Abstract:
Defective Characters exist frequently and broadly in images such as license plates, electricity, water meters, street boards, etc.
Thus, building robust recognition systems or enhancing the accuracy and robustness of the existing recognition systems to recognize such
characters on images is a challenging research topic in image processing and computer vision. This paper Investigates and adopts ReId
dataset for all the experimental work and introduces two deep learning models (CNN5-BLSTM and CNN7-GRU) based on convolutional
recurrent neural networks (CRNN) to address the problem of defective characters sequence recognition. The two proposed deep learning
models are segmentation-free, lightweight, End-To-End trainable, and slightly different from each other. The models are evaluated on
testing data of ReId dataset, and the achieved accuracies are 95% of characters’ sequence accuracy and 98% of character-level accuracy.
Moreover, their performance on ReId dataset outperforms other models’ performance in the literature.
