Interleaved group convolution network for hyperspectral image classification

Researches have shown that using convolution neural network (CNN) on spatial-spectral domain can improve the performance of hyperspectral image (HSI) classification in recently years. However, due to the existence of spectral redundancy and the high dimensional kernels used in 3D-CNN, the HSI classification models are often heavy with a huge number of parameters and high computation complexity. Motivated by the lightweight model, this paper introduced a modular convolution structure named three-dimensional interleaved group convolution (3D-IGC). This structure contains two successive group convolutions with a channel shuffle operation between them. First group convolution extracts feature on spatial-spectral domain. Then the channel shuffle enables cross-group information interchange. After this, the second group convolution perform the point-wise convolution. We proved that an IGC is wider than a normal convolution in most cases by inferred formula. The empirical results demonstrate that the increment of width in 3D-IGC model is beneficial to HSI classification with the computation complexity preserved, especially when the model has fewer parameters. Compared with the normal convolution, the 3D-IGC can largely reduce the redundancy of convolution filters in channel domain, which greatly decreases the number of parameters and the computation cost without losing classification accuracy. We also considered the effects of the 3D-IGC on deep neural networks, therefore we used the 3D-IGC to modify the residual unit and get a lightweight model compared with ResNets.