Spatial-spectral separable convolutional neural network for cell classification of hyperspectral microscopic images

In recent years, with the development of computer science, deep learning plays a critical role in the classification of hyperspectral bloodcell images. However, traditional deep learning models require a large amounts of manually annotated training data, and ignore the nature of "graph-spectral uniformity" property of hyperspectral image. As a result, these methods can not explore the intrinsic information of hyperspectral images. In addition, traditional convolutional neural network methods have too many parameters, which takes a great deal of time to be trained. Aiming at these two shortcomings, a spatial-spectral separable convolutional neural network (S3CNN) is proposed to improve the classification performance of bloodcell hyperspectral image and reduce the complexity of the model.First, due to the spatial consistency of the hyperspectral bloodcell image distribution, a spatial-spectral combined distance (SSCD) was proposed to select the spatial-spectral nearest neighbor of each pixel and expand the training samples. At the same time, in the following neural network model, a group of depth convolution and point convolution are used to replace classical convolution and optimize the complexity of the model.The experimental result on bloodcell1-3 and bloodcell2-2 datasets show that the overall classification accuracies reaches 87.32% and 89.02%, respectively. Compared with other classification algorithms of bloodcells, the proposed S3CNN achieves much higher classification accuracy. The training time of the separable convolution model is 27% less than that of the classical convolution model.Experimental results show that the proposed S3CNN is an effective method to improve the classification performance of hyperspectral bloodcell and reduce model training time. © 2022, Science Press. All right reserved.