Research On Face Detection Method Based On Fully Convolutional Neural Network

Face detection refers to automatically determine whether there is a face in a given image by using machine learning and computer vision methods,and return the position of each face if there is a face.The performance of face detection algorithms is easily affected by the large intra-class differences caused by many factors,such as scale,posture,occlusion,makeup,and illumination changes which makes it be a difficult problem and also a hot topic in computer vision area.Face detection can be applied to intelligent video analysis,security monitoring,authentication and other fields.Therefore,it is of great theoretical and practical significance to carry out research on automatic face detection technology.After reviewing previous methods,we group current face detection methods into two main categories:the candidate box based methods and the direct regression methods.Current methods can easily detect face with clear appearances,but still struggle in cope with blurry and small faces in real application scenes.In this thesis,we analyzed the factors affecting the performance in direct regression face detection methods and implements a direct regression face detection method.The experimental analysis shows the effectiveness of the proposed method.The main research work and innovations of this thesis are as follows:1)Current non-maximum suppression(NMS)algorithm have the miss detection problem of nearby occluded faces,to solve this issue,a weighted non-maximal suppression algorithm is proposed in this thesis.During the clustering process in NMS,we decrease the scores of candidate bounding boxes based on the intersection over union ratios(IoU)computed with the bounding box of highest score,then filter out those bounding boxes lower than a predefined threshold.This method preserves multiple bounding boxes on the same face area,avoids the miss detection of nearby faces,and increases the recall rate.2)Current direct regression face detection methods only use a single layer which lack of expressive ability,so a hierarchical multi-level feature fusion method is proposed in this case.We construct a feature pyramid by applying summation operations to features of two adjacent layers,and then stack all layers to get a global level feature map.To select positive examples,we propose a sampling methods which can adaptively choose positive examples based on a template.To do hard example mining,we use the IoU ratio as a distance metric,and use a weighted method to increase the weight of the difficult faces in the final loss function.