Learning Versatile Convolution Filters for Efficient Visual Recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Efficient deep learning models are crucial for cost-effective hardware.
• Existing methods focus on reducing filter size, sparsity, or quantization.
• A novel approach is needed to enhance filter capabilities without increased resource demands.

Purpose of the Study:

• To introduce versatile filters for constructing efficient convolutional neural networks.
• To enhance the capability of filters by integrating information from diverse receptive fields.
• To reduce memory and computation costs in visual recognition tasks.

Main Methods:

• Filters are treated from an additive perspective using binary masks to derive secondary filters.
• Versatile filters are investigated from both spatial and channel perspectives.
• Theoretical analysis of network complexity and an efficient convolution scheme are introduced.

Main Results:

• Versatile filters inherit primary filter storage but enhance computational capability.
• Secondary filters integrate information from different receptive fields.
• Comparable accuracy is achieved with reduced memory and computation costs.

Conclusions:

• The proposed versatile filters offer an efficient approach to enhance convolutional neural networks.
• This method achieves comparable accuracy to original filters with significant resource savings.
• The versatile filters are effective for various visual recognition tasks on benchmark datasets.