Deep convolutional neural networks for regular texture recognition.

Area of Science:

• Computer Vision and Image Analysis
• Machine Learning and Pattern Recognition

Background:

• Regular textures are prevalent in man-made environments and natural imagery.
• Accurate recognition and localization of textures have numerous practical applications.

Purpose of the Study:

• To model and classify regular and irregular textures using deep convolutional neural networks (CNNs).
• To evaluate the performance of standard CNN architectures and Fisher Vector (FV) representations for texture classification.

Main Methods:

• Developed a novel database for regular texture classification.
• Trained classic CNN models (e.g., Inception, ResNet) via fine-tuning on the new database.
• Extracted features using Fisher Vector pooling from the last convolutional layer and used Support Vector Machines (SVM) for classification.

Main Results:

• Standard CNNs achieved sufficient accuracy in recognizing regular textures.
• Fisher Vector representations combined with SVM demonstrated high performance for both regular and irregular texture classification.
• CNNs exhibited sub-optimal performance on long-range patterns due to local feature pooling in fully-connected layers.

Conclusions:

• Deep CNNs are a viable method for regular and irregular texture classification.
• Fisher Vector representations offer a transferable and effective approach for texture analysis.
• Limitations of CNNs in capturing long-range dependencies were identified.