Anomaly detection and localization are important problems in computer vision. Recently, Convolutional Neural Network (CNN) has been used for visual inspection. In particular, the scarcity of anomalous samples increases the difficulty of this task, and unsupervised leaning based methods are attracting attention. We focus on Student-Teacher Feature Pyramid Matching (STPM) which can be trained from only normal images with small number of epochs. Here we proposed a powerful method which compensates for the shortcomings of STPM. Proposed method consists of two students and two teachers that a pair of student-teacher network is the same as STPM. The other student-teacher network has a role to reconstruct the features of normal products. By reconstructing the features of normal products from an abnormal image, it is possible to detect abnormalities with higher accuracy by taking the difference between them. The new student-teacher network uses attention modules and different teacher network from the original STPM. Attention mechanism acts to successfully reconstruct the normal regions in an input image. Different teacher network prevents looking at the same regions as the original STPM. Six anomaly maps obtained from the two student-teacher networks are used to calculate the final anomaly map. Student-teacher network for reconstructing features improved AUC scores for pixel level and image level in comparison with the original STPM.