A Convolutional Neural Networks-Based Approach for Texture Directionality Detection

Area of Science:

• Computer Vision
• Image Analysis
• Machine Learning

Background:

• Texture directionality is a crucial image characteristic with significant applications.
• Existing methods like Fourier-based and interpolated grey-level co-occurrence matrix (iGLCM) have limitations in speed or robustness.
• Convolutional neural networks (CNNs) offer a potential new approach for this task.

Purpose of the Study:

• To evaluate the effectiveness of CNNs for texture directionality detection.
• To compare CNN performance against established methods (Fourier, iGLCM, local gradient orientation).
• To assess the robustness and speed of CNN-based texture directionality detection.

Main Methods:

• Development of a synthetic texture dataset with controlled directionality and perturbations for training.
• Definition and testing of various shallow and deep CNN architectures.
• Comparative analysis of CNNs against iGLCM, Fourier, and local gradient orientation methods.

Main Results:

• CNNs achieve accuracy comparable to iGLCM, outperforming Fourier and local gradient orientation methods.
• CNNs demonstrate significantly higher computational speed compared to other methods.
• The best-performing CNN architecture shows generalization capabilities on real-life image data.

Conclusions:

• CNNs represent a viable and efficient approach for texture directionality detection.
• The speed and generalization of CNNs warrant further research and development in this area.
• CNNs offer a promising alternative for applications requiring rapid and accurate texture analysis.