Enhancement of low-light images is a challenging task due to the impact of low brightness, low contrast, and high noise. The inability to collect natural labeled data intensifies this problem further. Many researchers have attempted to solve this problem using learning-based approaches; however, most models ignore the impact of noise in low-lit images. In this paper, an encoder-decoder architecture, made up of separable convolution layers that solve the issues encountered in low-light image enhancement, is proposed. The architecture is trained end-to-end on a custom low-light image dataset (LID), comprising both clean and noisy images. We introduce a unique multi-context feature extraction module (MC-FEM) where the input first passes through a feature pyramid of dilated separable convolutions for hierarchicalcontext feature extraction followed by separable convolutions for feature compression. The model is optimized using a novel three-part loss function that focuses on high-level contextual features, structural similarity, and patch-wise local information. We conducted several ablation studies to determine the optimal model for low-light image enhancement under noisy and noiseless conditions. We have used performance metrics like peak-signal-to-noise ratio, structural similarity index matrix, visual information fidelity, and average brightness to demonstrate the superiority of the proposed work against the state-of-the-art algorithms. Qualitative results presented in this paper prove the strength and suitability of our model for real-time applications.
INDEX TERMSEncoder-decoder Architecture, Separable Convolution, Dilated Convolution, ASPP, Perceptual loss, Low-light Image Enhancement I. INTRODUCTION L OW light image enhancement is an active area of research that enables the acquisition system to capture superior quality images even under low-light conditions. Its applications include autonomous driving, photography, military, object detection, and surveillance. Low-light image enhancement (LIE) algorithms consider several factors like color contrast, brightness, image resolution, and dynamic range. Several researchers use image processing techniques to enhance low-light images, such as histogram equalization (HE) [1] that tend to equalize the dynamic range of intensities. Also, histogram equalization methods focus only on increasing the image contrast, whereas they fail to address the actual illumination issues. As mentioned in [2], linear and non-linear low-light image enhancement methods have simple yet fast implementation but don't consider the image distribution, leading to limited enhancement ability. Spatial filters and other image processing techniques have shown 18 improvement in the low-lit image quality, but these methods 19 fail to work under noisy context. Noise, typically, is amplified 20 by the filters, as they rely on a small neighborhood. These 21 reasons justify the need for deep-learning in enhancing low-22 light images, as these methods can dynamically learn and 23 handle noisy as well as cl...
