Face Detection in Unconstrained Environments Using Modified Multitask Cascade Convolutional Neural Network

Due to occlusion and variation of poses, facial detection is a challenging task to accomplish. In facial detection, occlusion has always been a standing challenge. Facial occlusion like sunglasses, scarf, and mask and pose variation are crucial factors that affect the performance of face detection. Undesirably, in the real-world scenario, occlusions are a very common situation that arises in the face detection and recognition problem. To deal with this problem, we put forward the modified multitask cascade convolutional neural network (M-MTCNN) with a slight modification. MTCNN is a trainable unit which may be included in present CNN architectures. With end-to-end training supervised by only the private identity labels, Mask Net learns a correct way of adaptively generating different feature map masks for various occluded face images. This paper deals with an efficient method for the detection of numerous occluded and pose variation faces. In addition to the marking of the face with a square box, there are five landmarks drawn (the two eyes, one in the nose, and two identifying the lips). Also, using the landmark points of the eyes, we have tried to mark the eyes using the eye landmarks as the center point. Using the landmarks of the lips, we also have drawn a straight line marking the edge of the lips. The presented method has been tasted on WIDER database and obtained efficient detection of multiple occluded faces.