Hyperspectral anomaly detection via memory?augmented autoencoders

Recently, the autoencoder (AE) based method plays a critical role in hyperspectral anomaly detection domain. However, due to strong generalised capacity of AE, abnormal samples are usually reconstructed well along with normal background samples. Thus, order separate anomalies from by calculating reconstruction errors, it can be greatly beneficial reduce AE capability for sample while maintaining performance. A memory-augmented (MAENet) is proposed address this challenging problem. Specifically, MAENet mainly consists an encoder, memory module, and decoder. First, encoder transforms original data into low-dimensional latent representation. Then, representation utilised retrieve most relevant matrix items matrix, retrieved will used replace encoder. Finally, decoder reconstruct input using items. With strategy, still cannot. Experiments conducted on five real sets demonstrate superiority method.