To reproduce the tactile perception of multiple contacts on the human tongue surface, it is necessary to use a pressure measurement device with high spatial resolution. However, reducing the size of the array sensing unit and optimizing the lead arrangement still pose challenges. This article describes a deconvolution neural network (DNN) for improving the resolution of tongue surface tactile imaging, which alleviates this tradeoff between tactile sensing performance and hardware simplicity. The model can work without high-resolution tactile imaging data of tongue surface: First, in the compression test using artificial tongues, the tactile image matrix (7 × 7) with low resolution can be acquired by sensor array with a sparse electrode arrangement. Then, through finite element analysis modeling, combined with the distribution rule of additional stress on the two-dimensional plane, the pressure data around the existing detection points are calculated, further expanding the tactile image matrix data amount. Finally, the DNN, based on its efficient nonlinear reconstruction attributes, uses the low-resolution and high-resolution tactile imaging matrix generated by compression test and finite element simulation, respectively, to train, and outputs high-resolution tactile imaging information (13 × 13) closer to the tactile perception of the tongue surface. The results show that the overall accuracy of the tactile image matrix calculated by this model is above 88%. Then, we deduced the spatial difference graph of the resilience index of the three kinds of ham sausages through the high-resolution tactile imaging matrix.
Keywords: deconvolution neural network; finite element analysis; food oral processing; food texture; tactile image; tongue.
© 2023 Wiley Periodicals LLC.