Lightweight deep neural networks: Optimization of vehicle classification using ICBAM based on depthwise separable...

Area of Science:

• Computer Science
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Vehicle classification is crucial for intelligent transportation systems.
• Existing deep learning models face challenges with computational efficiency and generalization.
• There is a need for lightweight and accurate models for practical deployment.

Purpose of the Study:

• To propose DSICBAMNet, a novel lightweight and efficient deep neural network for vehicle classification.
• To enhance computational efficiency and generalization ability compared to existing models.
• To validate the model's performance on benchmark datasets.

Main Methods:

• Developed DSICBAMNet, integrating Depthwise Separable Convolutions (DSC) and an Improved Convolutional Block Attention Module (ICBAM).
• DSC reduces computational complexity and parameters.
• ICBAM improves overfitting resistance and feature weighting via Dropout and optimized attention mechanisms.

Main Results:

• DSICBAMNet achieved 97.36% accuracy on the MIO-TCD dataset (286 samples) and 96.51% on the Stanford Cars dataset (1,060 samples).
• Demonstrated superior performance compared to five classic models (e.g., AlexNet, MobileNetV2).
• Grad-CAM and confusion matrix analysis confirmed effective focus on key regions and consistent classification.

Conclusions:

• DSICBAMNet offers a promising solution for efficient and accurate vehicle classification in intelligent transportation.
• The model's lightweight design and high performance make it suitable for resource-constrained environments.
• Validated practical applicability and value in intelligent transportation scenarios.