A Noise-robust Feature Fusion Model Combining Non-local Attention for Material Recognition

Abstract

Material recognition, as an important task of computer vision, is hugely challenging, due to large intra-class variances and small inter-class variances between material images. To address those recognition problems, multi-scale feature fusion methods based on deep convolutional neural networks are presented, which has been widely studied in recent years. However, the past research works paid too much attention to the local features of the image, while ignoring the non-local features that are also crucial for fine image recognition tasks such as material recognition. In this paper, Non-local Attentional Feature Fusion Network (NLA-FFNet) is proposed that combines local and non-local feature of images to improve the feature representation capability. Firstly, we utilize the pre-trained deep convolutional neural network to extract the image feature. Secondly, a Multilayer Non-local Attention (MNLA) block is designed to generate a non-local attention map which represents the long-range dependencies between features of different positions. Therefore, it can achieve stronger noise-robustness of model and better ability to represent fine features. Finally, combined our Multilayer Non-local Attention block with bilinear pooling which has been proved to be effective for feature fusion, we propose a deep neural network framework, NLA-FFNet, with noise-robust multi-layer feature fusion. Experiment prove that our model can achieve a competitive classification accuracy in material image recognition, and has stronger noise-robustness at the same time.