Face occlusion detection based on multi-task convolution neural network

With the rise of crimes associated with ATM, security reinforcement by surveillance techniques has been in high agenda for both academia and industries. Though cameras are generally installed in ATMs to capture the facial images of users, the function is only limited to recording for follow-up criminal investigations, which could become useless when a criminal's face is occluded. Therefore, face occlusion detection has become very important to prevent crimes connected with ATMs. Traditional approaches to solve the problem typically consist of a succession of steps such as localization, segmentation, feature extraction and recognition. This paper proposes robust and effective facial occlusion detection based on convolutional neural networks (ConvNets) with multi-task learning. Covering of different facial parts, namely, left eye, right eye, nose and mouth, can be predicted by the multi-task CNN. In comparison with previous approaches, CNN is optimal from the system point of view as the design is based on end-to-end principle and the model operates directly on the image pixels. We created a large scale face occlusion database, consisting of over fifty thousand images, with annotated facial parts. Experimental results revealed that the proposed method is extremely effective.