# PAN++: Towards Efficient and Accurate End-to-End Spotting of Arbitrarily-Shaped Text

**Authors:** Wenhai Wang, Enze Xie, Xiang Li, Xuebo Liu, Ding Liang, Zhibo Yang, Tong Lu, Chunhua Shen  
**arXiv:** [2105.00405](https://arxiv.org/abs/2105.00405)  
**Date:** 2021-05-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.00405v1/

## Key points

- Scene text detection and recognition have advanced, but spotting arbitrarily-shaped text remains a challenge
- PAN++ framework redefines text line as central text kernel with peripheral pixels
- Kernel representation accurately describes arbitrary text and distinguishes adjacent text
- Pixel-based representation allows for real-time prediction by single fully convolutional network
- Components of PAN++ include FPEMs for feature enhancement, PA for lightweight detection head, and attention-based recognition head with Masked RoI
- PAN++ introduces major extensions in text recognition module and overall end-to-end text spotting framework compared to previous versions like PSENet and PAN
- Extensive experiments on benchmark datasets show effectiveness of PAN++
- Achieves high performance in speed and accuracy across various benchmarks

## Summary

Scene text detection and recognition have seen significant advancements in recent years, but the efficient and accurate end-to-end spotting of arbitrarily-shaped text remains a challenge. In response to this, a new framework called PAN++ has been proposed. This  framework is based on a  that redefines a text line as a central text kernel surrounded by peripheral pixels. Through systematic comparisons with existing scene text representations, it has been shown that the kernel representation not only accurately describes arbitrarily-shaped text but also effectively distinguishes adjacent text. One of the key features of PAN++ is its pixel-based representation, which allows for prediction by a single fully convolutional network, making it suitable for real-time applications. The framework includes several components designed to enhance performance: a feature enhancement network consisting of stacked Feature Pyramid Enhancement Modules (FPEMs), a lightweight detection head with Pixel Aggregation (PA), and an attention-based recognition head with Masked RoI. Compared to previous versions such as PSENet and PAN, PAN++ introduces major extensions in the text recognition module and the overall end-to-end text spotting framework. The architecture has been revamped to integrate a tailored feature extractor (Masked RoI) and a lightweight text recognition head. Additionally, improvements have been made to the text detection module through systematic comparisons with other existing representations, simplification of FPEM into a more effective module, and enhancing PA to be aware of background elements. Extensive experiments conducted on challenging benchmark datasets such as Total-Text, CTW1500, ICDAR 2015, and MSRA-TD500 demonstrate the effectiveness of PAN++. On the Total-Text dataset, PAN++ achieves an end-to-end text spotting F-measure of 68.6%, outperforming previous methods like ABCNet while maintaining faster inference speeds. Furthermore, it achieves competitive results on other benchmarks including multi-oriented and long-text datasets. In summary,  offers an efficient and accurate solution for  of arbitrarily-shaped text in natural scenes. Its innovative kernel representation and tailored components contribute to high performance in both speed and accuracy across various benchmark datasets.