Overview - ICDAR 2023 Competition on Reading the Seal Title

Overview

Based on the flourish of deep learning method, we have witnessed the maturity of regular and general OCR technology, including scene text detection and recognition. However, as a common element which can be seen everywhere in official and financial scenarios, seal title text has not gain its attention. And the task of reading seal title text is also faced with many challenges, such as variable shapes of seal (for example, circle, ellipse, triangle and rectangle), curved text, background noise and overlapped text, as shown in Figure 1- Figure 3. In order to promote the research of seal text, we propose the competition on reading the seal title.

Considering there are no existing datasets for seal title text reading. We construct a dataset including 10,000 real seal data, which covers the most common classes of seal. In the dataset, all seal title texts are labeled with text polygons and text contents. Besides, two tasks are presents for this competition: (1) Seal title text detection; (2) End-to-end seal title recognition. We hope that the dataset and tasks could greatly promote the research in seal text reading.

Various kinds of shape:

The shape of seals can vary from ellipse, rectangle to triangle:

 

Figure 1. Different shapes of seals

Curved Text:

Curved Text is widely distributed in the seal datasets:

Figure 2. Seals with curved texts

Overlapped Text:

Most seals in daily occasion have overlapped texts:

Figure 3. Seals with overlapped texts

 

Note: The competition report can be found at here.