# PIMNet: A Parallel, Iterative and Mimicking Network for Scene Text Recognition

**Authors:** Zhi Qiao, Yu Zhou, Jin Wei, Wei Wang, Yuan Zhang, Ning Jiang, Hongbin Wang, Weiping Wang  
**arXiv:** [2109.04145](https://arxiv.org/abs/2109.04145)  
**Date:** 2021-09-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2109.04145v1/

## Key points

- The field of [redacted] has seen a surge in interest due to its wide range of applications
- Advanced methods utilize autoregressive models with attention mechanisms for sequential text generation
- Non-autoregressive models offer faster inference times but sacrifice accuracy compared to autoregressive models
- PIMNet introduces a novel approach that leverages parallel attention mechanism and iterative generation for balancing speed and precision
- PIMNet uses an additional autoregressive decoder during training alongside the parallel decoder for improved accuracy without pre-training requirement
- Extensive experiments demonstrate the effectiveness and efficiency of PIMNet in achieving competitive performance with fast inference times
- Code for PIMNet is available at https://github.com/Pay20Y/PIMNet for further exploration and implementation

## Summary

The field of  has seen a surge in interest in recent years due to its wide range of applications. Many advanced methods currently utilize an  with  to generate text sequentially from left to right. While these approaches have demonstrated impressive performance, their speed is often hindered by the one-by-one decoding strategy they employ. On the other hand, non-autoregressive models offer faster inference times by predicting results in parallel; however, they tend to sacrifice accuracy compared to their autoregressive counterparts. To address this trade-off between accuracy and efficiency, this paper introduces a novel approach called the . PIMNet leverages a parallel attention mechanism for quicker text prediction and an iterative generation mechanism to enhance the accuracy of predictions. By fully exploring context information in each iteration, PIMNet aims to strike a balance between speed and precision. A key innovation of PIMNet lies in its use of  during training. This involves incorporating an additional autoregressive decoder alongside the parallel decoder, where the latter mimics the autoregressive decoder by aligning its outputs with those of the hidden layer. With a shared backbone between the two decoders, PIMNet can be trained end-to-end without requiring pre-training. During inference, the branch associated with the autoregressive decoder is removed to further boost speed. Extensive experiments conducted on public benchmarks demonstrate the effectiveness and efficiency of PIMNet in achieving competitive performance while maintaining fast inference times. The authors have made their code available at https://github.com/Pay20Y/PIMNet for further exploration and implementation by interested parties. This innovative network architecture holds promise for advancing  technology and opening up new possibilities for real-world applications.