Research On Face Segmentation Based On Deep Learning

Face is a very important biological feature of the human body.Face detection,face segmentation and other face processing technologies have been widely used in security,communication,medical treatment and other fields,and have practical significance for in-depth research.The aim of face segmentation is to find the contour of the face and separate the face from the image background.As the basic algorithm of many applications,face segmentation requires high accuracy.Due to the pose,face shape,hair style,camera angle and other reasons,the face contour in the image has great geometric differences,these factors affect the segmentation accuracy.At present,the inaccurate face segmentation caused by inaccurate face detection still exists,and it is difficult to segment small face accurately.The work of this dissertation is as follows:(1)To improve the detection stage,a method of three fine adjustments is proposed.During the third adjustment,the center of the detection box is fixed and adjust the border directly to obtain more accurate face detection results,then we can get more accurate face segmentation boundary.The experimental results show that the scheme is effective.(2)On the basis of three fine tuning,a new structure is designed,which integrates the deep and shallow features of the network.The method combines channel attention and spatial attention,and uses Depthwise convolution is used to provide the corresponding attention weights for each channel.Under the guidance of attention mechanism,the fused features are more advantageous to the precise segmentation task.Experiments show that this method is accurate and efficient in the face segmentation task.(3)In order to train the model effectively,the training set needs multiple face images with different area ratios.Therefore,a face data set with more abundant face dimensions is combined,and the data set is labeled by hand.The experimental results are compared with the classical instance segmentation model and the latest face segmentation model.It is proved that the method proposed in this dissertation is better,it has good accuracy and high efficiency.(4)The proposed segmentation model is verified in instance segmentation and face segmentation.The results and accuracy of object detection,instance segmentation,face detection and face segmentation are tested by experiments.It is proved that the proposed scheme has good results in instance segmentation and face segmentation tasks.