Image Generation Model for Person Posture Interpolation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image interpolation techniques struggle with accurately interpolating large lost portions of a person in images, leading to unnatural postures and decreased analysis accuracy, especially when objects or strong light affects the detection-target person.
Innovation Solution
An image processing apparatus and method that constructs an image generation model using machine learning with both sample images of a person and human body information, such as posture data, to interpolate lost parts of a person's image, ensuring natural posture preservation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of moving object
If machine learning is performed merely to produce a target image from an input image without considering posture, then a lost portion can be interpolated even if the area is large, but the posture of the person after interpolation may become unnatural
Solution Approach 1:
The patent changes the parameters used in machine learning from only image data to include both image data and posture data (skeleton information). By incorporating posture parameters as additional constraints during the learning process, the system maintains natural human poses while interpolating large lost portions, resolving the contradiction between handling large areas and preserving posture accuracy
Solution Approach 2:
The patent introduces posture information (skeleton data) as an intermediary element that mediates between the input image and the generated output image. This intermediary constraint ensures that the interpolated regions maintain anatomically correct human poses, preventing unnatural postures while filling large lost areas
2Manufacturing precision
If image interpolation is performed merely to achieve local region continuity, then the texture can be continuous at boundaries, but the larger the area of lost portion becomes, the more difficult it becomes to accurately interpolate
Solution Approach 1:
The patent transitions from two-dimensional image pixel data to three-dimensional posture information (skeleton coordinates and joint angles). By utilizing this additional dimensional data, the system can accurately reconstruct large lost portions by leveraging the structural constraints of human anatomy, overcoming the limitations of purely two-dimensional texture interpolation
Data Source
AI summary
An image processing apparatus 10 includes an image generation model construction unit 11 configured to construct an image generation model for, using image data in which a part of a person is lost as an input image, generating an image in which the lost part of the person has been interpolated. The image generation model construction unit 11 constructs the image generation model by machine learning in which a first sample image including an image of a person, a second sample image in which a part of the person in the first sample image is lost, and human body information specifying a posture of the person in the first sample image are used.