Category-aware deep perspective transformation research

Read full text Summarise with VIRAT
Authors: Xiaofei Ji, Sixu Zhao, Jiangtao Cao
Cited by: 0
References: 47

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The perspective transform is widely applied in computer vision tasks, including document rectification, scene text reshaping, and 3D reconstruction. However, existing perspective correction methods typically treat corner localization and geometric transformation as independent problems, while the role of semantic understanding in guiding the correction process is largely overlooked. To address this limitation, a novel framework termed the Category-Aware Deep Perspective Transformation Network (CA-DPTN) is proposed, in which content-aware semantic understanding and geometric correction are effectively integrated into a unified paradigm to enhance both the accuracy and robustness of perspective transformations.A geometric pseudo-label generator is first designed to construct semantically guided perspective categories. Subsequently, a multi-head corner prediction module is introduced to achieve precise boundary localization while maintaining category-aware representation. Finally, an adaptive fusion mechanism is incorporated to dynamically regulate the influence of category information on the estimated transformation parameters.To evaluate the performance of the proposed framework on content-aware perspective correction tasks, a multi-scenario, multi-category in-house dataset is constructed. Furthermore, systematic evaluations are conducted on the public MIDV-500 benchmark dataset to assess the generalization capability of the model. Additional experiments are carried out on the DocUNet and DIR300 document dewarping datasets to further investigate performance boundaries and cross-task generalization ability. Experimental results demonstrate that CA-DPTN achieves favorable overall performance in terms of corner localization accuracy, inference efficiency, and adaptability to complex scenes, validating its effectiveness and general applicability in tasks requiring the integration of semantic understanding and geometric modeling.

Fields of Study & Subjects

Handwritten Text Recognition Techniques Advanced Neural Network Applications Advanced Image and Video Retrieval Techniques

Publication details

Journal / source
Neurocomputing
Volume
681
Pages
133418–133418
Publication date
21 March 2026
Language
EN