A novel sequential structure for lightweight multi-scale feature learning under limited available images.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep models with multi-scale feature learning improve performance but increase parameters quadratically, leading to overfitting with limited data.
• Lightweight models reduce overfitting but can underfit due to insufficient feature learning with limited data.

Purpose of the Study:

• To propose a novel lightweight model, the Sequential Multi-scale Feature Learning Network (SMF-Net), that addresses both overfitting and underfitting issues in deep learning.
• To enable efficient multi-scale feature learning with larger receptive fields using a sequential structure and minimal, linearly increasing parameters.

Main Methods:

• Introduced a novel sequential structure for multi-scale feature learning within the SMF-Net architecture.
• Developed a lightweight network designed to extract features with larger receptive fields efficiently.
• Evaluated SMF-Net on both image classification and segmentation tasks.

Main Results:

• SMF-Net achieved high accuracy on classification and segmentation tasks, outperforming state-of-the-art deep and lightweight models.
• The proposed model requires significantly fewer parameters and computations: 1.25M parameters and 0.7G FLOPS for classification, and 1.54M parameters and 3.35G FLOPs for segmentation.
• Demonstrated superior performance even with very limited available training data.

Conclusions:

• SMF-Net effectively alleviates overfitting and underfitting issues common in deep and lightweight models, respectively.
• The sequential multi-scale feature learning approach allows for efficient extraction of features with large receptive fields.
• SMF-Net presents a promising solution for resource-constrained environments and scenarios with limited training data.