一种融合可变形卷积与球面注意力的全景图像双重感知增强方法与装置
By employing a U-shaped encoder-decoder network in panoramic images, combined with deformable convolution and spherical attention, the problems of geometric distortion and insufficient global modeling capabilities in panoramic image processing are solved, achieving efficient image perception enhancement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU DIANZI UNIV
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-17
AI Technical Summary
When processing panoramic images based on equidistant cylindrical projection, existing technologies often fail to maintain geometric consistency due to the difficulty of traditional convolutional neural networks in maintaining geometric consistency. This results in severe polar distortion, which affects the accuracy of downstream perception tasks. Meanwhile, pure Transformer methods lack the local inductive bias of convolution and cannot effectively capture long-range global contextual dependencies in images.
A U-shaped encoding and decoding network is adopted, which combines deformable convolution and spherical attention. By mapping panoramic images to the vertices of a polyhedral mesh, deformable spherical convolution branches and deformable spherical local self-attention branches are used to extract local geometric deformation features and global context features in parallel. Feature fusion is performed through adaptive gating fusion blocks to achieve a deep integration of geometric adaptation capability and global modeling capability.
It enhances the geometric adaptation and global modeling capabilities of panoramic images, solves the geometric distortion problem existing in traditional methods, strengthens the adaptability to irregular geometric deformations, and improves the accuracy and efficiency of image perception tasks by dynamically fusing weights to balance the contributions of local and global features.
Smart Images
Figure CN122175811B_ABST