EnhanceNet: A Deep Neural Network for Low-Light Image Enhancement with Image Restoration

An important task in computer vision is improving low-light images, which aims to increase the quality of images collected in low light. Our study has introduced EnhanceNet, a deep learning-based technique, in this area. The DCE-Net low-light image enhancer architecture and the NAFNet baseline architecture make up this model. The model increases pixel intensity from a low-light image as input to enhance image quality, reduce grain, and remove blur. Altering the exposure, contrast, and gamma as well as adding filters like a histogram equalization or a simple linear transformation may all be used to achieve this. For image denoising and deblurring, we pass the enhanced image with a nonlinear activation-free network or NAFNet. The proposed combination model is trained on our own large dataset to uncover the association between low-light photographs and their corresponding enhanced versions. According to experimental results on numerous datasets like LOL, SICE, LIME, and NPE, the proposed approach provides state-of-the-art performance in terms of visual quality and quantitative measurements like PSNR value 60% more efficient than previous state-of-the-art model. The proposed EnhanceNet is computationally efficient since it has a small model size and a powerful design. This network can efficiently be implemented on edge devices for various applications like surveillance, automobile, and photographic camera. In this paper, the performance of the model on Android devices has been presented.