Dimensionality Reduction based on Spatial Features for Efficient Multivariate Image Classification

Abstract: Multivariate imaging advanced in recent years which prompted many applications for detailed understanding in the fields of satellite imaging, medical imaging, and microscopic imaging. To achieve more insights about it, various feature extraction techniques exist which utilize the ample spectral and spatial details in an image. But apart from feature extraction dimensionality reduction (DR) and efficient classification has become a key aspect in multivariate image analysis (MIA). Adding more and more variables in feature space of multivariate image results into high dimensionality which in turn increases the complexity in classification. Therefore, it becomes important to apply DR techniques before classification process. Most widely used DR method is Principal component analysis (PCA) which is linear DR method. The main disadvantage of PCA is that it does not consider the nonlinearity in data. The proposed new method is invariant to nonlinearity in data. To consider nonlinearity, Gabor filter is used to extract spatial features from multivariate data. Gabor filter based method performs dimensionality reduction of nonlinear multivariate images while improving the classification accuracy. Keywords - Multivariate image analysis (MIA), dimensionality reduction (DR), Gabor filter, Support Vector Machine (SVM), Convolutional Neural Network (CNN)