Font Size: 

In the actual production, a large amount of workers are needed for quality inspection, they inspect various defects by subjective eyesight and experience. This paper proposes a multi-class defect detection network for cookers(CDDNet). It also designs a training procedure to preserve end-to-end model accuracy post quantization. It applys CDDNet training and testing the ability of multi-classification and transfer learning, and show that multi-classification and transfer learning can be successfully applied using image data from an entirely different domain. The database is obtained from many days of automated camera recordings. As a result, the proposed quantization scheme improves the tradeoff between accuracy and device latency.

defects detection; CDDNet; multi-classification; transfer learning