FoveaBox with Consistent Label Assignment Distillation for Page Object Detection

With the increasing prevalence of digital documents replacing traditional paper documents, there is an urgent need to develop techniques to extract information from them. In particular, image-based documents pose a significant challenge in object detection, where the identification of objects such a...

Full description

Saved in:
Bibliographic Details
Main Authors: Nguyen, Truong, Le, Hai, Vo, Nguyen D.
Format: Conference Proceeding
Language:English
Subjects:
Data mining
Deep Learning
Document Object Detection
FoveaLAD
Knowledge Distillation
Object detection
Object recognition
Vietnamese Document Images
Visual systems
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:With the increasing prevalence of digital documents replacing traditional paper documents, there is an urgent need to develop techniques to extract information from them. In particular, image-based documents pose a significant challenge in object detection, where the identification of objects such as figures, tables, and captions can be difficult. The crux of object detection lies in the accuracy of label assignment, and we all know that the human visual system is highly capable and proficient in recognizing objects. In light of this, we propose a novel approach, FoveaLAD, which integrates FoveaBox drawed inspiration from the human eye, and Label Assignment Distillation to enhance the label assignment process. We evaluate FoveaLAD on the UIT-DODV-Ext image document dataset, which comprises three classes: Table, Figure, and Caption. Our findings demonstrate that FoveaLAD effectively selects different scales for objects of varying sizes and improves the accuracy of label assignment in object detection.
ISSN:2770-6850
DOI:10.1109/MAPR59823.2023.10288889