Facial Image Synthesis Using Depth-Image Feature Point Offsets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing applications require high-precision 3D models and complex architectures for generating 3D facial reconstructions, leading to slow processing speeds and inefficiencies in creating images from different angles.
Innovation Solution
A method that performs feature point recognition on a face image, determines position offset information, and adjusts facial feature points to generate a target depth image, allowing for direction deflection of the image without relying on 3D models, enabling efficient synthesis of images from different angles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If 3D facial reconstruction technology is applied to generate face images from different angles, then image processing functionality is improved, but device complexity increases due to high-precision 3D model creation and deformable 3D model requirements
Solution Approach 1:
The patent uses a reference face depth image as a template copy instead of creating a personalized 3D model. The method copies the depth structure from the reference image and adjusts it through feature point offset information to generate target depth images, avoiding the complexity of 3D model creation while maintaining the ability to generate face images from different angles
Solution Approach 2:
The patent changes the approach from 3D model parameter adjustment to 2D feature point position offset adjustment. By modifying the position parameters of facial feature points based on offset information from the target face image, the system achieves angle transformation without requiring deformable 3D models or complex parameter adjustments
2Measurement precision
If 3D model-based methods are used for image synthesis, then image processing accuracy is improved, but processing speed deteriorates due to complex architecture and large quantity of implementation requirements
Solution Approach 1:
The patent segments the face image into key feature points (eyes, nose, mouth, ears) and processes each feature point independently using position offset information. This segmentation approach maintains matching accuracy by focusing on critical facial landmarks while significantly reducing computational complexity compared to processing entire 3D models, thereby improving processing speed
Solution Approach 2:
The patent extracts only the essential element - the position offset information of facial feature points - from the complex 3D model framework. By taking out and using only this critical data component, the system achieves accurate facial feature matching without the computational burden of full 3D model processing, thus improving processing speed while maintaining accuracy
Data Source
Figure 1a~1b
Figure 2a~2b
Figure 3a~4
AI summary
Disclosed are an image processing method and an image synthesis method, an image processing apparatus and an image synthesis apparatus, and a storage medium. The method comprises: performing feature point recognition on a facial image to be processed, so as to obtain facial part feature points to be processed, wherein said facial part feature points are facial part feature points of said facial image (101); calculating feature point position offset information of said facial part feature points relative to reference facial part feature points, wherein the reference facial part feature points are facial part feature points corresponding to a reference facial image (102); on the basis of the feature point position offset information, performing position adjustment on facial part feature points in a reference facial depth image to obtain a target facial depth image corresponding to said facial image, wherein the reference facial depth image is a facial depth image corresponding to the reference facial image (103); and according to the target facial depth image, performing direction deflection on said facial image to obtain a target facial image (104).