University of British Columbia
Spotting shipping container codes with deep learning methods
Abstract
dc:descriptionAutomated container code reading systems are of great value to logistics management. However, previous container code spotting methods have limited recognition performance in complex outdoor environments of transportation industry. Therefore, this research focuses on designing high-performance shipping container code spotting (including detection and recognition) methods for automated container code reading. First, an adaptive deep learning framework is proposed for horizontal code spotting. Adaptive threshold and average score aggregation (ASA) are designed to remove the noisy text. The detected code regions are adaptively adjusted for code recognition in our framework. Experiments have been conducted on real horizontal code image datasets. The overall recognition accuracy has been improved from 90.17% to 93.33%. Second, a deep learning model is proposed for vertical code spotting. It aligns vertical features of container codes. In addition, it includes two recognition pipelines and an uncertainty and probability-based decision strategy to adaptively choose the final results from two pipelines. Experiments conducted on real vertical code datasets demonstrate that the overall recognition f1-score performance has been improved from 93.5% to 96.5%. Third, a visual and textual information fusion-based zero-shot deep learning framework is proposed to spot placard codes. In this framework, Logarithmic Weighted Cross-Entropy (LWCE) is designed to reduce the effects of imbalanced classes and Logarithmic Weighted Confidence Fusion (LWCF) is proposed to fuse the visual and textual information to handle unseen classes and enhance performance in seen classes. Our proposed framework has been validated on real industrial placard datasets. The overall end-to-end recognition f1-score performance has been improved from 79.97% to 92.77%. Overall, our deep learning-based code spotting methods improve the accuracy of automated container code recognition and facilitate automated container management.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Electrical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zhang, Ran
- Contributors dc:contributor
-
- Liu, Zheng (Professor of engineering)
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/83394