A Small Network MicronNet-BF of Traffic Sign Classification

Area of Science:

• Computer Vision
• Deep Learning
• Embedded Systems

Background:

• Traffic sign recognition is crucial for real-world applications.
• Deep neural networks offer high accuracy but pose challenges for embedded systems due to size and complexity.
• There is a need for small, accurate neural networks suitable for resource-constrained environments.

Purpose of the Study:

• To improve the MicronNet convolutional neural network for traffic sign recognition.
• To develop a smaller, more accurate neural network for embedded systems.
• To enhance the efficiency and generality of neural networks for real-world applications.

Main Methods:

• The study enhanced the MicronNet model by incorporating batch normalization (BN) and factorization techniques.
• Batch normalization was applied to reduce the effect of image brightness on feature recognition.
• Factorization was used to replace convolutional layers, reducing the number of parameters.

Main Results:

• The proposed MicronNet-BN-Factorization (MicronNet-BF) model achieved an accuracy of 99.383% on the German Traffic Sign Recognition Benchmark (GTSRB).
• This represents a significant improvement over the original MicronNet's accuracy of 98.9%.
• Batch normalization was identified as the most influential factor in the accuracy improvement.

Conclusions:

• The MicronNet-BF model offers a compelling solution for traffic sign recognition in embedded systems.
• The model demonstrates high accuracy, reduced parameters, and efficient training, making it suitable for practical applications.
• Further enhancements through data augmentation also contributed to improved accuracy.