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

## TL;DR

This paper introduces DSICBAMNet, a lightweight deep learning model for vehicle classification that balances accuracy and efficiency.

## Contribution

The novel DSICBAMNet combines depthwise separable convolutions and an improved attention module to enhance vehicle classification in resource-limited settings.

## Key findings

- DSICBAMNet achieves 97.36% accuracy on the MIO-TCD dataset and 96.51% on the Stanford Cars dataset.
- The model demonstrates strong generalization and efficient multi-scale feature extraction.
- Grad-CAM and confusion matrix analysis confirm the model's focus on key regions and consistent classification.

## Abstract

Vehicle classification is a core task in intelligent transportation systems, where high demands are placed on both computational efficiency and generalization ability in practical applications. Existing deep learning models often struggle to meet these requirements due to their high computational complexity and limited generalization. To address this challenge, this study proposes a lightweight and efficient deep neural network called DSICBAMNet, which achieves high classification accuracy while significantly improving computational efficiency. The design of DSICBAMNet is centered on two key components: Depthwise Separable Convolutions (DSC) and an Improved Convolutional Block Attention Module (ICBAM). The DSC module reduces the number of parameters and computational complexity by decomposing convolution operations, making it well-suited for resource-constrained deployment scenarios. Meanwhile, ICBAM addresses the shortcomings of traditional CBAM in terms of overfitting resistance and feature weighting strategies. By introducing Dropout regularization into the channel attention module, ICBAM enhances the model’s resistance to overfitting. Additionally, it optimizes the interaction mechanisms and weight distribution between the channel and spatial attention modules, enabling more accurate multi-class feature representation. The network achieves efficient multi-scale feature extraction by stacking multiple improved DSICBAM blocks while maintaining an overall lightweight structure. In experimental evaluations, DSICBAMNet was compared with five classic models, including AlexNet and MobileNetV2. Experimental results demonstrate that DSICBAMNet achieves outstanding performance on both the MIO-TCD dataset, with 286 test samples and an average classification accuracy of 97.36%, and the Stanford Cars dataset, with 1,060 test samples and an accuracy of 96.51%. Moreover, the combination of Grad-CAM visualizations and confusion matrix analysis validates the model’s ability to focus on key regions and maintain consistency in classification outcomes. These results underscore the model’s potential applicability and practical value in intelligent transportation scenarios.