Efficient End-to-End Deep Learning Model for Algerian Drug Label Recognition

Optical Character Recognition (OCR) is still a very active area of computer vision research. This work proposes a robust text detection/recognition framework for Algerian drug labels (ADL). It extracts the text from the drug label using the benchmark YOLOv8 object detector. Text recognition is done through our CRNN-gMLP model combining a convolutional recurrent neural network (CRNN) and a gate multi-layer perceptron (gMLP) network. The gMLP detects the spatial interaction between the target characters, which helps CRNN learn a more precise alignment at each decoding step. Our end-to-end YOLO-CRNN-gMLP framework locates label text and then recognizes words by converting text pixels into sequences of characters. To evaluate our approach, we created a new dataset specifically for ADL, which contains texts in various sizes and fonts for training and testing purposes. The experiment results showed that CRNN-gMLP achieved an acceptable text recognition rate even on the least readable labels.