detection technology based on deep learning has shown excellent performance in the field of object detection; however, it has not yielded expected results when used for synthetic aperture radar (SAR) ship detection. Herein, an SAR ship detection method based on a convolutional neural network is proposed for multiscale ship detection in multiple scenarios. Based on the single shot multiBox detector, we use Darknet-53 as the feature extraction network. A deep feature fusion network is added to generate new feature prediction maps with rich semantic information. In addition, we use a new two-class loss function in the training strategy to deal with the imbalance in the difficult and easy samples in the training process. The verification experiments are performed on the expanded public SAR ship detection dataset. The experimental results indicate that our proposed method has a good adaptability to SAR ship detection at different sizes in complex scenes.