Face Key Point Detection Using Depthwise Convolution for Low Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional face key point detection methods require a large computing amount, making them unsuitable for low-power devices and resulting in low accuracy, especially under extreme face pose conditions.
Innovation Solution
A face key point detection method using a depthwise convolution approach that converts multi-channel feature maps into target feature maps with a predetermined channel quantity, performing convolution operations with corresponding kernels to generate feature vectors and determine key point coordinates, reducing computational complexity while ensuring accurate extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional face key point detection methods using manually designed features and cascaded regression are used, then the computational complexity is reduced, but the prediction accuracy deteriorates, especially under extreme face pose conditions
Solution Approach 1:
The patent transforms the face key point detection problem from spatial coordinate prediction to channel dimension classification. By changing the parameter space from 2D/3D coordinates to channel indices, the method achieves both computational efficiency and high accuracy. The depthwise convolution operation processes each channel independently, reducing computational complexity while the channel-wise attention mechanism improves prediction accuracy by selectively emphasizing important features.
Solution Approach 2:
The patent moves the detection task from the spatial dimension to the channel dimension. Instead of predicting key point coordinates directly from spatial features, the method uses channel dimension to represent different key point positions. This dimensional transformation allows the use of efficient depthwise convolution operations while maintaining high prediction accuracy through channel-wise feature aggregation.
2Measurement precision
If deep learning-based face key point detection methods using convolutional neural networks are used, then the prediction accuracy is improved, but the computing amount increases, making them unsuitable for low-power devices
Solution Approach 1:
The patent segments the feature processing into independent channel-wise operations using depthwise convolution. Each channel is processed separately with its own convolution kernel, which reduces the total number of parameters and computational operations compared to standard convolutions. This segmentation enables accurate deep learning-based detection while reducing the computing amount to be suitable for low-power devices.
Solution Approach 2:
The patent changes the parameter representation from continuous spatial coordinates to discrete channel indices. This parameter transformation reduces the output dimensionality and computational burden while maintaining prediction accuracy. The channel dimension serves as a compact representation of key point positions, reducing the computing amount required for deep learning inference on low-power devices.
3Device complexity
If manual feature extraction methods are used for face key point detection, then the computational complexity is reduced, but the feature discrimination ability deteriorates, resulting in low accuracy
Solution Approach 1:
The patent introduces channel dimension as an intermediary between feature extraction and key point detection. The channel-wise features serve as intermediaries that capture discriminative information without requiring complex computational operations. This intermediary representation enables both low computational complexity and high feature discrimination ability by using channel statistics and depthwise convolution to capture meaningful patterns.
Data Source
AI summary
A face key point detection method includes determining, according to an image containing a face, a multi-channel feature map of the image; converting the multi-channel feature map into a predetermined channel quantity of target feature maps; performing a convolution operation on each target feature map in the predetermined channel quantity of target feature maps by using a convolution kernel corresponding to each target feature map; generating a feature vector corresponding to the image based on a result of the convolution operation on each target feature map; and determining key point coordinates of the face on the image according to the feature vector.