Steerable Wavelet Machines (SWM): Learning Moving Frames for Texture Classification.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Local organizations of image directions (LOIDs) are crucial for visual understanding.
• Existing methods like Local Binary Patterns (LBPs) and Scale-Invariant Feature Transform (SIFT) use hand-crafted representations.
• There is a need for learned, rotation-invariant texture representations.

Purpose of the Study:

• To propose novel texture operators that learn data-specific, rotation-invariant representations of LOIDs.
• To introduce steerable wavelet machines (SWMs) for enhanced texture characterization.
• To demonstrate the effectiveness of SWMs in natural texture classification.

Main Methods:

• Utilized steerable circular harmonic wavelets (CHWs) for initial texture representation.
• Employed moving frames (MFs) to preserve joint location and orientation information.
• Trained support vector machines to learn a multi-class shaping matrix for CHW representations, creating SWMs.
• SWMs involve linear operations (convolution, weighted combinations) and non-linear steermax operations.

Main Results:

• The proposed SWMs effectively encode class-specific LOIDs in a rotation-invariant manner.
• Experimental results show superior performance compared to recent approaches.
• The SWMs achieved high accuracy on texture classification tasks using Outex and CUReT databases.

Conclusions:

• Steerable wavelet machines offer a powerful approach for learning rotation-invariant texture representations.
• The proposed method significantly advances the state-of-the-art in natural texture classification.
• Learned, data-driven texture operators provide advantages over hand-crafted methods.