Image Anomaly Detection by Aggregating Deep Pyramidal Representations

Anomaly detection consists in identifying, within a dataset, those samples that significantly differ from the majority of data, representing normal class. It has many practical applications, e.g. ranging defective product industrial systems to medical imaging. This paper focuses on image anomaly using deep neural network with multiple pyramid levels analyze features at different scales. We propose based encoding-decoding scheme, standard convolutional autoencoders, trained data only order build model normality. Anomalies can be detected by inability reconstruct its input. Experimental results show good accuracy MNIST, FMNIST and recent MVTec Detection dataset.