Auto Eye Pixelate System Based On Convolutional Neural Network

In order to protect the privacy of people in the video,it is sometimes necessary to mosaic the human eye of the video to erase the identity information.It needs to point out the eye location manually to mosaic the human eye in the video,which causes low processing speed and a lot of manpower consumption.In recent years,with the rapid development of deep learning,computer vision technology such as face detection,has gradually developed and applied in practice.So the job to mosaic the human eye can also be handed over to the computer.In this paper,an automatic human eye mosaic system based on convolutional neural network is proposed,which can find the human eyes quickly and correctly and then mosaic them.It eliminates the manual operation and greatly improves the processing speed,making it convenient to protect personal privacy in videos.In order to achieve this function,the system first detects the human face,then locates and tracks some feature points around human eye.And finally mosaics the position base on the location of the feature points.For practical application,the automatic human eye mosaic system should have high processing speed and accuracy,the failure and error rate should be kept at a low level.To this end,the main research work and contributions of this paper are as follows:1?Improving the FaceBoxes human face detection algorithm that can run under the CPU in real time,and propose the PyramidFaces human face detection algorithm,which enable the system to detect face quickly and accurately.Improvements include: improved network structure,increased network depth and width,and removed hard training data.The experiment shows that the PyramidFaces greatly improves the accuracy,while the calculation time is only slightly increased,which meets the requirements of the system.2?Based on convolutional neural networks,a suitable human eye feature point location algorithm and human eye feature point tacking algorithm for this system are built.The human eye feature point locator can reliably and accurately locate the human eye feature points from the face region,while the human eye feature point tracker can quickly track the feature point of the current frame from the feature point of the previous frame.And when it can't track accurately,it can tell the system to stop tracking.By using a large scale 3D facial landmarkdata set and convolutional neural networks,the system can handle well with the difficulty of multi-pose localization.The blocked feature point can still be located even in high-angle face posture.3 ?Integrating the three algorithms proposed in this paper: PyramidFaces human face detection algorithm,human eye feature point location algorithm and human eye feature point tacking algorithm,an automatic mosaic human eye system is built.The testing result in various videos illustrates that this system can mosaic the human eyes with robustness and efficiency.The system greatly improves the processing speed of eye mosaic and decreases manual operations.