Subsurface Scattering#

Harnessing coherent illuminator properties to detect subsurface scattering

The coherent light output by a laser yields unique interference properties in comparison to incoherent light sources. This project harnesses these unique properties to discern bare skin from other categories present on the face within the constraints defined by a driver monitoring software (DMS) application context. More specifically, the project uses various laser speckle imaging techniques on vertical-cavity surface-emitting laser illuminated images to perform clustering of bare skin. Three distinct methods are investigated: beam profile analysis, laser speckle variation analysis and laser speckle contrast imaging (LSCI). All methods are evaluated and a comparison is made highlighting the respective benefits and drawbacks of each technique in a DMS application context. In the final implementation, a convolutional neural network is trained to use temporal LSCI processed images to effectively classify skin. A comparison of models trained on light-emitting diode (LED) versus VCSEL illuminated datasets is presented and a consistent improvement in classification performance is demonstrated for the VCSEL models. Finally, a comprehensive evaluation is provided into the limitations of this model.