Images of different object surfaces convey important information about haptically perceptible textures. The extraction of tactile information of different materials by making use of inexpensive technologies can have practical and commercial applications in e-commerce or robotics. However, differences in distance, rotation, lighting and focus conditions are hurdles which need to be overcome to extract robust image-based features that will allow a successful surface classification task.
In this work, eleven haptically relevant features are introduced, which have a low to invariant dependency on different camera conditions. These are used for a robust machine learning-based approach for surface classification.
A database of 690 images, corresponding to 69 different textures, is used to extract haptically relevant features. Perceptually-relevant image features such as roughness, softness and regularity are used to correctly classify the textures. The extracted features are perceptually relevant so that they can be also used in future work for the retrieval of the most similar textured surface to a classified one.
Experimental results and the evaluation of a cross-validated naive Bayes classifier show that the proposed approach allows for the successful classification of textured surfaces under varying camera conditions, a maximum prediction accuracy of 85.8% being achieved. When a subset of 6 features is selected, a classification accuracy of 82.5% is obtained.
Table of Contents
1 Introduction
2 Related Work
2.1 Textural Properties
2.2 Structure of Texture Patterns
2.3 Features Defined by Tamura
2.3.1 Coarseness
2.3.2 Contrast
2.3.3 Directionality
2.3.4 Linelikeness
2.4 Other Features
2.4.1 Complexity
2.4.2 Regularity
2.5 Shortcomings
3 Recording Procedure
3.1 Original Database
3.2 Magnified Database
4 Haptically Related Textural Features
4.1 Improved Features
4.1.1 Coarseness
4.1.2 Directionality
4.1.3 Linelikeness
4.1.4 Complexity
4.1.5 Regularity
4.2 New Features Definitions
4.2.1 Edginess
4.2.2 Color Distance
4.2.3 Roughness
4.2.4 Glossiness
4.2.5 Softness
5 Subjective Experiment
6 Results
6.1 Statistical Evaluation of Features
6.1.1 Feature Scaling and Cross Validation
6.1.2 Statistical Analysis
6.1.3 Feature Quality Identification
6.2 Feature Selection for High-Dimensional Data Classification
6.3 Classification Accuracy
6.3.1 Original Database
6.3.2 Magnified Database
6.4 Discussion
7 Conclusion
Research Objective and Scope
This work explores the automated classification of various textured surfaces by extracting haptically relevant image features. The primary research question addresses whether robust image-based feature extraction can successfully classify materials under varying camera conditions, such as distance, rotation, and lighting, to support applications in e-commerce and robotics.
- Development of eleven haptically relevant texture features.
- Creation of two comprehensive image databases (original and magnified).
- Implementation of robust machine learning-based surface classification.
- Visio-haptic subjective experiment to establish human perception ground truth.
- Evaluation of classification accuracy and feature subset selection strategies.
Excerpt from the Book
2.3.1 Coarseness
Coarseness is, along with contrast and directionality, one of the most fundamental textural features as mentioned in [TMY78] and is related to the appearance of large structured elements in an image. For example, in the case of two patterns differing only in scale, the magnified one is coarser, whereas when it comes to patterns that have different structures, the bigger its element size is or the less often its elements are repeated, the coarser it is perceived by humans, as shown in Figure 2.1.
The algorithm starts by computing and storing the average gray level (AGL) of subwindows of size 2k × 2k , with k ∈ {1, 2, 3, 4, 5}, which are centered at every pixel within the image, where the AGL is computed as: AGLk = (sum of pixels) / 2^(2k).
Afterwards the symmetric subtraction of average gray levels is undergone. This is performed for non-overlapping, adjacent subwindows in the horizontal and vertical directions with respect to the current pixel for the complete range of sizes, as shown in Figure 2.2b. Afterwards, the results are stored. For example, in the vertical direction, this is computed as Ek,v(x, y) = |AGLk(x, y + 2k−1) − AGLk(x, y − 2k−1)|. For each pixel, the k value which corresponds to the maximum gradient of average gray levels of the adjacent subwindows in all directions is selected and Sbest(x, y)=2k is obtained. The coarseness feature (Fcrs) is determined as the average of Sbest over the entire picture.
Summary of Chapters
1 Introduction: Provides an overview of the importance of haptic texture perception in image processing and establishes the goal of using visual data to classify material textures.
2 Related Work: Reviews existing texture analysis methods and defines established textural features such as coarseness, contrast, directionality, and regularity.
3 Recording Procedure: Details the acquisition of the two image databases, including the use of smartphones and magnifying attachments to capture texture details under various conditions.
4 Haptically Related Textural Features: Presents improvements to classical features and introduces new descriptors like edginess, glossiness, and softness to better match human tactile perception.
5 Subjective Experiment: Describes the design and execution of a human user study to gather ground truth data on texture roughness via visual and tactile inspection.
6 Results: Evaluates the classification performance using various machine learning classifiers and analyzes the effectiveness of feature selection methods.
7 Conclusion: Summarizes the findings and discusses the potential for future improvements in automated haptic texture classification.
Keywords
Texture classification, Haptics, Image processing, Feature extraction, Coarseness, Contrast, Directionality, Linelikeness, Complexity, Regularity, Edginess, Glossiness, Softness, Naive Bayes, Machine learning
Frequently Asked Questions
What is the core purpose of this research?
The work aims to extract image-based features that correlate with human tactile perception, enabling automated classification of textured surfaces for applications like robotics or e-commerce.
Which specific textural characteristics are analyzed?
The study focuses on eleven features, including traditional properties like coarseness and directionality, and newly defined ones such as edginess, glossiness, and softness.
What is the primary goal of the feature extraction process?
The goal is to generate robust, rotation-invariant, and illumination-invariant features that allow for reliable machine learning-based classification of textures.
Which machine learning methods are employed?
The research primarily utilizes Naive Bayes classifiers, Quadratic Discriminant Analysis (QDA), and Linear Discriminant Analysis (LDA) to evaluate classification accuracy.
How is the classification performance measured?
Performance is measured using misclassification error (MCE), classification accuracy (CA), and a goodness of features criterion (GFC) to assess feature quality.
What are the fundamental challenges addressed by this thesis?
The thesis addresses issues such as the "curse of dimensionality," redundancy in textural data, and the sensitivity of feature calculations to varying capture conditions like light and camera distance.
Why are magnified images included in the study?
Magnified images are used to capture fine-grained intrinsic details—such as fiber orientation or small gaps in foams—that are invisible in standard wide-angle smartphone photography.
How does the subjective experiment influence the results?
The experiment provides human-labeled roughness data, which acts as a reference for optimizing the computational roughness feature to ensure it aligns with human tactile impressions.
- Citation du texte
- Albert Iepure (Auteur), 2015, Image Based Haptic Feature Extraction, Munich, GRIN Verlag, https://www.grin.com/document/493690
