Abstract: To improve the clarity of the haze image and realize the generalization of the dehazing ability, a network structure with feature enhancement and multi-scale loss constraint is proposed, which is trained by an incremental training method. The network consists of a teacher network and a student network. The student network enhances the features by learning the attention information of the labeled samples extracted by the teacher network and uses the multi-scale semantic features of the labeled samples as the soft targets. Besides, a multi-scale semantic feature loss measurement mechanism is established, which cascades with the global pixel difference loss to construct the loss function for feature and pixel levels. According to the incremental training method, the teacher network guides the student network to balance the relationship between the new and old knowledge of different datasets. Thus, the network can improve the generalization ability of the supplementary dataset quickly, while retains the original knowledge. Experiment results show that the proposed algorithm performs well in both subjective visual effects and objective evaluation indicators.

Key words: image dehazing, feature enhancement, multi-scale loss constraint, incremental training