| Summary: | The convolutional neural network (CNN) has a poor performance in nonuniform and edge regions due to the limitation of fixed receptive field. At the same time, feature stacking of input data can bring burden and overfitting to the network. To solve these problems, this article proposes a reg-superpixel guided CNN based on feature selection and receptive field reconstruction. First, a feature selection method is designed, which uses polarimetric SAR statistical distribution features to calculate distance and similarity, and selects features that are easier to identify to avoid the negative impact of low distinguishing features on classification. Second, the reg-superpixel, which means regular superpixel, is used to reconstruct the receptive field and represent the features of the central pixel. The classification result of the central pixel is extended to the whole superpixel during the test. This method can extend the pixel-level CNN network to superpixel-level. Finally, by using edge information of the small-scale superpixel and spatial information of the large-scale superpixel to adjust receptive field of the central pixel, the classification results with uniform smooth region and high edge fitting can be generated. Experimental results with four state-of-the-art methods on four datasets show that feature selection and multiscale reg-superpixel network is effective for polarimetric SAR classification problems.
|
|---|