Residual Attention Dual Autoencoder for Anomaly Detection and Localization in Cigarette Packaging

Anomaly detection and localization is an important problem in the industrial production process, such as cigarette packaging quality inspection. This task is challenging due to very few abnormal samples. This paper proposes Residual Attention Dual Autoencoder (RADAE) composed of the dual autoencoder and the attention map generator. Without the need of the abnormal training images, the proposed model can achieve anomaly detection and localization by unsupervised learning. The dual autoencoder is an encoder-decoder-encoder network that maps the input image to a lower dimension vector for reconstruction, and remaps the generated image to its latent representation. The data distribution of the normal samples is learned by training the dual autoencoder. Anomaly discrimination is realized by measuring the difference between the latent space of the original image and the generated image during detection. And the attention map generator is a CNN model, which uses the residuals of the original image and the generated image to obtain the attention map under the guidance of the attention loss function. The generated attention map is encouraged to focus on all regions in normal samples during training, and spontaneously focus on abnormal regions to achieve localization during testing. The experimental results demonstrate the reliability and superior performance of the proposed model for anomaly detection and localization in the cigarette packaging dataset.