Robust and fast face detection using CNN based facial component heat map and face bound regression

Abstract

Face detection is the first step in a wide range of face applications, such as facial keypoint localization, face recognition, facial expression analysis and other topics. As a special task of object detection, a broad study of face detection has been conducted. The standard face detection dataset has several challenges for face detection. In particular, it contains faces at the wide range of scales, illumination and requires more precise localization due to occlusions and pose changes. Although recent studies show that deep learning approaches can achieve impressive performance on most computer vision tasks, face's localization caused by the above mentioned problems were not be fully resolved yet. There are additional challenges such as the repetitive convolution operation in the sliding window method and the resizing procedure due to the fixed size constraint of the fully-connected layers, which make it have a high computational complexity. Our main inspiration is to construct fully convolutional network. The architecture takes arbitrarily-sized input and produces last layer's feature maps of corresponding size represented as facial component heat map. Add to this we adopt multi-task learning technique. To address these challenges, we propose three-fold: (1) multi-scale proposal network for facial component heat map, (2) detection network with face bound regression, (3) hard sample mining to enhance the effectiveness of detection network by selecting the informative samples. We adopt the standard evaluation protocol to strictly validate our method. Experimental results present that our method outperforms other existing methods having used the standard FDDB dataset for face detection with 83.29% in the true positive rate at 1,034 false alarms. From the comparison of runtime efficiency, we can verify our method with competitive computational complexity. Our approach casts a light on face detection commonly deemed as a barrier which could not find the breakthrough in a trade-off between the performance and processing speed.