An integrated approach for advanced vehicle classification.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Conventional Convolutional Neural Networks (CNNs) face a trade-off between receptive field size and computational cost.
• Existing methods like adding layers or expansion filtering increase computational burden or result in sparse sampling.
• The grid effect in expansion filtering limits receptive field coverage.

Purpose of the Study:

• To propose a novel multilevel Discrete Wavelet CNN (DWAN) model.
• To address the trade-off between receptive field size and computational efficiency in low-level vision tasks.
• To improve the performance and efficiency of visual tasks through enhanced feature capture.

Main Methods:

• Introduced a four-level discrete wavelet transform into the CNN architecture.
• Integrated the Convolutional Block Attention Module (CBAM) for efficient multiscale feature capture.
• Implemented a shrinkage subnetwork to reduce feature map size, widening sensory field coverage with lower computational cost.

Main Results:

• The DWAN model achieved wider sensory field coverage with reduced computational cost.
• Demonstrated significant performance gains in a fine-grained automotive image classification task.
• The model accurately identified subtle features and differences in automotive images, improving classification accuracy.

Conclusions:

• The proposed DWAN model effectively balances receptive field size and computational efficiency.
• DWAN shows effectiveness and robustness in vision tasks, particularly for fine-grained classification.
• The model provides a strong foundation for generalization to practical computer vision applications.