C-CNN: Contourlet Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Texture classification faces challenges due to scale variations and pattern clutter.
• Traditional methods rely on frequency-domain spectral analysis.
• Convolutional Neural Networks (CNNs) show promise in spatial-domain texture analysis.

Purpose of the Study:

• To develop a novel network architecture for robust texture classification.
• To combine spectral and spatial domain features for richer information.
• To improve feature representation learning for image analysis.

Main Methods:

• Application of contourlet transform for spectral feature extraction.
• Development of a spatial-spectral feature fusion strategy.
• Integration of statistical features via statistical feature fusion.
• Classification using fused features.

Main Results:

• The proposed contourlet CNN (C-CNN) architecture effectively learns sparse and discriminative features.
• Experiments on multiple texture and remote sensing datasets show superior performance.
• The C-CNN achieved higher classification accuracy compared to existing methods.
• The approach requires fewer trainable parameters.

Conclusions:

• Combining spectral and spatial features via the C-CNN enhances texture classification performance.
• The C-CNN offers an effective solution for challenging texture classification problems.
• The method demonstrates significant improvements on benchmark datasets.