An AI Based Lightweight Image Processing Model for Resource-constrained Architecture

Abstract: The quality of images captured by budget-friendly smartphones degrades significantly in low-light conditions due to hardware limitations. To resolve this, we present a lightweight, end-to-end deep learning framework designed to function as a software-based Image Signal Processor (ISP) for on-device enhancement. Our approach is centered on a U-Net architecture, trained on a hybrid dataset combining specialized low-light pairs (LoL) and general high-quality photographs (MIT-Adobe FiveK) to ensure robust and aesthetically pleasing results. The model, which contains only 2.90 million parameters, is optimized using a composite loss function balancing pixel-wise accuracy and structural integrity. Quantitative evaluation shows our model achieves a highly competitive PSNR of 17.24 dB on the LoL Dataset. A key finding from our ablation studies reveals that for a network of this scale, a simpler architecture without residual connections performs marginally better, providing a valuable insight for future lightweight model design. Overall, our work demonstrates a superior trade-off between performance and computational efficiency, establishing a promising foundation for bringing superior photographic computation on a variety of mobile devices.

Keywords: Deep Learning, U-Net, Mobile ISP, CNN, Lightweight Neural Networks, On-Device AI, Computational Photography, Edge Computing, Low-Light Image Enhancement.