Revolutionizing Scene Text Recognition: Unleashing the Power of Dual Step Attention Mechanism in the Encoder

Text recognition in natural scenes presents a significant challenge in computer vision despite its widespread applications in real-life scenarios. Deep learning advancements have notably enhanced the precision of scene text recognition. The prevailing framework for scene text recognition currently revolves around the encoder–decoder architecture, with Connectionist Temporal Classification (CTC) and attention mechanisms serving as popular text recognition techniques. However, existing methods rooted in the encoder–decoder architecture primarily integrate the attention mechanism into the decoder, limiting their capacity to identify characters accurately. Conversely, CTC-based methods exhibit faster processing times and impose fewer computational burdens, yet they are comparatively less effective than their attention-based counterparts. We introduce a new text recognition network based on a transformer-based network to address these limitations. Our proposed network introduces the Dual Step Attention Mechanism (DSAM), wherein the encoder's attention mechanism captures the overall text location within the image. In contrast, the decoder's attention mechanism precisely identifies the location of each character in the word image. In addition, we have employed a bidirectional encoding to capture the capture character sequence in both directions. To validate the efficacy of our approach, we conducted a comprehensive performance evaluation comparing our proposed network with state-of-the-art methods across six diverse text datasets, namely IIIT5K, SVT, IC13, IC15, SVTP, and CUTE. The results indicate that our model outperforms existing approaches across various metrics.