Multi-Scale Face Detection Network for Mobile Face Unlock
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face detection technologies are inefficient for face unlock applications on mobile devices due to varying face scales and computational constraints, failing to accurately detect tiny faces while incurring high computational costs.
Innovation Solution
A face detection device utilizing a convolutional neural network with a frontend feature extractor assembly and backend detector assembly, comprising multiple feature extractors and detectors, optimized for efficient face detection by reducing computational cost and improving accuracy through multi-scale feature extraction and bounding box classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current face detection technologies are used to detect faces with various scales in outdoor environments, then face detection accuracy is improved, but computational cost increases significantly
Solution Approach 1:
The face detection device is segmented into multiple independent feature extractors (first, second, and third feature extractors) that process images at different scales separately. Each feature extractor handles a specific scale range, allowing the system to maintain high detection accuracy for faces of varying sizes while reducing overall computational cost by avoiding unnecessary processing of all scales simultaneously.
Solution Approach 2:
Different feature extractors are designed with locally optimized characteristics tailored to specific scale ranges. The first feature extractor uses larger convolution kernels for small faces, while the second and third feature extractors use progressively smaller kernels for medium and large faces respectively. This local optimization ensures high accuracy for each scale while minimizing computational waste on irrelevant scale ranges.
2Measurement precision
If expensive computational methods are used to improve face detection accuracy, then detection precision increases, but the system becomes unsuitable for mobile devices with constrained computational power
Solution Approach 1:
The system dynamically selects which feature extractors to activate based on the detected face scale. When a face is detected at a particular scale, only the corresponding feature extractor is fully activated, while others remain dormant or are processed with reduced precision. This dynamic adaptation maintains high accuracy for the detected scale while significantly reducing computational load on mobile devices with constrained resources.
Solution Approach 2:
The convolution kernel sizes are systematically changed across different feature extractors to match the scale requirements. The first feature extractor uses larger kernels (e.g., 7x7) for small faces, while the second and third use smaller kernels (e.g., 5x5, 3x3) for larger faces. This parameter optimization enables the system to achieve mobile-device-friendly computational costs while maintaining detection accuracy across various face scales.
3Measurement precision
If multi-scale feature extraction is performed to detect tiny faces, then detection accuracy for small faces improves, but computational overhead increases
Solution Approach 1:
Instead of applying all feature extractors to all images uniformly (excessive action), the system applies only the necessary feature extractors based on the image content and detected face scales (partial action). When tiny faces are detected, only the first feature extractor is activated; when medium or large faces are detected, the second and third feature extractors are activated instead. This partial application reduces computational overhead while maintaining high accuracy for the detected face scale.
Data Source
AI summary
A face detection device based on a convolutional neural network is provided. The device includes a feature extractor assembly and a detector assembly. The feature extractor assembly includes a first feature extractor, a second feature extractor and a third feature extractor. The first feature extractor is used to apply a first set of convolution kernels on an input grayscale image thereby generate a set of basic feature maps. The second feature extractor is used to apply a second set of convolution kernels on the set of basic feature maps and thereby generate more than one set of intermediate feature maps, which are concatenated. The third feature extractor is used to perform at least one convolution operation on a concatenated layer. The detector assembly includes at least one detector whose input is derived from one of the second feature extractor and the third feature extractor.