Research on a lightweight electronic component detection method based on knowledge distillation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Accurate and rapid detection of electronic components is vital for efficient electronic component assembly.
• Existing methods may lack the necessary balance between precision and computational complexity.

Purpose of the Study:

• To propose a lightweight electronic component detection method using knowledge distillation.
• To enhance the performance of a compact student model by learning from a larger teacher model.

Main Methods:

• Construction of a lightweight student model for electronic component detection.
• Implementation of a novel knowledge distillation technique combining feature and channel distillation.
• Training the student model to capture rich class-related and inter-class features from the teacher model.

Main Results:

• Student model parameters reduced by 55% (13.32 M) and FLOPs by 35% (28.7 GMac) compared to the teacher model.
• Mean Average Precision (mAP) improved by 3.91% on Pascal VOC and 1.13% on the electronic components dataset.
• Achieved a detection precision (mAP) of 97.81% with a speed of 79 FPS.

Conclusions:

• The proposed knowledge distillation method effectively balances model precision and complexity.
• The lightweight student model enables fast and accurate electronic component detection.
• This method offers a superior solution for real-time electronic component detection in assembly lines.