Human Segmentation Using Template and Deep Learning Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods for human segmentation in images rely on fixed human body templates that fail to accurately segment humans due to orientation and presence of only facial data, leading to inaccurate results.
Innovation Solution
A method and electronic device that perform human segmentation by combining a pre-defined full human body template with a deep neural network (DNN) trained model, using face detection to generate composite human segmentation, which improves accuracy and handles varying orientations and presence of full or partial human bodies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If fixed human body templates are applied based only on face detection, then the segmentation process is simple and fast, but the segmentation accuracy deteriorates when the image includes only facial data or when human orientation varies
Solution Approach 1:
The patent combines two different segmentation approaches: (1) template-based segmentation for fast processing, and (2) deep learning-based segmentation for high accuracy. The system merges the results of both methods to produce a final segmentation that benefits from both speed and accuracy, resolving the contradiction between processing speed and segmentation accuracy.
Solution Approach 2:
The patent dynamically changes the segmentation parameters based on the detected human orientation and the presence of facial data. When face detection is positive but body detection is negative, or when human orientation varies (tilted, rotated), the system adjusts the template application parameters or switches to deep learning-based segmentation, thereby maintaining accuracy across different scenarios while managing processing complexity.
2Device complexity
If fixed human body templates are used without considering human orientation, then the device complexity is low, but the segmentation accuracy deteriorates for tilted or rotated faces
Solution Approach 1:
The patent introduces dynamic orientation detection and adjustment mechanisms. The system detects the orientation of detected humans (including tilted or rotated faces) and dynamically adjusts the template application or switches to deep learning-based segmentation accordingly. This dynamic adaptation maintains segmentation accuracy for varied orientations without significantly increasing device complexity.
3Measurement precision
If deep learning models are used for human segmentation, then segmentation accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies deep learning-based segmentation selectively rather than universally. The system first attempts faster template-based segmentation and only invokes the computationally intensive deep learning model when necessary (e.g., when template-based methods fail to achieve sufficient accuracy or when human orientation varies). This partial application of the more accurate but slower method reduces overall processing time while maintaining high accuracy when needed.
Data Source
AI summary
An electronic device and a method for human segmentation in an image are provided. The method includes obtaining the image comprising at least one human, performing face detection to detect the human in the image, generating a first human segmentation from the image by comparing a pre-defined full human body template and the detected human, generating a second human segmentation from the image, by inputting the image to an artificial intelligence trained model and obtaining a composite human segmentation representing the detected human by using both of the first human segmentation and the second human segmentation.