Identification of face along with configuration beneath unobstructed ambiance via reflective deep cascaded neural networks

Abstract Detection and alignment of face under nonrestricted environment are forcing us to make a lot of effort due to variation in threat, light, and occlusion. Many of the current studies teach us that these two tasks (face alignment and face detection) can be achieved by the deep learning approach. In this research paper, we study and implement a cascaded neural network framework using deep learning to develop the basic correlation among them so that their performance can be enhanced. Our framework particularly uses a streamed or cascaded structure that includes three storage of designing deep convolutional network with great attention and predicts face and landmarks in a nonsmooth to smooth manner. Including under the learning process we suggested a new online hard sample mining strategy without manual sample selection. These methods achieved more accuracy as compared to leading techniques available on making a lot of effort on Face Detection Dataset and Benchmark and WIDER, FACE Benchmark for face detection and also the Annotated Facial Landmark Wild Benchmark for face alignment in real-time performance. We have used modified convolutional neural networks (CNNs) for the detection and alignment of the face in this task. In our method, we have used various CNNs. Each of the networks performs different tasks to get better performance. The output of first network is used as an input to another network, this method results in overall increase of accuracy. Our proposed model has an accuracy of about 95.4%.