다중 패치를 이용한 트랜스포머 기반 장면 텍스트 인식

Abstract

In this paper, we explore the application of Vision transformer (ViT) to the scene text recognition task. As a popular research direction in computer vision, Scene text recognition enables computers to recognize or read the text in natural scenes, such as object labels, text descriptions, and road text signs. At present, the traditional convolutional neural network-based model has better performance. Still, in the face of complex backgrounds and irregular scene text pictures, the performance of the convolutional neural network-based model is challenging to improve in curved text, diverse fonts, distortions, etc. With the application of transformers in computer vision, the model structure based on transformers has also significantly been developed. Although the current transformer-based model can obtain the performance of the model structure similar to CNN, it is currently in the early stage of application, and there is much room for research and improvement. We propose a multi-scale vertical rectangular patch model (MSVSTR) for transformer-based feature extractor to be more suitable for text images. By only arranging the patches in a single direction, when the image is cropped through the patch, it can be more suitable for the distribution form of the text in the text image. At the same time, to be suitable for different numbers of characters in other texts and more robust feature extraction, vertical rectangular patches of different scales are applied to crop the image. Our structure performs better through various ablation experiments than similar transformer-based STR models. At the same time, experiments show that our structure can perform seven benchmarks well.