AUTOMATED DEFECT DETECTION BASED ON TRANSFER LEARNING AND DEEP CONVOLUTION GENERATIVE ADVERSARIAL NETWORKS

Yangbo Feng,∗ Tinglong Tang,∗ Shengyong Chen,∗ and Yirong Wu∗

Keywords

Defect detection, deep convolution generative adversarial network, transfer learning

Abstract

Machine learning and computer vision methods are widely used in defect detection. It is challenging to collect enough samples for training, and the collected samples are usually in unbalanced distribution. In this work, we use deep convolution generative adversarial networks (DCGANs) to augment the data and transfer learning method to avoid training from scratch. Particularly, only the instances of rare class are augmented using DCGAN so that the data distribution will be balanced, and transfer learning is adopted by fine-tuning the pre-trained VGG16 network which is improved by using focal loss function in our method. The proposed method is applied to detect penicillin bottle defects. Experiments show that the proposed method outperforms the traditional models.

Important Links:

Go Back