多語言場景文字偵測與辨識機制

Texts appearing in pictures are usually regions of interest, providing reliable and abundant information related to the images. Extracting such text information from images can thus facilitate many potential applications, such as instant translation, scene understanding, development of smart cities, robots, auto-pilots, and assisting vision-impaired people. Scene text analysis is thus one of the research focuses in the fields of pattern recognition and computer vision. Several challenges of scene text detection and recognition exist, including texts of various forms, multilingual texts, slanted texts, different sizes, occlusion, text-like outliers, lighting, shadows, etc. Deep learning is the mainstream methodology of scene text analysis to achieve reasonably high performance for future applications. This research utilizes deep learning techniques for developing multilingual scene text detection and recognition schemes. Based on our experience of holding “the AICUP Traditional Chinese Scene Text Recognition Competition,” we found that character detection is quite useful in Traditional Chinese scene text recognition and language determination. We thus develop a character detection model, which is trained via weakly supervised learning on string-labeled data. The experimental results show that the proposed scheme works well in detecting English letters, numbers, and Chinese characters, and outperforms existing methods when dealing with texts with multiple orientations. The methodology of character detection also enables lightweight character recognition models in scene text analysis tasks.