Multi-scale contextual information convolutional neural network for structures detection in high-resolution remote sensing image

Structures in high-resolution remote sensing images exhibit multi-scale characteristics. The detection network designed for large-scale objects directly applied to multi-scale structures such as wind turbines and transmission towers does not yield satisfactory results. To solve this problem, this paper proposes a structures detection network using multi-scale context information. Firstly, the VGG network is employed to generate multi-scale feature maps. Medium-sized feature maps are selected, because them contain more information about structures. To strengthen the location information of structures, Convolutional Block Attention Module is used in branches. Then, the feature maps at different scales are contacted together in the channel dimension to integrate multi-scale information. And small size feature maps are not conducive to the detection of structures, so the size of feature maps are increased through deconvolution. Finally, the network can be enhanced the learning of location information by Location Enhance Module. Then, the feature maps are input into RPN network to get the suggested region. The results show that this network has higher accuracy in structure detection compared with Faster Rcnn in remote sensing images.