Extraction and Recognition of Container Code Characters

Lau, Lok Jing (2022) Extraction and Recognition of Container Code Characters. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Lau Lok Jing.pdf Restricted to Registered users only Download (18MB)

Abstract

The need for an industrial competent Automatic Container Code Recognition (ACCR) is never as high before due to the revolution of container shipping. The productivity of internal trades relies on an efficient and effective systematic container management plan, and the increase in that has resulted in the demand for an ACCR system. In the recent development of deep learning, the emerged deep learning models not only addressed the interferences of irrelevant words in natural scenes that are highly challenging in traditional approaches but resulted in a simpler yet elegant pipeline. This research proposed an end-to-end, real-time deployable ACCR system that runs at an inference speed of 31.72 frames per second (fps) with an end-to-end accuracy of 89.20%. The ACCR system is composed of localization and recognition modules that implement Differential Binarization (DB) and Semantic Reasoning Network (SRN) respectively to address the Container Code Characters (CCCs) in Scene Text Recognition (STR). Besides, several other recently published neural networks such as Single-Shot Arbitrarily-Shaped Text Detector (SAST) and Convolutional RecurrentNeuralNetwork (CRNN) with Connectionist Temporal Classification (CTC) been studied and investigated to identify their inherent strength and weakness in addressing STR problems. For on-site integration, the research proposed an off-site machine and servers integration with Firebase Cloud storage for achieving high-computational power.