PCB Defect Recognition by Image Analysis using Deep Convolutional Neural Network

Printed circuit board (PCB) is one of the most important components of electronic products. The traditional defect detection methods are very difficult to meet the defect detection requirements in the PCB production process. In recent years, the Convolutional Neural Networks (CNNs) have developed rapidly and have shown greater advantages than traditional methods in the field of machine vision. In order to reduce the workload of manual inspection and improve production efficiency, the PCB defect image recognition method based on the convolutional neural network is studied in this paper. Three convolutional neural network classification models VGG16, InceptionV3 and ResNet50 are studied. The test results based on the PCB image data set show that the ResNet50 model has better PCB defect image classification capabilities than the VGG16 and InceptionV3 models. It also shows that the classification accuracy of the ResNet50 model can be improved by data augmentation methods, even without increasing the number of PCB image samples. Based on the analysis of the ResNet50 network structure, an improved network model structure is developed. In the ResNet50 model, a new CNN module Res2Net is introduced, which replaces the residual block of the original convolutional layer with a more layered residual connection structure. The Rectified Linear Unit (ReLU) function is used as the activation function behind each BottleNeck to improve the non-linear expansion capability of the network. The experimental results under the same conditions show that the improved ResNet50 model has better classification performance in PCB defect classification tasks.