Multiscale Curvelet Scattering Network.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Convolutional Neural Networks (CNNs) are standard for image classification but can be improved for feature representation.
• Existing methods focus on enhancing CNN feature extraction, particularly for edges and textures.
• Advanced feature representation is crucial for boosting image classification performance.

Purpose of the Study:

• To introduce a novel classification framework, the multiscale curvelet scattering network (MSCCN).
• To enhance image feature representation by integrating curvelet features into a scattering process.
• To improve classification accuracy over current state-of-the-art techniques.

Main Methods:

• Developed the multiscale curvelet scattering network (MSCCN) incorporating a multiscale curvelet-scattering module (CCM).
• Utilized multiscale geometric analysis and curvelet features to enhance scattering processes with directional information.
• Formulated a unified optimization structure for scattering processes and curvelet features, aggregating and learning multi-scale features.
• Proposed a single-level CCM for embedding into existing networks to improve feature representation quality.

Main Results:

• MSCCN demonstrated superior classification accuracy compared to existing state-of-the-art methods.
• The multiscale curvelet-scattering module (CCM) effectively improved feature representation quality.
• Experimental results validated the network's convergence, insight, and adaptability through loss function analysis, feature map visualization, and generalization analysis.

Conclusions:

• The proposed MSCCN framework offers a significant advancement in image classification.
• Integrating curvelet features within a scattering framework provides richer, more effective image representations.
• MSCCN shows strong potential for improving various image analysis tasks and can be adapted for use with other networks.