A light CNN based on residual learning and background estimation for hyperspectral anomaly detection

Existing deep learning-based hyperspectral anomaly detection methods typically perform anomaly detection by reconstructing a clean background. However, for the deep networks, there are many parameters that need to be adjusted. To reduce parameters of network and improve the performance of anomaly detection, a light CNN based on residual learning and background estimation was proposed. Different from traditional methods, the proposed method could directly learn anomaly features rather than background features. First, during the training stage, a background estimation method based on non-central convolution kernels was used to obtain the pseudo-background. Second, to purify the pseudo-background, a pair down-sampling method and a joint loss that combines cross-approximation background loss and consistency loss were proposed. Third, the anomaly matrix was obtained by the difference between the hyperspectral image (HSI) and the pseudo- background. Fourth, a light CNN with three layers was proposed to extract features of the anomaly matrix. Finally, during the prediction stage, anomaly detection results were calculated from the predicted anomaly matrix obtained by light CNN through the Mahalanobis distance. Experiments were conducted with multiple metrics on five real-world datasets. Compared with eight state-of-the-art methods, the proposed method achieved the superior performance in both qualitative and quantitative evaluations.