P‐98: Novel Solution for High‐Speed Detecting Very Small Defects in Display

Abstract

There are various inspection stages in the process of display manufacturing. Accurate detection of defects in various sizes is crucial for high yield rate. Detecting tiny defects in an image is a challenging task for object detection techniques based on convolutional neural networks because their feature information can be lost in deep convolutional layers. Also, our task requires defects in various sizes are detected in a high‐speed to achieve high yield rate. We tackled this challenging issues in our object detection problem by establishing three theories: Slicing and Robust of Outlier, Multi‐Inference. For the defect dataset that occurs in mass production, mAP@0.5 78.5% for Robust Dataset is increased to 89.5%, and the existing method, which take 4.7 sec with SAHI method, is reduced to 0.2 sec.