CED-Net: A more effective DenseNet model with channel enhancement

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) are increasingly used in various fields.
• Low-power platforms require efficient and smaller network sizes.
• Existing lightweight CNNs face challenges in balancing accuracy and computational cost.

Purpose of the Study:

• To propose CED-Net (Channel enhancement DenseNet), a more efficient densely connected network.
• To improve network accuracy and feature representation for resource-constrained environments.
• To reduce computational overhead (FLOPs) and parameter count in CNNs.

Main Methods:

• Developed CED-Net by integrating bottleneck layers with learned group convolution and a channel enhancement module.
• Designed CED-Net based on the CondenseNet architecture.
• Conducted experiments on CIFAR-10 and CIFAR-100 datasets.

Main Results:

• CED-Net achieved higher accuracy on CIFAR-10 (0.4%) and CIFAR-100 (1%) compared to CondenseNet.
• The proposed model maintains similar parameter counts and FLOPs to CondenseNet.
• Ablation experiments validated the effectiveness of the bottleneck layer in CED-Net.

Conclusions:

• CED-Net offers improved performance over existing lightweight CNNs.
• The integration of learned group convolution and channel enhancement effectively boosts network representation and accuracy.
• CED-Net presents a viable solution for efficient deep learning on low-power platforms.