CNN-Based Classification for Highly Similar Vehicle Model Using Multi-Task Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Vehicle make and model classification is vital for intelligent transportation systems (ITS).
• Accurate identification is challenging for vehicles with similar visual appearances.
• Existing methods struggle with fine-grained classification of visually alike vehicles.

Purpose of the Study:

• To develop a fine-grained vehicle classifier capable of distinguishing between makes and models of cars with high visual similarity.
• To leverage a multi-task learning approach within a convolutional neural network framework.

Main Methods:

• Utilized the VGG-16 architecture for feature extraction from vehicle images.
• Implemented a multi-task learning approach with two separate branches for make and model classification.
• Evaluated the model on the InaV-Dash dataset, specifically targeting Indonesian vehicle models.

Main Results:

• Achieved high accuracy rates: 98.73% for vehicle make and 97.69% for vehicle model.
• Demonstrated superior performance compared to baseline methods on visually similar vehicle classification.
• Successfully addressed the challenge of distinguishing between makes and models of cars that look alike.

Conclusions:

• The proposed multi-task learning convolutional neural network classifier effectively handles fine-grained vehicle classification.
• This method offers a significant improvement for intelligent transportation systems dealing with visually similar vehicles.
• The approach provides a robust solution for accurate vehicle identification in complex scenarios.