SARLBP and STMO-GA : two novel description and selection approaches for challenging feature classification problems

Texture analysis and classification is a much-researched area due to its significance within computer vision and pattern recognition applications. Broadly speaking, two approaches for texture classification are found in the literature today: Classical and Deep Learning (DL) based. Typical Deep Learning-based methods extract millions of features thru a training process that requires a very large number of annotated images to generate a sequence of filters which can then be applied to the raw images. This process can pose an issue in problems with limited training data. In contrast, Classical methods require experts to craft a set of features using their knowledge of the problem. This requirement can impose strict limitations on the types of datasets and hence, the problems for which the system is designed. In that sense, in order to be effective, Classical methods require: 1) the design of a comprehensive set of scale-invariant features, so it can capture a wide range of textures; 2) an efficient feature selection algorithm to avoid redundancy and waste of learning resources; and 3) the use of a powerful classification paradigm. In this research, we focused on Classical methods for texture classification, resulting in three major contributions. First, we propose a novel texture feature descriptor called Scale Adaptive Robust LBP (SARLBP) that enhances feature description both at macro and micro levels. This is accomplished by extracting information from very small to very large scales, while using a novel encoding scheme that dynamically determines the optimal scale for each radial direction based on the characteristics of the local and global surrounding areas in the image. Indeed, SARLBP overcomes the limitations in representation space and richness of information of other descriptors. Second, we present STMO-GA, for Sequential Transfer with Multi-Objective Genetic Algorithm, a sequential transfer approach combined with a multi-objective based genetic algorithm for feature selection. This method effectively selects a subset of features with minimal size while it maintains the optimal amount of information necessary for accurate classification. The method proved to be particularly effective for extreme problems in machine learning and pattern recognition involving very small instance datasets containing very large numbers of features. Finally, we propose the use of an Ensemble of Support Vector Machine (SVM) weaker learners to address the problem in classification of continuous classes -- e.g. to detect the percentage of crop residue in farming fields.

Name thesis:degree_name

Ph. D.

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Electrical and computer engineering (MU)

Grantor dc:publisher

University of Missouri--Columbia

Year dc:date.issued

2024

Author dc:creator

• Upadhyay, Parth Chandrakantbhai

Advisor dc:contributor.advisor

• DeSouza, Guilherme

Language dc:language.iso

eng, English

DOI

https://doi.org/10.32469/10355/104751

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/104751