Resolution Learning in Deep Convolutional Networks Using Scale-Space Theory.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) typically have fixed resolution limitations due to architecture design.
• Tuning resolution hyper-parameters in modern CNNs is complex and dataset-dependent.
• Existing methods struggle to adapt network resolution dynamically.

Purpose of the Study:

• To develop a novel layer for CNNs that learns optimal resolution directly from data.
• To eliminate the need for hard-coded resolution hyper-parameters in network architectures.
• To improve the adaptability and performance of CNNs across diverse datasets and input sizes.

Main Methods:

• Utilized scale-space theory for self-similar filter parametrization.
• Employed the N-Jet, a truncated Taylor series, to approximate filters using learned Gaussian derivative filters.
• Optimized the Gaussian basis parameter σ, controlling filter detail and spatial extent, with respect to the loss function.

Main Results:

• The proposed N-Jet layer demonstrated comparable performance to state-of-the-art architectures.
• The N-Jet layer automatically learned the correct resolution at each layer.
• Learning σ proved particularly beneficial for tasks involving inputs of multiple sizes, such as classification and segmentation.

Conclusions:

• The N-Jet layer offers a flexible and data-driven approach to resolution in CNNs.
• Automatic resolution learning via the N-Jet layer simplifies network design and improves performance.
• This method shows significant promise for handling variable input resolutions in deep learning applications.