LDF-BNN: A Real-Time and High-Accuracy Binary Neural Network Accelerator Based on the Improved BNext.

Area of Science:

• Computer Vision
• Machine Learning
• Hardware Acceleration

Background:

• Deep Neural Networks (DNNs) excel at feature extraction for industrial defect detection but are computationally intensive for edge devices.
• Traditional Binary Neural Networks (BNNs) offer efficiency but suffer from accuracy degradation.

Purpose of the Study:

• To develop an efficient Binary Neural Network (BNN) with a layered data fusion (LDF) mechanism to address accuracy and computational challenges in industrial defect detection.
• To design a hardware accelerator architecture optimized for the proposed LDF-BNN.

Main Methods:

• Constructed a Layered Data Fusion BNN (LDF-BNN) based on BNext, incorporating a layered data fusion mechanism to minimize bandwidth pressure and accuracy loss.
• Designed an efficient hardware accelerator architecture featuring multi-storage parallelism to enhance computational efficiency and performance for complex BNN models.

Main Results:

• The LDF-BNN achieved 72.23% accuracy, 72.6 FPS, and 1826 GOPs on the ImageNet dataset.
• Demonstrated high applicability to industrial defect detection, achieving 98.70% accuracy on the Mixed WM-38 dataset.
• Outperformed existing methods in comprehensive comparisons, balancing accuracy and computational efficiency.

Conclusions:

• The proposed LDF-BNN effectively mitigates accuracy degradation issues in BNNs while maintaining low computational and memory requirements.
• The developed hardware accelerator architecture enhances the performance of LDF-BNN models, making them suitable for real-time industrial defect detection on edge devices.