SRRNet: A Transformer Structure With Adaptive 2-D Spatial Attention Mechanism for Cell Phone-Captured Shopping Receipt Recognition

Shopping receipts, which are regarded as a kind of consumption proof provided to consumers, contain important information for trade. The digitalization of shopping receipts by extracting text information from images can provide smart retail with precise data analysis for commodity management and supply chain optimization. Despite the fact that traditional optical character recognition (OCR) systems have performed well on document template-based recognition, accurate recognition of receipts taken by cell-phones remains difficult due to the uncertainty of the shooting environment. To address irregular text recognition for receipts, in this research we propose a transformer-based text recognition network model by developing an adaptive 2D spatial attention module to extract the 2D correlation information of image features. We examined the performance of our proposed model on both public benchmarks and a large-scale real-world shopping receipt text recognition dataset. Results demonstrate the efficacy of the proposed method in comparison to extant approaches.