A Lightweight Denoising Convolutional Neural Network for On-Device Artifact Suppression

Image compression for mobile and streaming applications often introduces blocking, blurring, and ringing that degrade visual quality and harm downstream vision tasks. This work presents a lightweight on-device restoration model based on a Denoising Convolutional Neural Network (DnCNN) that is optimized for efficiency using structured pruning, 8-bit integer (INT8) quantization, and architectural slimming, followed by perceptual fine-tuning in MATLAB. The model was trained on the Berkeley Segmentation Dataset 400 (BSD400) and evaluated on Set5, Set14, and Berkeley Segmentation Dataset 68 (BSD68). We report standard full-reference metrics, namely Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), a perceptual metric, Learned Perceptual Image Patch Similarity (LPIPS); and no-reference metrics, Natural Image Quality Evaluator (NIQE) and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE). On average, the compressed model attains about 29.0 dB PSNR and 0.83 SSIM, while reducing model size by about 52% to 1.0 MB and cutting CPU inference time by about 70% compared with the uncompressed DnCNN baseline. These results show that the compressed and perceptually fine-tuned DnCNN suppresses compression artifacts effectively while meeting the memory and latency constraints of mobile and embedded platforms, providing a practical receiver-side solution that remains compatible with legacy codecs.