Multilingual Text Recognition using Advanced Deep Learning Techniques

Abstract: Reading licence plates, recognising street signs, retrieving photographs, doing mobile visual searches, identifying image-based relocation, and aiding visually impaired individuals are just a few of the many uses of text recognition in natural scene images in computer vision systems. The proliferation of digital multiscript documents, photos, and videos has made the detection and identification of text in multilingual scenes an increasingly difficult undertaking. By creating methods for word detection from real-world scene pictures, this study hopes to solve the current problem of scene text extraction. There are a lot of obstacles for texts in natural photos, such as images with wavy messages, images with rotated texts and orientation, images with noise, images with a diversity of typefaces, and photographs with wild backgrounds. By using atrous separable convolution, the improved system, light-YOLOv3, outperforms the original YOLOv3 in terms of accuracy and agility. One model for language identification that uses neural networks is FastText, which can learn several language representations of text automatically. This paper introduces many methods using CNN to enhance text detection. Our study's results show that deep learning techniques and embedding visualisation are effective in multilingual text classification, and we achieved this by using light-YOLOv3, which replaced standard convolution with atrous separable convolution. This reduced the volume of computations and the number of parameters for multi-lingual scene text detection