Self-intersection Attention Pooling Based Classification for Rock Recognition

Rock recognition is a critical step for intelligent control system in tunnel boring machine (TBM). We build the rock recognition dataset by the visual monitoring system in TBM. After that, we present an attention-fusion network (AFN) for rock recognition classification. Our model is trained with the recognition object (rock) image, and gets the rock attention map through the self-intersection operation. To complete the classification task, we propose the attention block in the attention-fusion classifier, which is effective to fuse and encode the feature and attention in an end-to-end manner. We confirm the effectiveness of the self-intersection operation and attention block by visualized experiment and contrast experiment. At the same time, experimental results of the rock recognition dataset indicate that AFN outperforms state-of-the-art methods, including LSTM (mask-guided), ABN (attention-based), ResNetSO (global-feature), both in accuracy and realtime on the rock recognition tasks.