Connecting The Past And The Present: Histogram Layers For Texture Analysis


Feature engineering often plays a vital role in the fields of computer vision and machine learning. A few common examples of engineered features include histogram of oriented gradients (HOG) (Dalal and Triggs, 2005), local binary patterns (LBP) (Ojala et al., 1994), and edge histogram descriptors (EHD) (Frigui and Gader, 2008). Features such as pixel gradient directions and magnitudes for HOG, encoded pixel differences for LBP, and edge orientations for EHD are aggregated through histograms in order to extract texture information. However, the process of designing handcrafted features can be difficult and time-consuming. Artificial neural networks (ANNs) such as convolutional neural networks (CNNs) have performed well in various applications such as facial recognition (Chan et al., 2015), semantic segmentation (Long et al., 2015), object detection (Zhao et al., 2019) and image classification (He et al., 2016a) with features learned by the network.

In this work, a new histogram layer is proposed to learn histogram features and maximize the performance of ANNs for texture analysis. Current approaches using ANNs, particularly CNNs, or handcrafted features do not perform well for some texture applications due to inherent problems within texture datasets (e.g., high intrinsic dimensionality, large intra-class variations) and limitations in methods that use handcrafted and/or deep learning features. CNNs, as currently constructed, cannot directly model statistical texture information in the data; and statistical texture serves a powerful descriptor that can be used in several applications. The proposed histogram layer can estimate bin centers and widths through 12 the backpropagation of errors to aggregate the statistical texture features from the data while also maintaining spatial information. Two aspects of texture analysis, classification and segmentation, are explored in this work. The improved performance of each network with the addition of the histogram layer demonstrates the potential for the use of this new element within ANNs



J. Peeples, "Connecting The Past And The Present: Histogram Layers For Texture Analysis", Ph.D. Thesis, Gainesville, FL, 2022. 
author = {Joshua Peeples},
title = {Connecting the Past and the Present: Histogram Layers for Texture Analysis},
school = {Univ. of Florida},
year = {2022},
address = {Gainesville, FL},
month = {March},