Skip to Main content Skip to Navigation
Theses

Human skin characterization and analysis based on hyperspectral reflectance using machine learning

Abstract : Skin, the largest and multi-functional organ of human body, has always been an important research object in many fields, such as cosmetics and computer graphics. Its appearance especially color can reflect certain diseases, such as melanoma and vitiligo, which has been widely investigated. In the past, we obtain the skin physiological information by biopsy. This method is usually invasive and takes long time. Recently, the inner information can be derived non-invasively benefiting from skin hyperspectral diffuse reflectance. However, it is still a challenging task since the accuracy and the efficiency cannot be ensured at the same time in applications. For example, the gold standard Monte Carlo method gives favorable estimations but costs much time. This thesis aims to build a detailed skin model and apply it for faster non-invasive determination of skin components. Moreover, an auxiliary method based on this model can identify the presentation attacks at high precision.Our skin model is composed of three sub-layers: the epidermis, the dermis and the subcutis. We first implement a GPU-based Monte Carlo method to reconstruct a skin diffuse reflectance database based on our skin model. The wavelength is randomly taken in the visible light range from 450 to 700 nm. Then, this database is used for training a forward artificial neural network to map optical parameters calculated from our skin model and skin diffuse reflectance. We compare the skin diffuse reflectance reconstructing capacity of forward network and Monte Carlo method, and find that they match well each other. It takes 19 ms for the forward network to reconstruct a reflectance spectrum for 450 to 700 nm with 1 nm interval. However, it takes 337 s for Monte Carlo method. Besides this, we analyze the impact of each skin parameters on reconstructed reflectance and then apply this forward network combined with a curve-fitting algorithm to extract skin parameters using NIST skin database. The results show that the forward network has acceptable accuracy for melanin, blood and oxygen saturation without limitations to fix the thickness of skin sub-layers. And the forward network method costs an average of 17 seconds to finish the extraction process. An inverse network, random forest and support vector regression are also studied. As shown in previous research, inverse networks have large errors in extracting skin parameters but at extremely high speed. In our research, we generate a skin diffuse reflectance database using proposed forward network instead of Monte Carlo method to reduce time cost. Two types of dimensionality reduction methods: low variance filter and principal component analysis are applied for further speeding up. The experiments show that the inverse net- work works better in extracting melanin content than random forest and support vector regression and has similar results to inverse Monte Carlo. Moreover, it only takes around 10 min to train the inverse network after having used dimensionality reduction methods and 12 ms to extract melanin content for one spectrum.As for the presentation attacks detection, we use two metrics RMSE and STD of fitting performance to classify if the object diffuse reflectance belongs to skin. By selecting the appropriate wavelength range, it has promising classification results. The vulnerability of face recognition system has been discussed a lot while detecting presentation attacks, especially the silicon face masks. Our method uses hyperpspectral reflectance to identify the non-skin objects because the absorption coefficients of several skin pigments are unique. We will collect more hyperspectral skin images and generalize our method for practical applications.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03412032
Contributor : Abes Star :  Contact
Submitted on : Tuesday, November 2, 2021 - 5:34:12 PM
Last modification on : Wednesday, November 3, 2021 - 3:52:58 AM

File

TH_T2753_sli.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03412032, version 1

Citation

Shiwei Li. Human skin characterization and analysis based on hyperspectral reflectance using machine learning. Other. Université de Lyon, 2021. English. ⟨NNT : 2021LYSEC004⟩. ⟨tel-03412032⟩

Share

Metrics

Record views

52

Files downloads

31