LiteOCR: A Lightweight Text Detection Framework Integrating Cross-Modal Semantic Distillation and Dynamic Resolution Decision-Making
TL;DR
LiteOCR is introduced—a lightweight OCR framework designed for edge intelligence deployment that comprehensively validates the effectiveness of cross-modal attention distillation in preserving semantic accuracy, dynamic resolution gating in enhancing computational efficiency, and semi-supervised feature alignment in reducing data dependency, providing a practical lightweight solution for high-precision OCR deployment at the edge.
Abstract
To address the urgent demand for efficient and robust optical character recognition (OCR) technology in the e-commerce sector, this paper introduces LiteOCR—a lightweight OCR framework designed for edge intelligence deployment. This framework integrates three core technologies: cross-modal knowledge distillation, semi-supervised learning mechanisms, and dynamic computation gating strategies. The framework aims to effectively overcome the significant limitations of traditional OCR systems, which suffer from insufficient robustness in complex environments and the difficulty of running resource-intensive large models efficiently on computationally constrained devices. Its core design principle focuses on achieving an optimal balance between adaptive capabilities for complex and variable visual conditions in e-commerce scenarios and real-time processing performance on edge devices, providing an effective solution for efficient text information extraction at the edge. Experiments demonstrate that on a custom-built e-commerce dataset, LiteOCR achieves an F1 score of 83.3%. Despite having only 15% of the model parameters compared to its teacher model (GOT-OCR 2.0), the performance gap narrows to just 1.8 percentage points. This comprehensively validates the effectiveness of cross-modal attention distillation in preserving semantic accuracy, dynamic resolution gating in enhancing computational efficiency, and semi-supervised feature alignment in reducing data dependency, providing a practical lightweight solution for high-precision OCR deployment at the edge.