|
Abstract
|
In recent years, deep learning methods have become one of the most popular algorithms in hyperspectral anomalydetection (HAD), and their detection performance is satisfying. However, most approaches focus on the representation of purebackgrounds and ignore the category diversity of backgrounds in a real hyperspectral scene. To some extent, the issue limits thedetection effect. To address this problem, a novel deep feature aggregation network (DFAN) is put forward in this paper, and itdevelops a new paradigm for HAD to represent multiple patterns of a hyperspectral image. The DFAN adopts an adaptiveaggregation model, which combines the orthogonal spectral attention module with the background-anomaly category statisticsmodule. This allows effective utilization of spectral and spatial information to capture the distribution of the background andanomaly. To optimize the proposed DFAN better, we design a novel multiple aggregation separation loss, and it is based on the intra-similarity and inter-difference of the background and anomaly. The constraint function weakens the potential anomalyrepresentation and enhances the potential background representation to improve the separation of background and anomaly.Additionally, the extensive experiment on the six real hyperspectral datasets demonstrates that the proposed DFAN achievessuperior performance for HAD
|