Improved Texture Feature Extraction and Selection Methods for Image Classification Applications

Classification is an important process in image processing applications, and image texture is the preferable source of information in images classification, especially in the context of real-world applications. However, the output of a typical texture feature descriptor often does not represent a wide range of different texture characteristics. Many research studies have contributed different descriptors to improve the extraction of features from texture. Among the various descriptors, the Local Binary Patterns (LBP) descriptor produces powerful information from texture by simple comparison between a central pixel and its neighbour pixels. In addition, to obtain sufficient information from texture, many research studies have proposed solutions based on combining complementary features together. Although feature-level fusion produces satisfactory results for certain applications, it suffers from an inherent and well-known problem called “the curse of dimensionality’’. Feature selection deals with this problem effectively by reducing the feature dimensions and selecting only the relevant features. However, large feature spaces often make the process of seeking optimum features complicated.

This research introduces improved feature extraction methods by adopting a new approach based on new texture descriptors called Local Zone Binary Patterns (LZBP) and Local Multiple Patterns (LMP), which are both based on the LBP descriptor. The produced feature descriptors are combined with other complementary features to yield a unified vector. Furthermore, the combined features are processed by a new hybrid selection approach based on the Artificial Bee Colony and Neighbourhood Rough Set (ABC-NRS) to efficiently reduce the dimensionality of the resulting features from the feature fusion stage.

Comprehensive experimental testing and evaluation is carried out for different components of the proposed approach, and the novelty and limitation of the proposed approach have been demonstrated. The results of the evaluation prove the ability of the LZBP and LMP texture descriptors in improving feature extraction compared to the conventional LBP descriptor. In addition, the use of the hybrid ABC-NRS selection method on the proposed combined features is shown to improve the classification performance while achieving the shortest feature length. The overall proposed approach is demonstrated to provide improved texture-based image classification performance compared to previous methods using benchmarks based on outdoor scene images.

These research contributions thus represent significant advances in the field of texture-based image classification.