Pattern Recognition Approach Using Modified Convolutional Neural Network with Data Texture Excitation

Convolutional neural networks (CNNs) have attracted wide attention in the field of pattern recognition because of their local feature extraction ability. However, fine-pattern recognition remains challenging when similar samples originate from different inter-classes. To address this issue, a texture excitation CNN (TE-CNN) is proposed. First, feature maps are extracted by a standard CNN framework. Second, a microscopic texture feature inspired by human visual perception is parameterized in parallel. Finally, the CNN feature maps are actuated by the texture feature to improve the separability of similar inter-class samples. Thus, TE-CNN enhances fine recognition from a more microscopic perspective than standard CNNs, without significantly increasing model complexity. The proposed method was evaluated on a benchmark bearing fault dataset and a real near-infrared spectroscopy (NIRS) dataset.