Skin Segmentation Based On Conditional Generative Adversarial Networks And Face Color Classification

Computer vision is the hotspot of research in the field of artificial intelligence,and face image is an important research object in computer vision research.There are many attributes in the face images that can be used for classification and identity authentication.As one of the most important attributes of face,face color is more stable and easier to obtain compared to other biological attributes.Also,face recognition classification is of great significance for face tracking and expression recognition.In this paper,the whole face color classification system is divided into three parts: skin segmentation,face color feature extraction and face color classification.In the skin segmentation experiment,we built a multi-scale feature fusion network based on the Conditional Generative Adversarial Networks.The multi-scale feature fusion network is used as a generator,and the face image is segmented with different resolutions feature maps from ResNet101,and the segmentation result is input to the discriminator.The experimental results show that our segmentation method is very effective.Due to the weaker distinguishability of color features,only using color features for face color classification is unsatisfactory.Therefore,in this paper,the texture features and the color features is combined as the feature for face color classification.We input the extracted face color features into the Support Vector Machine and the BP neural network.In the face color classification experiment,we use 1000 skin images as the training dataset and 128 images as the test dataset.The face color is divided into three categories: deep,medium and light.The experimental results show that the recognition performance is the best when using the normalized RGB color feature,HSV color feature,Lab color feature,texture feature of gray level co-occurrence matrix and texture feature of LBP.The average classification accuracy rate of Support Vector Machine reached 87%,and the average classification accuracy rate of the BP neural network classifier reached 95%.the main contributions can be summarized as follows:1.The theoretical principle and structural model of convolutional neural network are introduced in detail.The application of the full convolutional neural network in image semantic segmentation is also introduced.2.In the skin segmentation experiment,a multi-scale feature fusion method based on conditional generation antagonism network is proposed.Traditional algorithms only use the last low-resolution feature map,and the low-resolution feature map will lose some important information in the high-resolution feature map,which may be very useful for the location of pixels.Therefore,the multi-scale feature fusion network in this paper study the feature maps of each scale,and the extracted features is fused to achieve skin segmentation.At the same time,because of the good performance of conditional generation antagonism network,this paper employs the structure of conditional generation antagonism network and takes the multi-scale feature fusion network as the generator.3,Since the poor ability of discrimination for color features,it leads to a bad effect of the traditional methods which only use color features for face color classification.In this paper,the color features and the texture features are fused as the face color feature for classification.Eventually,we achieve a good performance of face color classification.