Abstract: Multi-class object detection of remote sensing images is a challenging subject since objects in remote sensing images are dense,multi-oriented and multi-scale.This paper presents a new framework for object detection in remote sensing images.The framework enhances object information,suppresses non-object information,and improves the ability of feature representation by extracting salient features and the relationship between different convolutional channels.Meanwhile,multi-scale features are added to the convolution module to capture more context information without adding extra parameters to the model.To solve the problem of variable object angle in rencote sensing image,we add angle information to the Region Proposal Network (RPN) to get oriented rectangular object proposals.In the training stage,the attention loss function is added to guide the significance of network learning.The proposed framework is validated on the public remote sensing image data set,and the experimental results on the horizontal boxes task and the oriented boxes task prove the effectiveness of the proposed method.

Key words: Channel self-learning, Multi-scale feature, Object detection, Remote sensing image, Saliency learning