Multi-lingual Scene Text Detection Containing the Arabic Scripts Using an Optimal then Enhanced YOLO Model

In the past few years, notable progress has been achieved in the field of deep learning, particularly in the realm of identifying text within images of natural scenes, owing to the advancements in machine learning and artificial intelligence. The effectiveness of deep learning and text detection in the wild, especially when dealing with Arabic language, is frequently hindered by the scarcity of diverse datasets encompassing multiple languages and scripts, which poses an additional challenge. Despite significant advancements, this shortage continues to be a limiting factor. The YOLO (You Only Look Once) deep learning neural network has gained widespread popularity for its adaptability in tackling various machine learning tasks, notably in the field of computer vision. The YOLO algorithm has garnered growing recognition for its remarkable capability to address intricate issues when dealing with images taken in natural environments, managing noisy data, and surmounting the diverse challenges encountered in the wild. Our experiments provide a concise evaluation of text detection algorithms centered around convolutional neural networks (CNNs). Specifically, we concentrate on different versions of the YOLO models, applying identical data augmentation methods to both the SYPHAX dataset and the ICDAR MLT-2019 dataset, both of which encompass Arabic scripts within images of natural scenes. The objective of this article is to pinpoint the most efficient YOLO algorithm for recognizing Arabic script in the wild, and subsequently, to improve upon the best-performing model. Additionally, we aim to investigate potential research directions that can further enhance the capabilities of the most robust architecture in this domain.