SEASTR: Spatial-Enhanced Attention Scene Text Recognition

Scene text recognition has emerged as a significant research area in computer vision, and Transformer Decoder-Encoder-based architecture has gained widespread popularity. However, these methods are often complex and require larger models. In our research, we discovered that a simple modification to the attention method in the decoder can lead to significant improvements and facilitate simple and efficient experimental results. Specifically, we introduce a novel Spatial-Enhanced attention scene text recognition network (SEASTR), which employs a spatial attention module to improve the current attention-based decoding module. This module enables better focus on the target area at each decoding time step by strengthening the connection between feature spaces, thus reducing recognition errors due to attention drift. The spatial attention module we propose is both simple and highly effective. Comparative experiments revealed that the model equipped with the spatial attention module performed well, delivering excellent results across multiple public datasets. Notably, on regular datasets IIIT5K, SVT, and IC13, the addition of the spatial attention module enhanced the accuracy of the model by 5.1 %, 2.1%, and 1.6%, respectively. On the irregular datasets SVTP and CUTE80, the accuracy improved by 1.1% and 10.3%, respectively. The SEASTR model fine-tuned with real datasets improved by 7.4%, 6.8%, 2.4%, 6.3%, 7.2%, and 16.6% on the IIIT5K, SVT, IC13, IC15, SVTP, and CUTE80, respectively.