Lightweight Deep Learning Model and Novel Dataset for Restoring Damaged Barcodes and QR Codes in Logistics Applications

This study introduces a lightweight deep learning model and a novel synthetic dataset designed to restore damaged one-dimensional (1D) barcodes and Quick Response (QR) codes, addressing critical challenges in logistics operations. The proposed solution leverages an efficient Pix2Pix-based framework, a type of conditional Generative Adversarial Network (GAN) optimized for image-to-image translation tasks, enabling the recovery of degraded barcodes and QR codes with minimal computational overhead. A core contribution of this work is the development of a synthetic dataset that simulates realistic damage scenarios frequently encountered in logistics environments, such as low contrast, misalignment, physical wear, and environmental interference. By training on this diverse and realistic dataset, the model demonstrates exceptional performance in restoring readability and decoding accuracy. The lightweight architecture, featuring a U-Net-based encoder-decoder with separable convolutions, ensures computational efficiency, making the approach suitable for real-time deployment on embedded and resource-constrained devices commonly used in logistics systems. Experimental results reveal significant improvements: QR code decoding ratios increased from 14% to 99% on training data and from 15% to 68% on validation data, while 1D barcode decoding ratios improved from 7% to 73% on training data and from 9% to 44% on validation data. By providing a robust, resource-efficient solution for restoring damaged barcodes and QR codes, this study offers practical advancements for enhancing the reliability of automated scanning systems in logistics operations, particularly under challenging conditions.