Image segmentation using sparse logistic regression with spatial prior

We propose a supervised learning based method for image segmentation. The method is based on feeding artificial features to a framework of logistic regression classifier with ℓ1 norm regularization and Markov Random Field prior, which has been studied, e.g., in hyperspectral image segmentation. The novelty of our approach stems from the use of a generic artificial feature set and the embedded feature selection property of the sparse logistic regression framework, which avoids application specific feature engineering. The proposed method generates a large set of artificial features and passes them to a ℓ1 regularized logistic regression classifier. Finally, a spatial prior imposes additional homogeneity to the classification result. We study the performance of the proposed method for two application cases, and show that the segmentation results are accurate even with simple models with high degree of sparsity.