Efficient ultra-lightweight convolutional attention network for embedded identity document recognition system

Abstract

With the rapid development of IoT, identity document recognition has been widely applied in various fields. Efficient recognition systems are crucial for deployment on resource-constrained embedded devices, but many deep learning models suffer from high computational complexity. We propose an efficient character recognition system with a two-stage framework: a document number detection network and an ultra-lightweight attention-based recognition network named EULCAN (Efficient Ultra-Lightweight Convolutional Attention Network). EULCAN's feature extraction module employs a novel Dense Simplified Convolutional Attention Module (DSCAM) and a Dual Dimensionality Reduction Block (DDRB) to capture discriminative features efficiently. DSCAM combines an Efficient Bottleneck Convolution Block and a Simplified Channel Attention Block, significantly reducing computational costs while maintaining accuracy. For sequence transcription, a simple fully connected layer coupled with a Connectionist Temporal Classification (CTC) layer is used for robust recognition. Evaluated on the BDCI benchmark and a real-world SUST dataset, EULCAN achieves competitive accuracies of 97.1% and 95.3%, respectively, while maintaining only 2.8 M parameters and 0.497 GFLOPs. Compared to MobileNetV3, the second most lightweight deployment-ready model, EULCAN improves accuracy by 11.7%, while its parameter size is only 0.6% of OmniParser, the most accurate model. Furthermore, the proposed identity document recognition system has been successfully deployed in real-world scenarios. On the RK3588S2 development board, EULCAN achieves an impressive inference speed of 65 FPS, demonstrating its practicality for embedded IoT applications. The source code is publicly available at https://github.com/ymxb1/EULCAN.