Face Detection Reparameterization for Multi-Scale Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional reparameterization algorithms for face detection models are limited in extracting face semantic features during training, leading to poor performance improvement.
Innovation Solution
Implement a face detection method where different convolution kernels in the reparameterization module correspond to different scales of faces, allowing them to extract more diverse face semantic features by adjusting model parameters based on gradient weights for each scale.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the existing convolutional reparameterization algorithm increases the number of convolution kernels during training, then the face detection model performance can be improved without increasing inference consumption, but the ability to extract diverse face semantic features remains limited
Solution Approach 1:
The patent segments the face detection task by dividing detection boxes into multiple groups based on scale characteristics (large-scale, medium-scale, small-scale faces). Different convolution kernels are assigned to different scale groups, allowing each kernel to specialize in extracting features for specific face scales. This segmentation enables diverse semantic feature extraction while maintaining the reparameterization benefit of fused parameters during inference.
2Speed
If convolution kernels are fused during inference to maintain efficiency, then computation speed is preserved, but the diversity of extracted semantic features is reduced
Solution Approach 1:
The patent implements a dynamic feature extraction mechanism where the system adaptsively selects which convolution kernel to use based on the scale of the detection box being processed. During training, multiple specialized kernels learn different scale-specific features, and during inference, the appropriate kernel is dynamically selected rather than always using a fused single kernel. This dynamic selection preserves feature diversity while maintaining inference efficiency.
Data Source
AI summary
A face detection method, a computer-readable storage medium, and an electronic device are provided. The method includes: obtaining the to-be-detected target image; and detecting, using a preset face detection model, the face in the target image; where, a reparameterization module of the face detection model during training includes a plurality of convolution kernels each corresponding to an individual scale of faces. In this manner, during the training of the face detection model, different convolution kernels in the reparameterization module will respectively correspond to faces at different scales, so that different convolution kernels can extract diverse face semantic features in a more extent, thereby effectively improving the performance of the face detection model and obtaining more accurate face detection results.