TbsNet: the importance of thin-branch structures in CNNs.

Area of Science:

• Computer Vision
• Deep Learning
• Neural Network Architecture Design

Background:

• Convolutional Neural Network (CNN) performance is complex, depending on factors like depth, width, and structure.
• Optimizing these factors for efficient CNN architecture design is a significant challenge.

Purpose of the Study:

• To analyze key factors influencing CNN performance.
• To propose strategies for efficient convolutional network construction.
• To introduce TbsNet, a novel, simplified network architecture.

Main Methods:

• Implemented lightweight operators (asymmetric, pointwise, depthwise, group convolution) to reduce computation and feature redundancy.
• Designed TbsNet with a plain, simplified structure, avoiding reparameterization and extraneous branches.
• Conducted extensive experiments evaluating network depth and width.

Main Results:

• TbsNet achieved high Top 1 Accuracy: 97.02% on CIFAR-10, 83.56% on CIFAR-100, and 86.17% on ImageNet-1K.
• Tbs-UNet demonstrated a Dice Similarity Coefficient (DSC) of 78.39% on the Synapse dataset, outperforming TransUNet.
• TbsNet shows competitive performance against state-of-the-art networks like ResNet, ConvNeXt, and MobileNet.

Conclusions:

• TbsNet offers an efficient and effective architecture for computer vision tasks.
• The simplified, thin-branch structure minimizes computational cost and feature redundancy.
• TbsNet is suitable for downstream applications including medical image segmentation.