Scalable Online Convolutional Sparse Coding.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Convolutional Sparse Coding (CSC) enhances sparse coding by learning shift-invariant dictionaries.
• Existing CSC algorithms are often computationally expensive due to batch processing.

Purpose of the Study:

• To develop an efficient online learning algorithm for CSC.
• To address the computational cost and scalability limitations of current CSC methods.

Main Methods:

• Reformulated the CSC objective for efficient frequency-domain convolution.
• Employed the alternating direction method of multipliers (ADMMs) for optimization.
• Developed an online learning approach requiring smaller history matrices.

Main Results:

• The proposed online CSC algorithm demonstrates faster convergence.
• Achieved superior reconstruction performance on benchmark and large-scale datasets (e.g., ImageNet).
• Outperformed state-of-the-art batch and online CSC methods in scalability and efficiency.

Conclusions:

• The novel online CSC algorithm is computationally efficient and scalable.
• Theoretical analysis confirms dictionary convergence to a stationary point.
• The method offers a practical advancement for CSC applications.