GCN-YOLO: YOLO Based on Graph Convolutional Network for SAR Vehicle Target Detection

Abstract

Recently, deep convolutional neural networks have been widely applied in target detection of synthetic aperture radar (SAR) images. However, the regular convolution kernel cannot effectively establish dependency between features of SAR image with geometric distortion. Meanwhile, SAR images contain a small number of vehicle targets, and the imbalance problem between foreground-background class is serious during training. To solve these problems, we propose a you only look once (YOLO) detector based on graph convolutional network (GCN) called GCN-YOLO. First, a multilayer GCN model called vision GNN (ViG) is used as feature extractor to model the local area and build long-term dependencies between features. In addition, a convolutional block attention module (CBAM) is embedded into the last layer to enhance semantic features. Then, we introduce the VariFocal loss (VFL) as confidence loss to relief the imbalance problem between positive and negative samples. The experimental results on the miniSAR data demonstrate the effectiveness of the proposed method.