal displays (LCD) method based on a convolutional neural network (CNN). Many previous LBD algorithms controlled the backlight intensity relying on hand-crafted features within a local block, that is, statistical information of pixel values in each block. However, they have a lack of generalization ability due to the use of hand-crafted features, which are usually not adaptive to the input properties. Also, they usually disregarded the diffusion property of the backlight that may affect the neighboring blocks. In this respect, we propose a CNN-based LBD algorithm to alleviate these problems. To address the lack of generalization ability of hand-crafted features, we adopt a CNN-based approach that learns the features and thus provides appropriate backlight intensities for the given inputs. Also, the diffusion property of light and leakage property of liquid crystal are considered when training the network, thereby alleviating the loss of details while achieving the high contrast ratio. Experiments show that the proposed method outperforms both quantitatively and qualitatively compared to the other LBD algorithms. Specifically, for the images from the DIV2K dataset, the proposed method achieves at least 1dB enhancement in PSNR, showing the generalization performance.