Abstract: A building detection algorithm in SAR images based on a lightweight network is proposed to address the problems caused by the large quantity of model parameters and memory resources involved in deep convolutional neural networks. First, the algorithm is based on the rotating detector R-centernet, and the traditional convolution in backbone network is replaced by Ghost convolution. Then, the Ghost-ResNet is constructed to reduce the number of model parameters. Second, a channel attention module that fuses width and height information is developed to enhance the network's ability to locate significant regions in the images accurately. An up-sampling method named CARAFE is used to replace the DCN module in the network, and the feature map information is fully combined in the up-sampling process to improve target detection. Finally, the improved R-centernet is used to train and test the SAR rotating building dataset. Based on the experimental results, compared with the R-centernet, the improved algorithm has increased detection accuracy by 3.8%, recall by 1.2%, and detection speed by 12 frames per second.

Key words: lightweightnetwork, buildingdetectioninSARimage, rotatingtargetdetection, Ghostconvolution, channelattention, CARAFEup-sampling