A Face Detection Framework Based on Deep Cascaded Full Convolutional Neural Networks

摘要

In practical applications, the collected face images are often affected by the surrounding environment, resulting in multiple expressions, multiple poses, occlusion, light intensity, and complex background issues in face detection. Hence this paper presents a novel face detection framework based on deep cascaded full convolutional neural networks (CNNs) to solve the mentioned issues in face detection. This frame work supports face detection, and positioning of face key points at the same time by using its 3-order cascaded CNN architecture. The 3-order cascaded architecture is a combination of three phases of network layers; phasel (I-Net, Initial Network), phase2 (A-Net, Advanced Network), and phase3 (U-Net, Ultimate Network). In the CNN design, a depthwise separable convolution is used instead of the traditional convolution, and also a convolutional layer with stride of 2 is used instead of the pooling layer. A bottleneck structure (1x1 convolution kernel structure) is used for convolution layer, and the network uses a convolutional layer instead of the fully connected layer (FC). The carried-out experiments proved that the presented framework is an effective one for face detection applications.