An LBP-HOG Descriptor Based on Matrix Projection For Mammogram Classification (1904.00187v4)

Abstract: In image based feature descriptor design, local information from image patches are extracted using iterative scanning operations which cause high computational costs. In order to avoid such scanning operations, we present matrix multiplication based local feature descriptors, namely a Matrix projection based Local Binary Pattern (M-LBP) descriptor and a Matrix projection based Histogram of Oriented Gradients (M-HOG) descriptor. Additionally, an integrated formulation of M-LBP and M-HOG (M-LBP-HOG) is also proposed to perform the two descriptors together in a single step. The proposed descriptors are evaluated using a publicly available mammogram database. The results show promising performances in terms of classification accuracy and computational efficiency.