Imaginary Face Generation for Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Face recognition systems in high-risk factories face inefficiencies due to limited training data diversity and sensitivity to ambient light, especially when personnel wear helmets and goggles, leading to poor recognition accuracy.
Innovation Solution
An imaginary face generation method and system that combines face color and depth images to create diverse virtual faces through image mixing and landmark alignment, enhancing training data diversity for face recognition systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of training face images is increased to improve recognition accuracy, then face recognition accuracy improves, but the complexity of data collection and processing increases
Solution Approach 1:
The patent creates virtual face images by copying and transforming existing real face images through depth map generation, color transfer, and blending operations. This copying approach generates additional training data without requiring physical collection of more real face images, thereby improving recognition accuracy while avoiding the complexity of expanding data collection infrastructure
Solution Approach 2:
The patent introduces depth maps as an intermediary element between real face images and virtual face images. By generating depth information from color images and using it to create realistic 3D effects in virtual faces, the system achieves diverse training data generation without directly manipulating large volumes of raw image data, simplifying the processing pipeline
2Reliability
If face recognition is performed in high-risk factory environments with helmets and goggles, then personnel safety is ensured, but face recognition accuracy deteriorates
Solution Approach 1:
The patent transforms the face image representation by converting 2D color images into 3D-like virtual images with depth information. This parameter change in image dimensionality and structure enables the recognition system to capture facial features that are less affected by occlusions from helmets and goggles, maintaining safety compliance while improving recognition accuracy
Solution Approach 2:
The patent adds a depth dimension to face images by generating depth maps and creating virtual 3D face representations. This dimensional transformation provides additional spatial information that helps the recognition system identify facial features even when partially occluded by safety equipment like helmets and goggles
3Productivity
If deep learning models use more diverse training data to improve recognition efficiency, then recognition efficiency improves, but the time and resources required for data preparation increase
Solution Approach 1:
The patent generates diverse training data by copying existing face images and applying various transformations including depth map generation, color space conversions, and blending operations. This copying-based approach creates numerous virtual face variants from a small set of real images, improving recognition efficiency without requiring proportional increases in data preparation time
Solution Approach 2:
The patent enables the training data generation process to be self-service by automatically generating virtual face images from real face images through automated depth estimation, color transfer, and blending pipelines. This self-service approach eliminates the need for manual data annotation and preparation, significantly reducing data preparation time while maintaining high diversity in training data
Data Source
AI summary
A face depth image is normalized and color-transferred into a normalized face depth image. The face color image and the normalized face depth image are mixed into a face mixed image. A plurality of face mixed images of several different users are processed with face landmark alignment and mean, and then are synthesized with the face mixed image of another user into an imaginary face.