BOLD-net: Brightness enhancement for old images using deep curve estimation and attention modules

Abstract

In this paper, a novel deep-learning network for the brightness enhancement of old images is proposed. Although the task for brightness enhancement of old images is similar to the low-light enhancement problem of modern images, the causes of darkness and image characteristics are significantly different. Unlike modern images, bright degradation in old images is due to low-quality camera sensors, unsophisticated techniques that capture the images, and the harsh environment where the photos/films were stored. Though existing low-light enhancement networks show good reconstruction capabilities for modern images, they result in overexposed, color-distorted outputs if applied to old images. To resolve these, we propose a novel deep-learning network with a combination of two convolutional neural networks, which estimates curve maps to adjust the dynamic range, and an attention-guided illumination map to adjust the illumination of the given input image. In addition, channel attention blocks in the illumination map estimation network are further performed to reduce image noise and prevent over- or underexposure. The networks perform in parallel to enhance the brightness of the image and preserve inherent colors and texture details. The proposed method was evaluated on the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS) and compared with the state-of-the-art methods on a dataset containing old video frames from different decades. Experimental results show that the proposed method outperforms state-of-the-art methods in terms of visual quality for brightness enhancement on old photo and video datasets.