Enhancing Low-Light Images with a Lightweight CNN-Based Visual AI Approach

This paper proposes a novel concatenated method for Deep Curve Estimation (DCE), inspired by Zero-DCE, targeting low-light image enhancement. Our streamlined approach utilize sanultra-light weight deep network for image-specific curve estimation, enabling dynamic range correction for superior image quality. Unlike prevalent Generative Adversarial Networks (GAN) and Convolutional Neural Networks (CNN) methods relying on paired images, our technique trains without need for paired/reference images. Through exhaustive experimentation with convolutional layers, loss functions, filters, and epochs, we optimize our method for enhanced Peak Signal to Noise Ratio (PSNR), yielding superior image quality. The result is an exceptionally light model, surpassing existing methods and displaying real-world applicability. With a focus on light weight architecture and superior enhancement, our approach provides a promising avenue for practical deployment.