A novel 3D shape recognition method based on double-channel attention residual network

Learning 3D features by deep networks has achieved a successful performance up to now. However, data imbalance and low-resolution voxels still remain and influence the performance of 3D shape recognition. To resolve these issues, we propose double-channel attention residual network (double-RVCNN) as a novel deep network model with residual structure based on multi-head self-attention mechanism. Double-channel structure adopts double channels to input data including voxels and 3D Radon feature matrices, aiming to fully utilize the local and global features. The multi-head self-attention mechanism can integrate the relatively important contents of the input data through multiple heads structure, which can enrich the information processing ability and stabilize the training process of our network. Residual structure with cross-entropy loss and center loss as weighted loss function can avoid information loss to a great extent. Experimental results show that the values of mean average precision (MAP) are 83.31% and 74.04%, the values of classification accuracy are 90.53% and 85.09% on ModelNet10 and ModelNet40 datasets respectively, which demonstrates that our method performs a better 3D shape recognition accuracy than compared methods on test datasets.