BARGAIN: A Super-Resolution Technique to Gain High-Resolution Images for Barcodes

Barcode scanners are common sights in every retail store. These scanners are used to detect and recognize barcodes present on each item during checkout by customers. Store associates and sometimes robots are employed to take pictures of the inventory present in a store; the barcodes that appear on the packages of the items are used for inventory management. However, these tasks are sometimes made difficult by distortions in the barcode images, e.g., bad lighting, images captured from large distances, poor resolution cameras, etc. In order to alleviate these problems, we propose a super-resolution technique that is specifically tuned to convert low-resolution barcode images into high-resolution ones; these high-resolution images are subsequently sent for the downstream barcode detection and recognition steps. We propose a new loss function for training an improved deep neural network architecture that recovers barcodes with sharp boundaries in the super-resolved images. Extensive experiments show that our model achieves better accuracy (around 11.27% improvement in decoding accuracy) and visual improvements against state-of-the-art methods in terms of barcode recognition accuracy.