Answers are grounded in this paper and cite their sources.
Single-image motion deblurring restores detail degraded by camera or object motion. We present the Directional Convolutional Neighborhood Attention Transformer (DiCoNAT), a two-stage residual restoration network. Within each stage, native and 90° rotated views share one encoder–decoder mapping. Inverse-aligned bottleneck features are added before a decoder with dilated neighborhood attention predicts image updates. The stages have identical topology and independent parameters. Trained on GoPro, DiCoNAT achieves 32.86 dB peak signal-to-noise ratio (PSNR) and 0.9603 structural similarity (SSIM) on GoPro, and 30.66 dB PSNR on HIDE without further training. The network contains 7.5 million trainable parameters. With 3000 training epochs, the one-, two-, and three-stage configurations achieve 32.09, 32.86, and 32.25 dB PSNR, respectively. The ResBlock-encoder/DiNAT-decoder allocation achieves 32.86 dB, compared with 31.61 dB for the reversed allocation. Additional controls assess rotation angle, bottleneck fusion, frequency-loss weight, dilation, and the number of attention heads. On restored GoPro images, YOLOv8-nano class-averaged mAP@50 increases from 54.4 for blurred inputs to 63.8. In a 35-participant study covering 30 scenes, DiCoNAT receives the highest mean perceptual rating among the evaluated methods, at 3.18.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: