LAUSR

Distributed tactile pressure recognition is an important task in intelligent industry. In this article, convolutional neural network (CNN) has attracted attention for their ability to automatically extract features and complete recognition. However, there are still the following weaknesses. The pressure information is low-resolution based on sensor density, resulting in limited configurable network depth. The current CNNs are deep network for high-resolution image, so they are not applicable for pressure task. What CNN has learned from the pressure is not clear. Therefore, a bilinear feature fusion CNN (BF-CNN) is proposed. The bilinear features are used to improve the feature extraction capability by depicting second-order characteristics, and fusion strategy is applied to promote the efficiency of feature utilization through the complementarity of hierarchical features. To understand BF-CNN intuitively, we develop the fused gradient class activation map (F-Grad-CAM) algorithm to visualize the regions of interest of the samples in detail, using multi-scale features to provide high-resolution heat maps. We also construct a pressure image dataset of 26 (A-Z) letter-shapes with complex contours collected by a 16 × 16 pressure array to validate the proposed models. The results show our work achieved highest accuracy of 98.23% and provided a more detailed location on activated areas.