CNN Aided Surface Inspection for SMT Manufacturing

Abstract

Automated optical inspection (AOI) is a visual defect inspection system. The semiconductor industry has a strong dependency on AOI for defects screening. Conventional AOI is inadequate for some inspections, especially surface defects like crack, chip and void, and the algorithms are inefficient in isolating the defects from product variants. Convolutional Neural Network (CNN) had been broadly studied and adopted to replace the conventional AOI in surface inspection. There are many CNN architectures developed in the past decade for image classification, such as AlexNet, GoogLeNet, ResNet, VGGNet, etc.; each with its own strength in terms of accuracy and speed. The training process could be speeded up too using techniques such as transfer learning from pre-trained CNN models. Newer techniques in vector programming on kernels, e.g., Single Instruction Multiple Data (SIMD) and depth wise separable method can further increase the efficiency of convolutional layer activation functions. CNN algorithms for surface inspection are found to be very promising, with defect classification able to achieve accuracies of 91-99% on the wide range of products. The CNN result outperforms conventional surface inspection methods like edge detection and machine learning algorithms.