Unsupervised anomaly detection in real industrial scenarios is challenging since the small amount of defect-free images contain limited discriminative information, and anomaly defects are unpredictable. In this paper, a dual-siamese network is designed to simultaneously detect and locate anomalies in images. It first uses a pre-trained convolutional neural network (CNN)-based siamese architecture to embed discriminative features of normal samples and synthetic defective samples. A dense feature fusion (DFF) module is employed to obtain the dense feature representation of dual input. The following siamese network of perceptual defects is proposed to reconstruct and restore the dual-dense features of the previous stage. Compared to the existing methods that only employ a single residual map, the restoration of dense feature maps is proposed to locate the anomalies better. The experimental results on the MVTec AD dataset demonstrate that our method achieves state-of-the-art inspection accuracy and has potential for industrial application.