Genetic programming-based learning of texture classification descriptors from Local Edge Signature

Highlights

• Local Edge Signature is a geometric-insensitive operator that represents texture.

• Local Edge Signature is based on edge pixels’ arrangements and orientations.

• A genetic programming technique learns automatically a global texture descriptor.

• A tree representation of individuals generates global texture features.

• Generated descriptor needs few training images to extract discriminative features.

Abstract

Describing texture is a very challenging problem for many image-based expert and intelligent systems (e.g. defective product detection, people re-identification, abnormality investigation in medical imaging and remote sensing applications$\dots$) since the process of texture classification relies on the quality of the extracted features. Indeed, detecting and extracting features is a hard and time-consuming task that requires the intervention of an expert, notably when dealing with challenging textures. Thus, machine learning-based descriptors have emerged as another alternative to deal with the difficulty of feature extracting. In this work, we propose a new operator, which we named Local Edge Signature (LES) descriptor, to locally represent texture. The proposed texture descriptor is based on statistical information on edge pixels’ arrangement and orientation in a specific local region, and it is insensitive to rotation and scale changes. A genetic programming-based approach is then fitted to automatically learn a global texture descriptor that we called Genetic Texture Signature (GTS). In fact, a tree representation of individuals is used to generate global texture features by applying elementary operations on LES elements at a set of keypoints, and a fitness function evaluates the descriptors considering intra-class homogeneity and inter-class discrimination properties of their generated features. The obtained results, on six challenging texture datasets (Brodatz, Outex_TC_00000, Outex_TC_00013, KTH-TIPS, KTH-TIPS2b and UIUCTex), show that the proposed classification method, which is fully automated, achieves state-of-the-art performance, especially when the number of available training samples is limited.

Introduction

Texture description and classification play a fundamental role in many computer vision applications, such as surface inspection of materials, medical imaging, image recovery, object and scene recognition. Due to the importance and the abundance of application fields, several texture classification approaches have been proposed during the last years. However, extracting highly representative and robust texture characteristics to describe textured images is a difficult problem that needs further investigation. In fact, many classical descriptors have proven to be efficient when describing texture. For instance, edge-based descriptors have focused on calculating first-order or second-order statistical measures of edge distributions. Indeed, gradient magnitudes and directions are calculated in a local neighboring in order to extract the edge distribution. Other descriptors focused on spatial frequencies of texture primitives. For example, the Local Binary Pattern (LBP) descriptor and its variants tried to model local texture by calculating the difference between a central pixel intensity and its neighbors’ intensities within a local support region. All these descriptors tried to describe local texture by means of features based on the occurrence of corners, the direction of edges, the shape of textures, the difference of pixel intensities and/or the spatial frequencies of texture primitives. Although some of these descriptors have managed to achieve good results, they still suffer from weakness against textures with scale, rotation or/and deformation variations. The most likely causes of this weakness are that these descriptors lacked to model the texture as seen by human visual sense and also focused very locally while forgetting that texture is defined in a global way. Thus, many textural features were defined to mimic the way human perceive texture, such as coarseness, contrast, complexity, busyness, shape, directionality and texture strength. Indeed, some works (Payne et al., 1999, Benjamin, 2006) compared the computational with human ranking for the texture classification, and they stated that there was a good correspondence between the two. A major disadvantage of almost human-like approaches is that they do not have general applicability. The human perception mechanism, in comparison, seems to work well for almost all types of textures. There is some agreement between most researchers about the main categories of texture classification, but they also note that humans tend to combine rather than use one single method. In (Payne et al., 1999), authors have carried out sets of human studies, where volunteers were asked to select which images, in order, from a set of 111 images, they considered to be most like a given target image. Fig. 1.a shows a target image and Fig. 1.b refers to the most like image according to human subjects. Most LBP-based descriptors would not be able to classify these two images as being of the same class. Indeed, these descriptors are totally decorrelated from the way humans perceived the resemblance between the two textures. In fact, the correlation between the two textures in Fig. 1 lies in the curvatures of the contours and the frequency of the discontinuities along the different directions. Thus, texture classification approaches still need to deal with the following two main challenges: how to describe locally a complex texture with relatively low dimensional measures while remaining insensitive to changes that may occur? and, how to aggregate these local texture measures to obtain a global texture description? The contributions of this work come to deal with these two issues. Firstly, a new local texture measure, named Local Edge Signature (LES), is proposed to describe local texture with a 6-dimensional vector. This local texture representation uses statistics on structural information in a specific local distribution of pixels around a central pixel. The local distribution is designed to have rotation and scale insensitive statistical information describing the local texture. Thus, the proposed LES descriptor tries to combine different aspects of the human way to perceive texture in a low-computational feature that has general applicability. Secondly, a genetic programming approach, which we called GTS for Genetic Texture Classification, automatically evolves robust texture descriptor from a small number of training instances. In fact, to obtain a global feature, tree representation using arithmetic and comparison operators is proposed in order to aggregate local edge signatures on a set of keypoints. Furthermore, a fitness function considering the intra-class homogeneity and inter-class discrimination properties of the features is designed. The significance of the proposed method lies in the generated descriptor that extracts discriminative and geometric-insensitive texture features from a small training set, without the need for an expert intervention. This makes it particularly appropriate for expert and intelligent systems dealing with numerous real-life problems of society and industry that include unconstrained content-based image classification (Swalpa, Dipak, Shiv, Siddhartha, & Bidyut, 2020) and retrieval (Majumdar, Chatterji, & Avijit, 2020), biomedical image analysis (Jouirou, Baâzaoui, & Barhoumi, 2019) and multi-spectral remotely sensed imaging (Wang, Yu, & Fang, 2020).

The rest of this paper is organized as follows. We present a brief literature review on existing methods for texture classification in Section 2. The proposed method is detailed in Section 3. The results are presented and discussed in Section 4. Finally, in Section 5, we conclude the proposed work and present some ideas for future studies.