DualConv: Dual Convolutional Kernels for Lightweight Deep Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Network (CNN) architectures are computationally intensive and memory-heavy, limiting their use in embedded systems.
• Efficient deep learning models are crucial for resource-constrained environments.

Purpose of the Study:

• To propose a novel lightweight CNN architecture using dual convolutional kernels (DualConv).
• To demonstrate the effectiveness of DualConv in reducing computational cost and parameters while maintaining accuracy.

Main Methods:

• Introduced DualConv, which combines 3x3 and 1x1 convolutional kernels processing the same feature map channels simultaneously.
• Utilized group convolution techniques for efficient filter arrangement.
• Integrated DualConv into various CNN models like VGG-16, ResNet-50, YOLO-V3, and MobileNetV2 for testing.

Main Results:

• DualConv significantly reduces computational cost and parameters in deep neural networks.
• Achieved slightly higher accuracy than original models in some classification tasks.
• Reduced MobileNetV2 parameters by 54% with a minimal accuracy drop (0.68%) on CIFAR-100.
• Improved MobileNetV1 accuracy by 4.11% on CIFAR-100.
• Enhanced YOLO-V3 object detection speed and accuracy by 4.4% on PASCAL VOC dataset.

Conclusions:

• DualConv is an effective method for creating lightweight and efficient CNNs suitable for embedded systems.
• The proposed method offers a flexible way to improve existing CNN architectures for various computer vision tasks.