Fast and Efficient Face Detector Based on Large Kernel Attention for CPU Device.
Shuaihui Qi,Xiaofeng Song,Zhiyuan Li,Tao Xie
DOI: https://doi.org/10.1007/s11554-023-01326-3
IF: 2.293
2023-01-01
Journal of Real-Time Image Processing
Abstract:Efficient face detectors with low computational cost and fast speed are still a pressing problem despite the significant progress made in uncontrolled face detection. To address this issue, we propose two lightweight face detectors for CPU device, named speed-priority face detector (SPFD) and accuracy-priority face detector (APFD). In the case of SPFD, we propose a simplified version of FaceBoxes by reducing convolutional layers and channel number for reducing model complexity and computational cost, as well as replacing max pooling layers with group convolutional layers to learn local information and features’ integration. Additionally, large kernel attention modules based on prior knowledge of face size and network architecture are applied to increase the expression ability of features and capture the key information from long distances. Finally, an iterative retrained method is designed to further enhance the accuracy without increasing any cost during model testing. Regarding APFD, a new anchor generation strategy is utilized to find out more faces based on SPFD. Extensive experiments conducted on WIDER FACE validation dataset indicate that our detectors exceed FaceBoxes comprehensively in terms of accuracy and speed. Specifically, SPFD outperforms FaceBoxes by 4.1%, 8.6%, and 9.6% on WIDEFACE validation dataset while achieving the fastest detection speed on CPU device for VGA-resolution images. The average speed can reach 45FPS on CPU devices, while the size of its parameters is only 0.66 times of FaceBoxes. Moreover, the APFD outperforms many famous and lightweight face detectors and attains superior accuracy (easy: 91.4%, medium:88.1%, and hard: 64.7%) with 20FPS; it achieves the best trade-off between accuracy and speed for face detection.