Hair Region Segmentation Using Face Key-Point Probability Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image segmentation methods for hair region detection suffer from low precision and efficiency, particularly in varying lighting conditions and when dealing with the inherent characteristics of hair and face, leading to poor robustness and excessive calculations.
Innovation Solution
A method involving face detection, key-point detection, and image segmentation using a deep learning model trained with sample images, label mask images, and a probability distribution map of hair, which includes averaging labeled data to generate a prior probability distribution for hair segmentation, thereby improving precision and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image segmentation methods are used for hair region detection, then the segmentation can be performed, but the precision is low and the efficiency is poor
Solution Approach 1:
The patent segments the image processing task into multiple stages: face detection to locate the face region, key-point detection to identify facial landmarks, and then hair segmentation only within the bounded face region. This multi-level segmentation approach improves both precision by focusing on relevant areas and efficiency by reducing the overall processing scope.
Solution Approach 2:
The patent applies different processing qualities to different regions: full-image processing for face detection, localized processing within the face bounding box for hair segmentation, and uses key-point information to enhance boundary accuracy. This local quality differentiation optimizes both precision and computational efficiency.
2Reliability
If the entire image is processed for hair segmentation, then complete coverage is achieved, but unnecessary calculations increase reducing efficiency
Solution Approach 1:
The patent extracts the face region from the entire image using face detection algorithms, then performs hair segmentation only within this extracted region. This extraction principle maintains segmentation completeness for the hair region while dramatically reducing calculation time by excluding irrelevant background areas.
Solution Approach 2:
The patent performs preliminary face detection and key-point identification before executing the hair segmentation. This preliminary action defines the precise processing region in advance, ensuring no hair pixels are missed while avoiding unnecessary processing of non-hair areas, thus balancing completeness and efficiency.
3Device complexity
If standard segmentation algorithms are used, then the process is simple, but the robustness to lighting and environmental changes is poor
Solution Approach 1:
The patent introduces face detection and key-point detection as intermediary steps between the input image and hair segmentation. These intermediaries provide robust landmark-based references that are less sensitive to lighting changes, thereby improving the overall robustness of the segmentation process while maintaining reasonable algorithmic complexity.
Solution Approach 2:
The patent changes the processing parameters dynamically by adapting the segmentation region based on detected face boundaries and key-points. This parameter adaptation allows the algorithm to maintain robustness across different lighting and environmental conditions by focusing computational resources on the relevant facial region regardless of environmental variations.
Data Source
AI summary
The disclosure relates to a method, an electronic device and a storage medium for segmenting an image. The method includes: obtaining an image to be segmented; determining a first face result by detecting a face in the image and a first key-point result by detecting one or more key-points of the face in the image; determining a first face region in the image based on the first face result and the second face result; and segmenting a hair region from the first face region by an image segmentation model, wherein the image segmentation model is trained based on sample images, label mask images and a probability distribution map of hair, and the label mask image comprises a hair region of the sample image, and the probability distribution map of hair comprises a probability distribution of hair in the sample images.