Abstract: In complex media environments, the quality of images produced by ground penetrating radar reflection echoes is poor. These images exhibit issues such as weak saliency of related abnormal regions and difficulty in feature identification, which affects both visual effects and the performance of advanced computer vision tasks. To address these issues, a saliency enhancement algorithm for abnormal regions in ground penetrating radar images based on hybrid supervised learning is proposed in this paper. The algorithm fully utilizes the labels of simulated enhanced images through a semi-supervised sub-network to allow the enhanced network to learn the distribution of simulated images, and uses real images to migrate the network to the distribution of real images. The unsupervised sub-network takes real images as input and uses simulated images as adversarial samples to train the model for saliency enhancement under real images. The two sub-networks can cyclically train the enhanced network, thereby achieving saliency enhancement of abnormal regions in real ground penetrating radar images with a unified training process. Experimental results show that the proposed method can effectively enhance the saliency of abnormal regions in ground penetrating radar images.