State-of-the-art model-driven Direction-Of-Arrival (DOA) estimation methods for mul-tipath signals face great challenges in practical application because of the dependence on the precise multipath model.In this paper,we introduce a framework,based on deep learning,for synchroniz-ing perturbation auto-elimination with effective DOA estimation in multipath environment.Firstly,a signal selection mechanism is introduced to roughly locate specific signals to spatial subregion via frequency domain filters and compressive sensing-based method.Then,we set the mean of the cor-relation matrix\'s row vectors as the input feature to construct the spatial spectrum by the corre-sponding single network within the parallel deep capsule networks.The proposed method enhances the generalization capability to untrained scenarios and the adaptability to non-ideal con-ditions,e.g.,lower SNRs,smaller snapshots,unknown reflection coefficients and perturbational steering vectors,which make up for the defects of the previous model-driven methods.Simulations are carried out to demonstrate the superiority of the proposed method.