New Benchmarks for Barcode Detection Using Both Synthetic and Real Data

Abstract

Given the wide use of barcodes, there is a growing demand for their efficient detection and recognition. However, the existing publicly available datasets are insufficient and of poor quality. Moreover, recently proposed approaches were trained on different private datasets, which makes the comparison of proposed methods even more unfair. In this paper, we propose a simple yet efficient technique to generate realistic datasets for barcode detection problem. Using the proposed method, we synthesized a dataset of $\sim$30,000 barcodes that closely resembles real barcode data distribution in terms of size, location, and number of barcodes on an image. The dataset contains a large number of different barcode types (Code128, EAN13, DataMatrix, Aztec, QR, and many more). We also provide a new real test dataset of 921 images, containing both document scans and in-the-wild photos, which is much more challenging and diverse compared to existing benchmarks. These new datasets allow a fairer comparison of existing barcode detection approaches. We benchmarked several deep learning techniques on our datasets and discuss the results. Our code and datasets are available at https://github.com/abbyy/barcode_detection_benchmark.