Research on Text Recognition of Ancient Books Based on Improved Lightweight CRNN: Integrating MobileNetV3 and Hybrid Attention Enhanced Architecture

Abstract

The traditional Convolutional Recurrent Neural Network (CRNN) architecture has demonstrated exceptional performance in text
sequence recognition, extracting visual features through convolutional layers, modeling sequence dependencies with recurrent layers, and
finally outputting results via transcription layers. However, for the specific task of ancient text recognition, existing models face multiple chal
lenges: ancient book images often suffer from severe paper degradation, blurred handwriting, and complex background interference; the com
plex structure of Chinese characters and the plethora of variant forms require a high level of precision in feature extraction; furthermore, the
massive demand for the digitization of ancient texts necessitates recognition systems with high inference speed to accommodate large-scale
deployment. To balance recognition accuracy and computational efficiency, this study proposes an improved lightweight CRNN architecture.
This architecture uses MobileNet V3 as the backbone network, enhances spatial feature acquisition through the integration of the Coordinate
Attention (CA) mechanism, and optimizes global relationship modeling with a "sandbox" structure composed of Bidirectional Gated Recur
rent Units (BiGRU) and self-attention encoders, achieving efficient and precise recognition of ancient texts.