Research On Face Detection Based On Convolutional Neural Networks

With the popularity of video equipment and the development of deep learning technologies,face detection algorithms based on convolutional neural networks are widely used in education,security,and transportation.Face detection still faces some challenges such as the differences among the video devices,the variousness among the faces,and the complexity of the scene where the images are captured.The research on face detection based on convolutional neural networks with real-time speed is carried out from three directions: ordinary faces,small faces and low-light dim faces.First of all,aiming at the problem that the recall is not ideal when the general object detection is applied to the face detection directly,a multi-scale parallel network structure from coarse to fine is proposed to detect faces of different sizes.In addition,the center loss function is introduced to reduce the distance within the class for improving the performance of this method.The experimental results show that the performance is close to the mainstream methods with real-time speed on the typical CelebA test set.Furthermore,a method based on enhanced feature module for detecting small faces is proposed.The multi-level network is used to enrich features of small faces at deep layer,and the feature enhanced module is used to extract the context information to further enhance the expressivity.In addition,a multi-stage loss function is used to balance the networks of different levels.The experimental results show that the method achieves competitive results with real-time speed on WIDER FACE and FDDB sets.Finally,a dim face detector based on image enhancement is proposed to solve the problem of dim faces caused by poor quality of images captured in low-light scenes.The image enhancement method with a fully convolutional neural network restores the color of the image and lights the image.Moreover,based on Retinex theory,a reflectance loss function is also proposed to improve the problem of uneven illumination in the training set for image enhancement,and to optimize the training of the enhanced network.The experimental results show that low-light images can be enhanced to obtain smooth,low-noise and real images.What's more,this image enhancement method combined with small face detector based on enhanced feature module can effectively detect faces in low light scenes which makes the detector more practical.