Patch-local self-similar network for ground terrain recognition

We proposed a patch-local self-similar network (PLoSNet) to exploit the inherent correlations between the patch-level features and the local features among the ground terrain texture for ground terrain recognition. Firstly, the early features extracted from the shallow layer of the backbone network were enhanced by using a texture statistical information enhancement module (TSIEM). Then we adopted the patch extraction method to generate patch-level features on the enhanced features. Next, the patch-level features and the local feature extracted from the last layer of the backbone network were fused by patch-level and local feature fusion module (PLFFM). Finally, the two modules served as a feature encoding layer could be easily integrated into the convolutional neural networks to achieve end-to-end training. In this paper, ResNet-18 was used as the backbone network, and the results showed the proposed PLoSNet has a superior performance over existing approaches on three challenging ground terrain/material recognition datasets.