A Comparative Study of Classical and Deep Learning Approaches for Non-Uniform Motion Blur Restoration

Abstract: Motion blur caused by fast or irregular relative motion between a camera and its subject remains a challenging image restoration problem, particularly when the blur is spatially non-uniform and follows a curved rather than linear trajectory — conditions under which standard deblurring assumptions break down. This work presents and rigorously compares two independent restoration pipelines: a classical approach combining per-tile blind kernel estimation (cepstrum-based and gradient-prior methods) with Richardson-Lucy deconvolution, and a lightweight deep learning model (NAFNet-lite, ~0.44M parameters) trained on a custom-generated synthetic dataset simulating non-uniform, curved motion blur via randomized Bezier-path kernels. Both pipelines are evaluated on a shared held-out benchmark using PSNR, SSIM, and an auxiliary readability-based metric. Results show the deep learning model consistently outperforms the classical baseline on pixel-fidelity metrics (SSIM 0.955 vs. 0.705), while the classical pipeline achieves comparable or better rates of individual-sample improvement despite requiring no training. An ablation study reveals that total-variation regularization, commonly used to suppress deconvolution ringing, measurably degrades fine-detail recovery in this setting. Further, an attempt to improve real-world generalization by sequentially fine-tuning the deep model on a real-world photographic blur/sharp dataset resulted in significant catastrophic forgetting of performance on the synthetic benchmark, highlighting a practical limitation of naive fine-tuning and motivating joint or replay-based training strategies as future work. A complete, interactive web application was developed to demonstrate both restoration pipelines side by side.

Keywords: Motion Deblurring, Blind Deconvolution, Non-Uniform Blur, Richardson-Lucy Deconvolution, Deep Learning, NAFNet, Synthetic Data Generation, PSNR, SSIM, Ablation Study, Catastrophic Forgetting, Transfer Learning, Web-Based Deployment.