Genetic programming-based learning of texture classification descriptors from Local Edge Signature
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.
Section snippets
Related work
Texture classification, which is an active research area in computer vision, strongly relies on texture description. Wang and He (He & Wang, 1990) defined texture as a variation in the intensity of an image, and Haralick (Haralick, 1979) described texture as a scale-dependent phenomenon that resulted from spatial interaction between primitives, or groups of similar pixels. Indeed, texture is commonly described using statistics in a local neighborhood. First-order statistics, especially
Proposed method
This section provides a detailed description of the proposed method for texture classification. Indeed, after presenting the process of extracting the local texture signature descriptor, we give an overview of the proposed genetic texture classification algorithm including coding of the genetic individuals, feature vectors’ extraction and fitness calculation.
Experimental results
In this section, the proposed texture classification method is tested and compared to nine relevant texture classification methods. All these methods, including the proposed one, have been implemented using the Anaconda Python distribution and the DEAP1 evolutionary computation framework. The implementation of the
Conclusion and future work
In this work, a robust method of automatic generation of texture descriptors through genetic programming is proposed. Texture is described locally using a new descriptor called Local Edge Signature (LES). We have reported in this descriptor, in a manner faithful to the human perception of the texture, statistics related to the arrangement of edges in a local region. The nature of the local neighborhood, as well as the statistics used in the proposed method, make the description of the texture
CRediT authorship contribution statement
Haythem Ghazouani: Conceptualization, Methodology, Software, Validation, Writing - original draft, Writing - review & editing. Walid Barhoumi: Conceptualization, Methodology, Software, Validation, Writing - original draft, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (48)
- et al.
Multi-view information fusion in mammograms: A comprehensive overview
Information Fusion
(2019) - et al.
Extended local binary patterns for texture classification
Image and Vision Computing
(2012) - et al.
Automatically evolving rotation-invariant texture image descriptors by genetic programming
IEEE Transactions on Evolutionary Computation
(2016) - et al.
Image descriptor: A genetic programming approach to multiclass texture classification. In 2015 IEEE Congress on
Evolutionary Computation
(2015) - et al.
Noise robust and rotation invariant framework for texture analysis and classification
Applied Mathematics and Computation
(2018) Upper and lower volumetric fractal descriptors for texture classification
Pattern Recognition Letters
(2017)- Benjamin, B., J. (2006). Texture synthesis and perception: Using computational models to study texture representations...
- Brodatz, P. (1966). Textures: A photographic album for artists and designers....
- et al.
A theoretical comparison of texture algorithms
IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-2
(1980) - et al.
Using basic image features for texture classification
International Journal of Computer Vision
(2010)