Design And Implementation Of Facial Skin Recognition Algorithm Based On Improved U-Net Image Semantic Segmentation

With the advancement of science and technology and the rapid development of medical cosmetology,the use of cosmetics has expanded rapidly.Among them,facial skin care products are the most popular products.However,different people have different skin conditions,and there are thousands of skin care products to choose from in the market.Consumers are prone to confusion and fatigue when choosing,and using inappropriate skin care products can also damage the skin.The common problems of facial skin are pores,spots,moles,oil,etc.Knowing your skin quality and using skin care products correctly is of great significance to skin care.At present,the evaluation of facial skin quality is still relatively backward,lacking professional diagnosis,which mainly relies on the professional knowledge and experience of physicians or beauticians to analyze.Inspired by the method of medical image segmentation,this paper adopts the idea of segmentation to perform semantic segmentation of human face and skin to detect and locate objects quickly.An improved model based on U-Net is proposed,and the specific research contents are as follows:(1)The current research on facial skin quality detection is still in its infancy,and there is no publicly available skin image dataset at home and abroad.First,the face data of 247 volunteers were collected by camera;then,the deep learning-based segmentation method batched the face data to extract the cheek parts and protect the privacy of the volunteers.Finally,the pores and moles on the cheek were analyzed.and other parts are manually annotated to produce a complete skin texture dataset including source images,source image annotations,and annotation masks.In addition,in order to meet the data set requirements of deep learning,a data expansion strategy was used for the label-based data,and finally 10,500 pairs of skin texture data sets were obtained.(2)For skin quality detection,which is a small target segmentation task,using conventional downsampling and convolution methods will lead to a large amount of information loss,and face the problem that the segmentation target is similar to the background,and the feature discrimination is insufficient.In this paper,an improved face skin texture recognition method(SSUNet)is proposed.In the coding part of this method,a Swin transformer block with long-term dependency characteristics is introduced as the backbone of the coding part to ensure that rich and highly discriminative feature information can be obtained in the shallow layer for subsequent segmentation.A global multi-scale feature extraction module is introduced at the bottom layer of the network to ensure that the feature information input into the upsampling recovery structure is complete.In the jump-connected part of the original U-Net,a channel transformation module is designed to convert the texture features of the shallow layers and then combine them with the semantic features of the deep layers,thereby further improving the segmentation accuracy.Through comparative experiments with four classic semantic segmentation networks such as U-Net and Deep Lab V3.The results show that compared with the classical semantic segmentation algorithm,the SSUNet segmentation model based on the improved U-Net has achieved a significant improvement in the recognition effect of facial skin pores.Among them,the accuracy rate increased by 0.12%～0.57%,the precision rate increased by8.22%～20.13%,the recall rate increased by 12.53%～24.889%,the F1 increased by11.01%～23.46%,and the average intersection ratio increased by 12.39%～24.94%.And the ablation experiment is carried out to prove the effectiveness of the improved module in this paper.(3)Quantify the skin pore information in the face image to make the evaluation result more scientific and reference value.The skin texture is divided into seven categories,and users can intuitively understand the skin texture of their face through the semantic segmentation renderings.At the same time,in order to fully verify the effectiveness and strong generalization performance of the deep learning model proposed in this paper,the public data set "ISIC2018" is used for experimental verification and compared with the experimental results of other literatures.The experimental results show that the Dice coefficient of the method proposed in this paper reaches 90.7%,the accuracy rate reaches 97.4%,and the sensitivity reaches 92.6%,which are better than the experimental results of other methods.