[Paper Review] Deep Dual Pyramid Network for Barcode Segmentation using Barcode-30k Database

Qijie Zhao, Feng Ni+3 more|arXiv (Cornell University)|Jul 31, 2018

Vehicle License Plate Recognition24 references3 citations

TL;DR

This paper proposes Barcode-30k, a large-scale synthetic dataset of 30,000 barcode and QR code images with pixel-level annotations, and introduces BarcodeNet, a deep dual pyramid segmentation network using Prior Pyramid Pooling (P3M) and Pyramid Refine Module (PRM) to improve boundary accuracy and semantic understanding. The model achieves 95.36% mIoU on the validation set and generalizes well to real-world images.

ABSTRACT

Digital signs(such as barcode or QR code) are widely used in our daily life, and for many applications, we need to localize them on images. However, difficult cases such as targets with small scales, half-occlusion, shape deformation and large illumination changes cause challenges for conventional methods. In this paper, we address this problem by producing a large-scale dataset and adopting a deep learning based semantic segmentation approach. Specifically, a synthesizing method was proposed to generate well-annotated images containing barcode and QR code labels, which contributes to largely decrease the annotation time. Through the synthesis strategy, we introduce a dataset that contains 30000 images with Barcode and QR code - Barcode-30k. Moreover, we further propose a dual pyramid structure based segmentation network - BarcodeNet, which is mainly formed with two novel modules, Prior Pyramid Pooling Module(P3M) and Pyramid Refine Module(PRM). We validate the effectiveness of BarcodeNet on the proposed synthetic dataset, and it yields the result of mIoU accuracy 95.36\% on validation set. Additional segmentation results of real images have shown that accurate segmentation performance is achieved.

Motivation & Objective

• To address the lack of large-scale, well-annotated datasets for barcode and QR code segmentation under real-world conditions such as small scale, occlusion, and illumination variation.
• To develop a deep learning-based semantic segmentation model tailored for digital signs with regular shapes and structural texture, which existing models fail to handle effectively.
• To reduce annotation cost through a novel image synthesis strategy that generates realistic barcode and QR code images with precise pixel-level labels.
• To improve segmentation accuracy—especially at object boundaries—by integrating both global semantic context and shallow structural features via a dual pyramid architecture.

Proposed method

• A data synthesis pipeline generates 30,000 realistic barcode and QR code images by overlaying real barcode/QR code templates onto diverse background images with controlled variations in scale, rotation, cropping, and lighting.
• The Prior Pyramid Pooling Module (P3M) enhances feature learning by incorporating global prior information to guide local location distribution, particularly benefiting large targets.
• The Pyramid Refine Module (PRM) refines segmentation boundaries by aggregating and refining shallow features from early convolutional layers, improving localization accuracy.
• BarcodeNet is built on a PSPNet-like backbone but extends it with P3M and PRM modules to jointly capture high-level semantics and low-level structural details.
• The network is trained end-to-end on the synthetic Barcode-30k dataset using cross-entropy and dice loss, with data augmentation to improve robustness.
• Model generalization is evaluated on real-world images collected from office environments, demonstrating strong zero-shot performance.

Experimental results

Research questions

• RQ1 Can a large-scale, synthetic dataset with precise pixel-level annotations significantly reduce annotation cost while maintaining real-world generalization for barcode segmentation?
• RQ2 How can a deep learning model effectively leverage both global semantic context and local structural features to improve segmentation of regular-shaped digital signs?
• RQ3 To what extent does integrating shallow features through a dedicated refinement module improve boundary accuracy in barcode and QR code segmentation?
• RQ4 Does a dual pyramid architecture—combining global prior guidance and multi-scale feature refinement—outperform standard semantic segmentation networks on this specialized task?

Key findings

• BarcodeNet achieves a mean Intersection over Union (mIoU) of 95.36% on the Barcode-30k validation set, significantly outperforming baseline models like FCN-8s and standard PSPNet.
• The Pyramid Refine Module (PRM) using two convolutional layers to refine features from Res1 (the shallowest layer) yields the best performance, achieving 91.63% mIoU, demonstrating the importance of shallow feature integration.
• Visual comparisons show that BarcodeNet effectively segments small, overlapping, and partially occluded barcodes, while reducing boundary blurring and missing detections common in baseline models.
• When applied to real-world images, BarcodeNet pre-trained on the synthetic Barcode-30k dataset produces highly accurate segmentation results, even in challenging conditions like truncation and small object scales.
• The synthesis strategy enables efficient creation of diverse, realistic training data with minimal manual annotation, supporting time-critical segmentation tasks.
• The method and model are generalizable to other regular-shaped targets such as logos, traffic signs, and billboards, suggesting broader applicability beyond barcodes and QR codes.