Improved CTPN and CRNN for blind reading based on attention mechanism and feature fusion

The primary function of a blind reader is to convert visual information into speech, thereby enabling blind individuals to access printed or electronic text content. Current blind reader technologies predominantly rely on online detection and recognition, which necessitates the use of cloud platforms. In this research, we present a blind reader technology that enhances the CTPN detection and CRNN identification algorithms, both of which are implemented on the Raspberry Pi 4B development board. This system is capable of performing text detection and recognition tasks in a completely offline environment. On the ICDAR2015 dataset, the enhanced CTPN model outperformed the original CTPN model in terms of accuracy, recall, and F-value, with improvements of 8.0%, 6.7%, and 10.1%, respectively. Similarly, on the ICDAR2017 RCTW dataset, the upgraded CRNN model surpassed the original CRNN model by 5%. The experimental results indicate that the Raspberry Pi-based blind reader technology proposed in this research can effectively assist blind individuals in reading books, achieving reasonably high levels of detection and recognition accuracy.