Extra-Feature Injected Reservoir Computing for Multi-Language Handwritten Digit Recognition Based on 2D Optoelectronic Memristor

The ability to extract and accurately distinguish multiple types of information from images is critical in addressing the increasing demands of data processing. Reservoir computing (RC) has emerged as a promising candidate due to its low training costs and rich reservoir states. Most previous studies mainly focus on the classification task for one type of data with a middle sample scale. Consequently, there is limited work on RC that can train with large samples and distinguish multiple features, the demand for high recognition accuracy requires to make innovative attempts at the RC model architecture. In this study, extra-feature injected RC based 2D memristor is demonstrated, enabling high-precision dual-function collaborative recognition of language and digits with large sample scale. The sensitive responsiveness of the device to photoelectric signals originates from the sulfur vacancies in rhenium disulfide. By using the RC model for handwritten digit recognition and the MIX-MNIST data set for classification, the innovative RC model improves recognition accuracy of language and numbers by ≈5% compared with traditional artificial neural networks, which are extremely tough with large sample data. This work provides insights into the innovative RC architecture to the recognition of complex image information.