Issue link: https://iconnect007.uberflip.com/i/1379105
JUNE 2021 I SMT007 MAGAZINE 69 e selection of a suitable model type, as demonstrated in the example in Table 1, is determined by the OT and IT requirement of a particular inspection use case. e nature of the object under inspection, the through- put requirement, and the IT operation budget are three important factors in determining the most suitable model type. Architecture Requirement for Large Scale Deployment e selection of a suitable model type and the training to optimize accuracy ensures a good AI model that will be ready for produc- tion deployment to manufacturing floor. How- ever, a new set of IT challenges arise in terms of practical deployment of the trained AI mod- els onto manufacturing floors. High availability Quality inspection is the final gate before product shipment to clients. Availability of the inspection system has direct impact to ship- ment schedule and, hence, revenue. e sys- tem must have 24/7 availability with minimum maintenance intervals. Scalability and performance Manufacturing operation can span multiple geographies. e system needs to be scaled to Tiny YOLO model is primarily optimized for speed and can be run anywhere but might not be as accurate as those optimized for accuracy, especially for use cases where small objects need to be classified. Single Shot Detector (SSD) model is suitable for real-time inference and embedded devices. It is almost as fast as YOLO but not as accurate as Faster R-CNN. Four models are trained and tested using training data set and test data set collected from actual quality inspection in IBM manu- facturing facility. is quality inspection use case requires classifying small objects with high accuracy. e key performance data is summarized in Table 1. e results were as expected, matching with the characteristics of different models. As this use case involves small object detection, SSD, and Tiny YOLO do not meet the accuracy requirement. erefore, only Faster R-CNN and Detectron are suitable for this use case. Both Faster R-CNN and Detectron have infer- ence time of 4-7 seconds per image, which also meets the operational requirement for this inspection. e difference in inference mem- ory usage between Faster R-CNN and Detec- tron has more impact in IT operation and cost than the quality requirement. e differences will be discussed later in this paper. Table 1: Model performance summary. * Tiny YOLO v2 was used. Due to the low accuracy, inference was not performed. The inference time is expected to be less than other model types. ** GoogLeNet was not selected for this test as the use case required the use of an object detection model that GoogLeNet does not support. *** YOLO V3 evaluation is ongoing at the time of writing this paper. Results will be shared in future publication.