The Bohai Sea is an important economic zone of China. Sea ice has been a significant threat to the human activities around the Bohai Sea. As the imaging capability of synthetic aperture radar (SAR) is independent of sun illumination and cloud condition, it is of great significance to detect the sea ice of the Bohai Sea from SAR images. Due to the limitation of the feature extraction mechanism, the accuracies of traditional sea ice detection methods need to be improved. Deep learning has a strong self-learning ability and is suitable for image detection. Here, we employ the well-known deep learning framework, U-Net, as the basic structure, and design a hybrid loss function to optimize the U-Net model, forming a hybrid loss U-Net model for sea ice detection in the Bohai Sea. The Sentinel-1 dual-polarization (VV and VH) SAR images are the inputs of the model. We compare the hybrid loss U-Net model with several traditional methods (Pulse Coupled Neural Network, Markov Random Field and Watershed Algorithm) and deep learning method based on CNN. Experiments show that the hybrid loss U-Net-based model achieves 97.567%, 98.769%, 98.767% and 98.771% in IoU, F1_Score, Precision and Recall respectively, outperforming the other methods. Compared with VV single-polarized input, the detection results of dual-polarized information input are 0.375%, 0.111%, 0.639% and 0.740% higher in F1_Score, Precision, Recall and IoU respectively. The detection results of the hybrid loss model are 1.129%, 0.947%, 1.794% and 2.231% higher than those of the non-hybrid loss function in F1_Score, Precision, Recall and IoU respectively. The model could effectively detect details such as ice water line, inter-ice water and ice gap. Our model is applied to detect the sea ice of a whole SAR image in the Bohai Sea, which can provide technical supports for sea ice monitoring, sea ice change analysis and sea ice prediction.