Unsupervised based method for rectangularization of stitched images
By using an unsupervised learning framework and an irregular boundary awareness mechanism, combined with the ResNet18 network and a multi-constraint loss function, efficient rectangularization of irregular boundary images is achieved. This solves the problems of relying on manual annotation and insufficient adaptability to complex scenes in existing technologies, and is applicable to multiple practical application scenarios.
CN122415786APending Publication Date: 2026-07-17ZHEJIANG UNIVERSITY OF MEDIA AND COMMUNICATIONS
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIVERSITY OF MEDIA AND COMMUNICATIONS
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-17
Smart Images
Figure CN122415786A_ABST
Abstract
基于无监督的拼接图像矩形化方法。本发明公开了一种无监督的拼接图像矩形化方法，在无标注数据的驱动下，通过网格变形将不规则边界图像规整为矩形，包括：S10，提取不规则拼接图像与二进制掩码的融合特征；S20，进行基于不规则边界感知的几何先验提取；S30，进行基于网格运动回归的图像扭曲；S40，进行基于多约束损失函数的矩形化结果优化。该方法能够在无监督学习框架下实现对不规则边界拼接图像的高效矩形化处理，在保持图像语义内容完整性的同时，输出边界规则的矩形图像，无需依赖人工标注数据集。
Need to check novelty before this filing date? Find Prior Art