DO-Conv: Depthwise Over-Parameterized Convolutional Layer.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are fundamental to modern computer vision.
• Conventional convolutional layers are the primary computational units within CNNs.
• Over-parameterization is a technique explored to improve model performance.

Purpose of the Study:

• To propose and evaluate a novel convolutional layer, Depthwise Over-parameterized Convolution (DO-Conv).
• To demonstrate that DO-Conv can enhance CNN performance across various computer vision tasks.
• To show that DO-Conv does not introduce additional computational cost during inference.

Main Methods:

• Augmenting standard convolutional layers with an additional depthwise convolution.
• Introducing learnable parameters through this depthwise convolution, creating an over-parameterized layer.
• Folding the depthwise convolution into the conventional convolution during inference to maintain computational efficiency.

Main Results:

• DO-Conv layers significantly boost performance in image classification, detection, and segmentation tasks.
• The proposed DO-Conv layers are shown to be an effective form of over-parameterization.
• Experimental results confirm performance gains without increased inference complexity.

Conclusions:

• DO-Conv offers a method to improve CNN performance by augmenting convolutional layers.
• The technique provides a practical enhancement as it is computationally equivalent to standard convolutions during inference.
• DO-Conv is presented as a viable and beneficial alternative to conventional convolutional layers for computer vision applications.