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

Abstract

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.

Design of New Wavelet Packets Adapted to High-Resolution SAR Images With an Application to Target Detection

“…Target Detection schemes [4] consist in deciding if a target of interest is present at a given position of the image. Segmentation [5], [6], aim's is to delimit the image into segments which are conceptually meaningful such as the boundary between land and sea. Finally classification [7], [8], [9], [10] allows to label part of the images with regards to an application of interest.…”

“…Fusion techniques on wavelet coefficients have been used in [34] in order to compute a change detection map. Wavelet decomposition associated with kurtosis statistics have been exploited in [6], [7] for both segmentation and classification purposes. In [35], [36], multiresolution information is used for target detection schemes in Gaussian context.…”

An Energy-Based SAR Image Segmentation Method with Weighted Feature

“…The differences among homogeneous regions in the high resolution SAR image exist not only in spectral features, but also in structural features such as boundary and texture features [6]. To segment a high resolution SAR image well, multiple structural features can be considered in the segmentation algorithm [7][8][9].…”

Image offset density distribution model and recognition of hand knuckle

“…This method is fast and can extract features better than other methods. Rahmani and Akbarizadeh [16] proposed a spectral clustering method using unsupervised feature learning (UFL).…”