Abstract: In response to the solution of the problem that the image defects of diode glass bulb are mostly detected by manual feature extraction and the recognition is of low accuracy, an improved VGG network for diode glass bulb image defect detection is proposed. Firstly, the glass bulb image is preprocessed. At the same time, the VGG-19 model of convolution neural network structure is pre-trained by the original large sample data set to obtain the pre-training model. Then, part of the weight parameters of the pre-training model, such as convolution and pooling, are transferred to the fixed layer of the improved network model by the method of transfer learning. In this model, the non-fixed layer is used for improvement and the full-connection layer structure of the network is re-set and optimized. Finally, the improved model is trained with the preprocessed image data set of the glass bulb, and the parameters and weights of the non-fixed convolution layer and the new full-connection layer are obtained. The results of the experiment on the data set of the diode glass bulb show that the method can effectively improve the accuracy of the classification and recognition of the diode glass bulb defect detection and the accuracy rate can reach 98.3%.

Key words: defect detection of diode glass bulb image, VGG-19 model, transfer learning, image classification, accuracy rate 
 
 defect detection of diode glass bulb image, VGG-19 model, transfer learning, image classification, accuracy rate