Uso de rede neural convolucional para identificação de substâncias utilizando imagens de sensores baseados na ressonância de plasmon de superfície

Abstract

Surface Plasmon Resonance (SPR) based sensors are well-established instruments that track the resonance position by the minimum reflectivity value hunting, which obtained by illuminating its multilayer structure. The reflected light is the signal captured by an image detector, which generates an image that can be used directly for sensing or used to generate SPR curves that graphically represent the resonance. The location of the position where the minimum light intensity occurs is called the resonance angle; and is used to classify and recognize substances. The present work creates a predictive model for identifying substances from SPR images. A dataset of simulated SPR images was created to test, validate, and train the VGG16 and VGG19 convolutional neural network (CNN) based models. Experimental tests with aqueous substances of different refractive index were performed. The results demonstrate the viability of the model, being able to reach 99.7% accuracy with the VGG19 architecture.