Face Detection And Recognition Based On Deep Learning

In recent years, Artificial Intelligence (AI), as an important branch of computer science, has received extensive attention from the academic and industrial circles. Face analysis is one of the most frequently used skills in human daily life. Teaching computer to perceive, analyze and understand human face is a key step to achieve Artificial Intelligence. Automatic face analysis technology, including face detection, face recognition and face attribute analysis, emerges as the times require. The related research of face analysis technology has important theoretical significance andapplication value. The technology can promote the research field of human cognitive science and image analysis,and it is also widely used in public security,data management, multimedia entertainment services etc.The face image will be affected by environmental factors such as illumination,occlusion and internal factors such as age, expression, which will brought great challenges to face analysis technology. Occlusion and blur require the face detection technology to be more robust. Because of the large intra class difference and the small inter class variation, strong discriminative features should be extracted by the algorithm automatically. Face detection cannot provide facial landmark for face alignment, which has a great effect on the precision of face recognition. So fast and accurate facial landmark detection under the unconstrained conditions is very important. Deep learning has achieved a great success in the field of computer vision and natural language processing. Deep learning especially the convolutional neural network (CNN) has been used in automatic facial image analysis.In this paper, we mainly study the application of convolution neural network in face detection, facial landmark location and face recognition. The main research work is as follows:1. Considering in some specific application, face image collected by the camera are blur, occlusion. What's more,a given image may contain multiple size and density faces. This paper puts forward the step by step training strategy.Using the training data prepared carefully to train the deep convolutional neural network model makes the final face detector can detect faces in multi pose, blurred and occlusion well. In addition, multi model fusion and multi scale detection is helpful to detect faces in various size and density.Experimental results on FDDB face detection benchmark and collective self randomly capture on the Internet verify the validity of our method.2. Facial landmark detection is a basic but crucial part of face image analysis.It is necessary to locate facial landmark quickly and accurately. Based on the multi-task convolutional neural network, this paper proposes a landmark detection method named leptosomic multi-task convolutional neural network.The "thin" network keeps low computational overhead and "tall" network ensures the learning ability. Effective data augmentation can reduce the risk of over-fitting and improve the generalization ability. The network can quickly and accurately complete the facial landmark detection task on the CPU, which can adapt to the mobile terminal applications well. The experimental results on AFLW database demonstrate the effectiveness of our method.3. Inspired by the human transfer learning mechanism and the top-down visual attention mechanism, this paper proposes a face recognition network based on transfer learning and specific learning. According to the transferability in different layers of source model, we transfer the source knowledge (for general object recognition) to target task (face recognition) selectively, which reduces the dependence of large scale training data and computing resources during learning. Combining current popular deep learning tricks such as Leaky ReLU,batch normalization,we train a specific network with global and local face patch to obtain knowledge specifically for face recognition.Experiments on LFW and CASIA-Webface dataset verify the effectiveness of feature extracted by TS-Net and the binary one in face verification and face identification tasks.