Unifying Feature Space from Heterogeneous Images for Unsupervised Change Detection

Domain transformation is playing an increasingly important role in heterogeneous change detection. Existent methods, however, cannot guarantee the consistency of reconstructed feature spaces. To solve this issue, we propose an unsupervised change detection method based on the unification of feature space (UFS). First, we construct a convolutional autoencoder based on adaptive instance normalization to unify feature space. Then, we use fuzzy local information c-means to reduce the over-reliance on reconstructed features. Finally, we design a dynamic superpixel-based label assignment (DSLA) rule to increase the number of reliable pseudo-labels used to learn a binary classifier to obtain the CD results. Experimental results on heterogeneous datasets demonstrate the effectiveness of UFS.