Face Detection In Unrestricted Environment

In recent years,pattern recognition has received widespread attention,especially face detection.Face detection is a large category in object detection.As an important step in face recognition,it is also an important part of human-computer interaction.It is widely used in various fields,such as street monitoring and access control.Considering the lighting changes,pose expressions,face deflections,occlusions,etc.,how to detect a face accurately from an image in unrestricted environment has become a hot spot in current research.The research content of this thesis is face detection algorithms in unrestricted scenarios,containing the traditional algorithm and deep learning.The following are the main research contents of this thesis:In some more complex scenarios,such as faces with many facial expressions,light interference,and complex backgrounds,face detection using only the Ada Boost algorithm may be missed.The skin color information is well clustered.A large number of very complex backgrounds can be eliminated by skin color segmentation.Therefore,a face detection algorithm combining skin color segmentation and Ada Boost algorithm is proposed.First,a new fusion HSV-YCb Cr color space is proposed.Skin color detection is performed on this color space,and morphological processing is used to generate candidate regions that belong to the skin color.A large number of regions that do not belong to the face are screened out.Then,based on the original Haar-like features,it is extended,and then the Ada Boost algorithm is used to detect the face area in the skin color candidate area obtained in the previous step.Experiments show that combining skin color information with Ada Boost algorithm to detect faces can reduce the impact of face detection in complex scenes to a certain extent,and improves the performance of face detection compared to other methods.For traditional methods,there are certain limitations to face detection in unrestricted scenarios.For example,the face deflection angle is too large,the background is very complicated,etc.The detection effect is often not ideal,and these limitations are less in deep learning.Single-stage neural networks require deep and complex structures to ensure the performance of the system,while cascade structures are often relatively simple and efficient.Therefore,a new two-stage convolutional neural network is designed to detect faces.After pre-processing the image,the image is transformed into an image pyramid and then input to a neural network.The first stage of the network aims to search the regression vector of the face window and its bounding box.The second stage uses Soft-NMS to merge highly covered face candidate regions.Comparison of AFW and FDDB datasets with other algorithms confirms the effectiveness of the algorithm.