RefConv: Reparameterized Refocusing Convolution for Powerful ConvNets.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are fundamental to deep learning.
• Existing CNN architectures often have limitations in representational capacity.
• Improving CNN performance without inference cost is a significant research goal.

Purpose of the Study:

• To introduce RefConv as a novel, plug-and-play module to enhance CNN performance.
• To demonstrate RefConv's ability to improve models without additional inference costs.
• To analyze the mechanisms by which RefConv enhances model capabilities.

Main Methods:

• Proposed reparameterized refocusing convolution (RefConv) as a replacement for standard convolutional layers.
• Developed a trainable Refocusing Transformation to modify inherited basis kernels.
• Applied RefConv to pre-trained CNNs for various computer vision tasks.

Main Results:

• RefConv improved performance across image classification, object detection, and semantic segmentation tasks.
• Achieved up to 1.47% higher top-1 accuracy on ImageNet.
• Demonstrated effectiveness against adversarial attacks without altering model structure or inference cost.
• Showed RefConv strengthens kernel spatial skeletons, reduces channel redundancy, and smooths the loss landscape.

Conclusions:

• RefConv is an effective module for enhancing CNN representational capacity.
• The plug-and-play nature of RefConv allows easy integration into existing models.
• RefConv offers a cost-effective method for boosting deep learning model performance.