Because of the influence of clutter and interference, intelligent detection is of great importance for high-frequency surface wave radar (HFSWR) in complex environment. In this paper, we propose a cascade HFSWR target detection strategy based on semi-supervised self-distillation ( $\sf S^{3}D$ ) learning algorithm. First, based on the results of constant false alarm rate (CFAR) detection, the $\sf S^{3}D$ algorithm proposed in this paper further classifies the real and other targets. Moreover, the non-maximum suppression (NMS) is combined to remove redundant target boxes, in order to achieve accurate location. For supervised learning, the classification needs to label all the samples, which is extremely difficult and time consuming, especially for HFSWR data under complex environment. While the $\sf S^{3}D$ algorithm can make use of both labeled and unlabeled samples. For labeled data, labels are directly used to restrict both the deep and shallow layers of the network, so that both layers represent the same semantic categories. For unlabeled data, the pseudo-label is retained only if the deepest layer produces a high-confidence prediction. The deep and shallow layers are then updated to predict the pseudo-label when fed a strongly augmented version of the same unlabeled samples. The above multiple consistency regularization methods can improve the generalization performance of the network effectively. Finally, the simulations conducted across public datasets show that compared with other semi-supervised learning methods, $\sf S^{3}D$ can better improve the performance of network generalization. Besides, the $\sf S^{3}D$ -based HFSWR target detection method can significantly improve the vessel detection performance by using only a minority of sample labels.