SCALE-AWARE MEMORY-AUGMENTED AUTOENCODER FOR HYPERSPECTRAL SUB-PIXEL ANOMALY DETECTION

Hyperspectral anomaly detection (AD) exhibits promising performances in hyperspectral image processing, and autoencoder (AE) has shown great potential with its network more challenging to reconstruct the anomaly. Unfortunately, as AD algorithm automatically detects abnormal targets that are significantly different from the background, abnormal targets which are less than one pixel on the image, mixing with the features of the background, have brought challenges to the tasks and AE also faces the problem of the increasing ability to reconstruct the anomaly. This article proposes a scale-aware memory-augmented AE(SMAE) to resolve the above two problems via sending the different scales features in the encoder to the corelative decoder by the memory module. Given an input, SMAE firstly obtains the embedding from the encoder and then uses it as a query to combine the most relevant memory items to retrieve the new features, which occurs in all depth of the network, prompting the ability to process diverse size targets, and memory-augmented design can strengthen the reconstructed errors for anomaly detection also increasing its sensitive to detect the sub-pixel anomaly targets. Experiments on various datasets prove the high precision of our proposed method.