| Key point detection of face,namely face alignment,has become a hot topic in the field of artificial intelligence research.Based on the object box obtained by face detection technology,the exact location of key points of face is further obtained.At the same time,as the front step of face verification,key point detection of face provides accurate facial key point location for the research of the expression analysis and face recognition.Key point detection of face is closely related to life,industry and business applications.Whether it is in the field of information security,such as monitoring equipment,security equipment,authentication equipment,or in the field of commercial applications,such as image processing,the research needs to take key point detection of face as technical support.Therefore,the research of key point detection technology of face has great application value and practical significance.In the paper,the following three technical improvements are made for the task of key point detection of face:(1)Firstly,the SSD detector based on the single step network is improved,and the feature layer containing more details which is closer to the underlying network is added for cascaded prediction to obtain the excellent performance of small object detection task for face.Secondly,the scale of the default box is adjusted to select a fixed proportion which is more suitable for the face shape to achieve a better fitting effect of the default box for the face.Finally,the improved scheme is compared and analyzed on the mainstream face data set.(2)The paper improves the NMS algorithm to suppress redundant prediction box of face.Firstly,the paper analyzes the problems that NMS algorithm may cause about the face detection of small objects.Secondly,an improved scheme is proposed to divide the prediction box of face into three segments according to the confidence score.The prediction box with the lowest confidence score will not be penalized or changed.The prediction box with the middle and high confidence score will be penalized in different degrees by the mechanism of rescoring with Gaussian function.Finally,in the mainstream face data set and the real-time scene under the USB camera,the overall improved SSD detector is compared and analyzed.(3)In the paper,the improved SSD is used to get the detection box of face.On the basis,the LBF key point detection algorithm with excellent speed and performance is improved.Firstly,the difference of random pixel pairs in the key region used in LBF algorithm is used as the local binary feature.Secondly,the paper proposes a multi angle initialization algorithm based on pixel difference to optimize the key point detection algorithm of LBF architecture.In the paper,five groups of key points with different tilt angles are used to initialize the face.The initialization scheme can achieve excellent fitting results for faces with different tilt angles.Then five groups of initial shapes are put into the network training to get five groups of predicted shapes.The variance value of the eye region with the largest pixel difference in each group of shapes is calculated.The prediction scheme with the largest variance value is selected as the key point position of the final prediction.Finally,the paper compares and analyzes the performance of the improved key point detection architecture of face in the mainstream face key data set and USB camera real-time scene. |
