For unmanned surface vehicles (USV), the ability to detect close-range obstacles is critical during the process of autonomous navigation. However, due to the variety of shapes and sizes of close-range obstacles, it’s difficult for existing methods to maintain high accuracy when detecting close-range obstacles. In response, we propose a network, a close-range obstacle detection network (CRODNet). CRODNet utilizes ResNet to extract local features, and uses attention mechanisms to capture global information. To fuse local features with global information, we propose a module, Feature Reweight Module (FRM), which adaptively preserves pixel values in localized regions of low-level feature images based on the threshold value. And we propose a module, Merge Module (MM), which uses confidence score maps generated from high-level feature maps as weights to fuse different feature images. In addition, we use the weighting of multiple loss functions to strike a balance between predicting classes of pixels and capturing the information of region layout. To validate the accuracy of CRODNet, we use the Maritime Object Detection and Segmentation dataset (MODs), for validation. The experimental results demonstrate that CRODNet achieves excellent performance in detecting close-range obstacles (89.9% F-score), which exceeds the state-of-the-art model by 2.3%.