Unsupervised feature learning based on sparse coding and spectral clustering for segmentation of synthetic aperture radar images

Synthetic aperture radar (SAR) image segmentation is fundamental for the interpretation and understanding of these images. In this process, the representation of SAR image features plays an important role. Spectral clustering is an image segmentation method making it possible to combine features and cues. This study presents a new spectral clustering method using unsupervised feature learning (UFL). In this method, the SAR image is primarily processed by the non-negative matrix factorisation (NMF) algorithm and then non-negative features containing spatial structure information are extracted. Afterwards, the extracted features are learned using a sparse coding algorithm to increase the discrimination power of the features. Sparse coding is an unsupervised learning algorithm which finds the patterns or high-level semantics of the data. Ultimately, the SAR image segmentation operation is performed by applying spectral clustering on learned features. In this method, sparse coding learns features and simultaneously creates the similarity function required in spectral clustering through the production of sparse coefficients. Therefore this method avoids the Gaussian similarity function, which has a problem with scale parameter adjustment that is one of the drawbacks of spectral clustering methods. The results demonstrate that, compared with wavelet and GLCM features, NMF features manage to obtain more meaningful information and provide a better SAR image segmentation result. The results have also demonstrated that SAR image segmentation using learned features is significantly improved compared with segmentation by unlearned features. The experimental results indicate the effect of UFL on SAR image segmentation.