Issue link: https://iconnect007.uberflip.com/i/1379105
68 SMT007 MAGAZINE I JUNE 2021 of the manufacturing process for a particu- lar product. e inspection is usually per- formed on raw material (also known as incom- ing material inspection) and finished product (also known as outgoing quality inspection). In a high-complexity manufacturing process, such as wafer fabrication and integrated circuit packaging, inspection is also performed on in- process products as early in-line quality feed- back. Quality inspection covers a wide range of items, from appearance, color, marking, and label, to defects and scratches. An inspection process consists of two steps: image acquisition and image examination. Tra- ditionally both steps are performed manually. e image is acquired directly from the prod- uct by human vision of the inspector and exam- ined spontaneously by human cognition. As a result, traditional quality inspection is labor intensive and has great dependency on human skill and competency. As feature sizes of interest become too small for human vision, advanced equipment such as magnifying lens, microscope, and techniques such as back lighting, dark field, and X-ray are employed to obtain the images that are examin- able by human vision. Automation is enabled in image acquisition equipment and significantly improves the throughput of inspection pro- cesses for mass production. A good example is automated optical inspection (AOI) equip- ment used in printed circuit board inspection that has optical solution to micrometers. With the development of computer vision, rule-based algorithms are employed to par- tially replace human cognition for image examination, which further improves the effi- ciency and throughput of the inspection pro- cess. However, rule-based algorithms have limitations in object detection and classifica- tion and are normally used as a "coarse" screen of the images under inspection. It still requires human cognition to a great extent to accurately classify the images under inspection. In recent years, neural network-based deep learning models have demonstrated high accu- racy in object detection and classification in the area of digital image processing. AI mod- els start to show great potential to replace human cognition in the quality inspection process through object detection and classi- fication. us, AI-assisted quality inspection became very promising to fully automate qual- ity inspection processes. AI Models for Quality Inspection While AI models are not the primary focus of this paper, a brief introduction to types of models popular for object detection and clas- sification is important in understanding how they are used in quality inspection, consider- ing both accuracy and performance. e following are just a few of the mature AI model types that have been widely used in image recognition application and use cases: • GoogLeNet • Faster R-CNN (region-based Convolutional Neural Network) • Detectron • Tiny YOLO (you only look once) • Yolo V3 • SSD (single shot detection) GoogLeNet employs a 22-layer Convolu- tional Neural Network for image classifica- tion only, this means GoogLeNet cannot iden- tify individual objects in each picture but is able to identify each image as a single category. GoogLeNet can be exported to run on edge devices making it highly portable. Faster R-CNN, Detectron, and Yolo V3 mod- els are optimized for accuracy. ese models use rectangular bounding boxes to label the objects. Detectron models can also use objects labeled with polygons (segmentation) for greater training accuracy. However, training a data set that uses polygon labels takes longer than training with rectangular bounding boxes. YOLO V3 is an object detection image analy- sis model that has higher accuracy than Tiny YOLO, but requires more computer resources for both model training and model inference.