【数据科学赛】ICDAR2023 Competitionon: Detecting Tampered Text in Images #文本篡改检测 #￥87000 #阿里云_text-manipulation-classification

CompHub 实时聚合多平台的数据类(Kaggle、天池…)和OJ类(Leetcode、牛客…）比赛。本账号会推送最新的比赛消息，欢迎关注！

更多比赛信息见 CompHub主页

以下内容摘自比赛主页（点击文末阅读原文进入）

Part1赛题介绍

题目

ICDAR2023 Competitionon: Detecting Tampered Text in Images

举办平台

阿里天池

主办方

背景

Texts in images efficiently deliver dense information and become one of the most common mediums for various applications, such as digital finance, electronic commerce, security audit, and qualification review. As texts contain important information, it is crucial that we can prevent text from tampering. In fact, a small change in a sentence might significantly twist the whole carried semantic information. However, most of the previous studies in Document Analysis and Recognition focus on detecting and understanding the content of texts. The authenticity of them is rarely discussed, raising growing concerns about information security in daily life.In recent years, image forensics have received increasing attention from both academia and industry aiming to defend malicious image manipulation. Most of the studies focus on natural images, in which the tampered subjects are usually the objects such as a human or a car. While tampered text detection is more challenging due to the unstructured presentation of the texts. For examples, the tampered areas can be very small (e.g., a character in a paragraph); the contrast between tampered regions and surroundings can be very low. However, most of the previous works are based on private datasets. The lack of open-source dataset also slows down the development of tempered text detection techniques.Therefore, we build a relatively large-scale dataset, namely Tampered Text in Images (TTI) which simulates the electronic commerce scenarios. The images are captured via several sources for diversity. It contains 19,000 text images in total and 15,994 images are manipulated using several types of manipulation techniques, including manual and automatic approaches. Each image is annotated with a binary mask indicating the tampered location. Correspondingly, two tasks are present for this competition: (1) text manipulation classification, (2) text manipulation detection.We hope that the dataset and task could help the research community and promote the research in text manipulation detection in images.

以图像为载体的文本有效地传递了密集的信息，成为数字金融、电子商务、安全审计、资格审查等各种应用的最常用媒介之一。由于文本包含重要信息，因此防止文本被篡改是至关重要的。事实上，句子中的一个微小的变化就可能极大地扭曲整个语义信息。然而，以往文献分析与识别的研究大多集中在文本内容的检测和理解上。它们的真实性很少被讨论，这引起了人们对日常生活中信息安全的日益关注。近年来，图像取证技术越来越受到学术界和业界的关注，旨在防御恶意图像篡改。

大多数研究都集中在自然图像上，其中被篡改的对象通常是人或汽车等物体。而由于文本的非结构化呈现，篡改文本检测更具挑战性。例如，被篡改的区域可以非常小(例如，段落中的一个字符);被篡改的区域和周围环境之间的对比可能非常低。

然而，以前的大部分工作都是基于私有数据集。开源数据集的缺乏也减缓了文本检测技术的发展。因此，我们构建了一个相对大规模的数据集，即模拟电子商务场景的篡改文本图像(Tampered Text in Images, TTI)。图像通过多个来源捕获，以实现多样性。它总共包含19,000个文本图像，15,994个图像使用几种类型的操作技术进行操作，包括手动和自动方法。 每个图像都用二进制掩码标注，表示被篡改的位置。相应的，本次比赛分为两个任务:(1)文本操作分类，(2)文本操作检测。我们希望该数据集和任务能够帮助研究界，促进图像中文本操作检测的研究。

Part2时间安排

• Feb 15, 2023: Start of the contest, training and validation set release

• Mar 15, 2023: Testing set release

• Mar 20, 2023: Testing set prediction submission deadline and Tech report submission

• Mar 25, 2023: Tech report submission deadline

• Mar 31, 2023: Final ranking announcement

Note: All deadlines are at Beijing time (UTC+8) 11:59:00 noon on the corresponding day unless otherwise noted. The competition organizers reserve the right to update the contest timeline if they deem it necessary.

Part3奖励机制

The prize for ICDAR2023 Competition on Detecting Tampered Text in Images is ￥87000 in total, sponsored by Alibaba Group.

• 1st Place: ￥40,000 CNY

• 2nd Place: ￥20,000 CNY

• 3rd Place: ￥10,000 CNY

• 4~6 Place: ￥3,000 CNY

• 7~10 Place: ￥2,000 CNY

Part4赛题描述

To evaluate the tampered text detection performance on various aspects, we introduce two common tasks on TTI dataset, i.e., Manipulation Classification and Detection.

Task 1. Text Manipulation Classification. The aim of this task is to determine whether the text image has been tampered.

此任务的目的是确定文本图像是否已被篡改。

Input: tampered text image or authentic image
Output: The confidence score for the class representing tampered text image.

Task 2. Text Manipulation Detection. The aim of this task is to detect tampered regions using pixel-level binary masks.

该任务的目的是使用像素级二进制掩码检测被篡改的区域。

Input: tampered text image or authentic image
Output: predicted masks for all test images, where the tampered pixels are 1 (white) and other pixels are 0 (black). The shape of mask should be equal to shape of corresponding image.