OFDN-YOLO: an optical film defect detection method based on improved YOLOv10

Abstract

Abstract Optical films refer to thin layers of materials with special optical properties that are coated or fabricated on optical components or independent substrates. In modern production processes, various defects often occur on its surface, which seriously affect the yield of optical films. This paper proposes a network model based on the improved YOLOv10 to achieve efficient and accurate optical film defect detection (OFDN-YOLO). Replace the Convolution layer in the backbone network of the YOLOv10 model with the Reference Convolution layer. Introduce the Content-Aware ReAssembly of Features sampling method and the Bidirectional Feature Pyramid Network pyramid network structure to its neck network. And the Super Token Attention global attention mechanism has been added. The improved network has enhanced the detection rates of various types of defects in optical films. The precision reaches 96.8%, the recall reaches 95.8%, and the mean average precision reaches 96.2%, meeting the requirements for the detection rate of film defects in actual industrial production. Among them, the minimum diameter of the detectable defects reaches 10 μm, achieving precise and efficient detection of multiple defects in optical films. Meanwhile, the stability of the improved model was verified through the illumination robustness test. The research conducted in this paper provides a new and effective solution for the field of optical film defect detection.