An Adaptive Edge-Guided Dual-Network Framework for Decoding-Oriented QR Code Motion Deblurring

Unlike natural image deblurring, which primarily prioritizes perceptual quality, Quick Response (QR) code deblurring aims to ensure successful decoding. QR codes are characterized by highly structured patterns with sharp edges, which provide strong structural priors for restoration. However, existing deep learning methods rarely exploit these priors explicitly. To address this limitation, we propose the Edge-Guided Attention Block (EGAB), which incorporates explicit edge priors into a Transformer architecture. Built upon EGAB, we develop the Edge-Guided Restormer (EG-Restormer) to improve the decoding rate for severely blurred QR codes. For mildly blurred inputs, we introduce a Lightweight and Efficient Network (LENet) that achieves rapid deblurring with minimal computational overhead. To leverage the complementary strengths of both networks, we further integrate EG-Restormer and LENet into an Adaptive Dual-network (ADNet), which dynamically selects the appropriate restoration branch according to the input blur severity of the input, making it particularly suitable for real-time QR code deblurring. Extensive experiments demonstrate the effectiveness of our approaches. Specifically, EG-Restormer achieves state-of-the-art decoding performance, improving the decoding success rate on QRData by 8.67 percentage points under GoPro-only training and 1.33 percentage points after GoPro pre-training followed by QRData fine-tuning. Moreover, ADNet reduces inference latency by 19% while maintaining high decoding accuracy, making it well-suited for real-time QR code deblurring.