Selfie Facial Acne Grading Via Cross-domain Alignment And Tone-sensitive

With the development of deep learning,computer-aided medical technology has shown strong vitality in disease diagnosis.Acne is the most common skin disease,and accurate grading of acne severity plays a critical role in assisting patients with precise treatment.Therefore,deep learning-based facial acne grading tasks have been widely studied.However,most existing studies focus on clinical scenarios with high-quality images.Unlike clinical scenarios,acne image data in selfie scenarios are scarce and of poor quality.Specifically,the research on facial acne grading based on selfie scenarios faces the following important challenges: first,there is no publicly available dataset of facial acne severity under selfie scenarios;second,the data size of facial acne grading-related datasets is limited;third,acne grading methods have not been based on skin color perception modeling.To address these issues,this paper conducts the following research:(1)To address the lack of acne severity grading datasets in selfie scenarios,this thesis collects three datasets for facial acne grading in selfie scenarios.Firstly,a fullface image dataset called "Acnehdu" is collected,which includes images of different skin tones and angles of shooting.Secondly,in order to reduce the risk of privacy leakage,two face block image datasets,"Acnehdu P" and "Acne PGP",are constructed by combining key point detection algorithms,which only contain facial blocks without key information such as human eyes.Compared with existing acne grading datasets,all three datasets contain information on the severity level and location of acne lesions in the images.(2)To address the limited sample size of acne datasets in selfie scenes,this paper proposes a cross-domain alignment-based acne grading model.An additional highquality clinical acne dataset is introduced,and a generative adversarial network is used for cross-domain data augmentation to expand the training data.There are distribution differences between the selfie dataset and the clinical dataset caused by factors such as shooting angle and image background.Therefore,domain adaptation techniques are used to reduce the differences between the two data domains and improve grading performance in selfie scenes.(3)To address the issue that the acne grading method is not based on skin tone modeling,this paper proposes two acne grading models that incorporate skin tone perception based on the cross-domain alignment of acne grading models.First,the dataset is divided according to skin tone categories,and corresponding sub-models are trained for each skin tone.To eliminate the influence of skin tone changes due to cross-domain data augmentation,a skin tone-independent universal sub-network is introduced,and weighted fusion is performed with the skin tone sub-networks to eliminate errors.Second,expert gate networks and skin tone gate networks are introduced.Multiple expert sub-networks are used to extract common features of skin tone,and weighted fusion is performed through corresponding expert gate networks.Multiple skin tone sub-networks are used to learn the underlying correlations between different skin tones,and weighted fusion is performed through corresponding skin tone gate networks for adaptive skin tone differences.In summary,this thesis proposes a selfie facial acne grading model based on cross-domain alignment and skin tone perception and demonstrates excellent performance on three selfie acne datasets collected in this paper.The cross-domain alignment method proposed in this paper has significant research significance in guiding the learning of unlabeled data with labeled data,and the skin tone perception module proposed guides the reverse research of skin modeling.In addition,the facial acne grading datasets constructed in the selfie scenarios provide a data guarantee for subsequent research on selfie acne grading.