LED-Display Defect Detection Based on YOLOv5 and Transformer

In the case of small and medium-sized enterprises (SMEs), it is not easy to find a solution for predictive maintenance or product failure diagnosis because they use fewer infrastructures than large companies and use a low-volume production method. In SME manufacturing facilities, people inspect products directly, what leads to oversights. Defective products are delivered to customers leading to causes and remedies for defects, apologies and compensation to customers Automatic detection of product defects in SMEs, especially in manufacturing, is a very important process and reduce costs as well as optimize corporate management. In the development of this technology, fault detection of LEDs on Display boards such as air conditioners and air purifiers was done by visual inspection in the past, but the image was obtained with a machine vision camera and the YOLOv5 algorithm was used to automatically detect defects. Adding a transformer to YOLOv5 make it is easy to understand the relationship between objects or the global context relating to the whole image, recognize complex patterns or structures, and reduce training time by using a framework that is well suited for parallel processing. Replacing the current module with a transformer in the YOLOv5 network helped developing an efficient algorithm, and the LED fault detection technology was tested in the production line. We propose an improved model with greater accuracy and speed than the existing YOLOv5 model. Semiconductor part LED-Display is made, and after training with our improved proposal, excellent performance is obtained. For the advancement of MES, it is applied to the control process in the surface mount technology industry with many small devices according to the production. The mAP performance of our proposed model was 0.994, a significant improvement over the 0.889 of YOLOv5 model. The Precision of the upgraded MODEL was 0.988, which largely improved the result compared the YOLOv5 model score (0.889), and allowed the detection of objects in Reel units to be applied in the field during object inference.