An image processing method, an image processing device, and an electronic device

By jointly learning networks to alternately perform depth estimation and semantic segmentation tasks and refine the prediction results at multiple scales, and by using semantic edges to align depth edges, the problems of insufficient information exchange and inconsistent boundaries between depth estimation and semantic segmentation tasks are solved, thereby improving accuracy and enhancing model generalization ability.

CN122289684APending Publication Date: 2026-06-26CHINA FAW CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINA FAW CO LTD
Filing Date
2026-03-30
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

In existing technologies, depth estimation and semantic segmentation tasks lack complementary depth information exchange, rely on accurate ground truth depth values, have blurred object boundaries during high-resolution prediction, poor handling of occluded regions, and limited model generalization ability.

Method used

A joint learning network is constructed to alternately perform depth estimation and semantic segmentation tasks. Attention is used to enhance the fusion of features, the prediction results are refined at multiple scales, and semantic edges are used to align depth edges. A local deformation function is constructed to adjust the depth edges.

Benefits of technology

It improves the accuracy and boundary consistency of depth estimation and semantic segmentation, reduces the dependence on ground truth depth, reduces occlusion artifacts, and enhances the generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122289684A_ABST
    Figure CN122289684A_ABST
Patent Text Reader

Abstract

This application provides an image processing method, an image processing apparatus, and an electronic device. The method includes: acquiring an initial image; inputting the initial image into a joint learning network, where the joint learning network alternately performs depth estimation and semantic segmentation tasks, and at each time step, performs attention-enhanced fusion of features obtained from the current task and features from the other task to obtain enhanced common features; refining the enhanced common features at multiple scales, and sequentially performing depth estimation and semantic segmentation to obtain a coarsely predicted depth map and a coarsely predicted semantic segmentation map; extracting semantic edges from the coarsely predicted semantic segmentation map and depth edges from the coarsely predicted depth map, and adjusting the depth edges through local deformation to align them with the semantic edges to obtain a final depth map and a final semantic segmentation map. This application improves the accuracy and boundary consistency of depth estimation and semantic segmentation.
Need to check novelty before this filing date? Find Prior Art