Transforming Scene Text Detection and Recognition: A Multi-Scale End-to-End Approach With Transformer Framework

Text is an essential means for humans to acquire information and engage in social communication. Accurate text extraction from images is crucial for various tasks in real-life scenarios and scene understanding. However, text detection and recognition in natural scenes are challenged by noise in the images, irregular distribution of text fonts, and degradation of image quality under complex acquisition conditions. These factors severely impact the accuracy of text recognition. Issues such as poor image quality, diverse text formats, and complex image backgrounds significantly affect the accuracy of the recognition, and these challenges remain urgent to be addressed in the field. To address these challenges, this paper proposes a transformer-based scene image text detection and recognition algorithm within a multi-scale end-to-end framework. Firstly, by integrating detection and recognition stages into an end-to-end framework, the process is simplified, reducing computation and errors. Subsequently, multi-scale characteristics are incorporated to effectively capture text information at various scales, enhancing recognition accuracy and robustness through feature fusion and anti-interference capability. Lastly, leveraging the transformer framework, the algorithm efficiently handles text information of different scales and positions, improving generalization ability. The self-attention mechanism, multi-layer stacking structure, and positional encoding in the transformer framework contribute to its effectiveness in processing diverse text information. Through validation, the proposed method demonstrates improved efficiency in scene text detection and recognition.