Feature pyramid network based on double filter feature fusion for hyperspectral image classification

Abstract

Deep learning has been widely used in hyperspectral image (HSI) classification due to its powerful feature extraction ability. However, most methods ignore the HSI features at the different image scales and lack shallow spectral-spatial information. For the above issues, the paper proposes an HSI classification method based on the feature pyramid network of double-filter feature fusion (DF3-FPN). Specifically, the feature pyramid is a multi-scale feature network with residual blocks as the backbone. It can not only learn the HSI features of different scales but also can mine the effective information that is easily missed between HSI scales by embedding the double-filter feature fusion module. And it could enrich the feature expression of HSI by interacting with the features of each scale. Next, a spectral-spatial fusion module is proposed to obtain shallow HSI features (including position, edge, etc.). Additionally, shallow information is added to the multi-scale features extracted by the feature pyramid network to make up for the lack of shallow information. Finally, the proposed DF3-FPN network can further boost the HSI classification performance. Besides, experimental results show that the proposed method achieves better classification performance on three benchmark datasets.