In industrial detection scenario, according to whether anomalies that do not exist in normal samples are introduced, anomaly detection problems can be divided into two categories: structural anomaly detection and logical anomaly detection. Logical anomaly detection places higher demands on the global understanding ability of the network. Faced with the problem that the existing unsupervised anomaly detection model has a good detection accuracy on structural anomalies, but cannot meet the requirements of logical anomaly detection, a dual autoencoder structure consisting of spatial reunion module and channel reunion module is proposed. Our method consists of three components: Initially, the parallel space channel dual autoencoder architecture is introduced, by obtaining feature vectors containing global information from spatial and channel directions, the long-range dependencies of the network is improved. Secondly a selective fusion module is designed to fuse the information of the dual autoencoder and amplify features containing important information to further improve the ability to express logical anomalies. Lastly cosine loss is proposed to the loss function between autoencoder and student network to avoid the network being sensitive to individual pixel differences, so as to focus on global differences. We conducted experiments on MVTec LOCO AD dataset, and achieved 89.4% in logical anomaly detection accuracy, 94.9% in structural anomaly detection accuracy, and 92.1% in average detection accuracy, surpassing the baseline method and other unsupervised defect detection methods, verifying the effectiveness and superiority of the method.