Inception Parallel Attention Network for Small Object Detection in Remote Sensing Images

Shuojin Yang, Liang Tian, Bingyin Zhou, Dong Chen, Dan Zhang, Zhuangnan Xu, Wei Guo, Jing Liu

Abstract: Small object detection in remote sensing images is a major challenge in the field of computer vision. Most previous methods detect small objects using a multiscale feature fusion approach with the same weights. However, experiments shows that the inter feature maps and the feature map in different scales have different contribution to the network. To further strengthen the effective weights, we proposed an inception parallel attention network (IPAN) that contains three main parallel modules i.e. a multiscale attention module, a contextual attention module, and a channel attention module to perform small object detection in remote sensing images. In addition, The network can extract not only rich multiscale, contextual features and the interdependencies of global features in different channels but also the long-range dependencies of the object to another based on the attention mechanism, which contributes to precise results of small object detections. Experimental results shows that the proposed algorithm significantly improves the detection accuracy especially in complex scenes and/or in the presence of occlusion.