A Novel Rotated YOLOv8 for Mini LED Multi-Class Defect Detection

In recent years, with the increasing demand for high-precision displays, Mini light-emitting diodes (Mini LED), as an important component of displays, have relatively high quality requirements for their production. However, various defects are inevitable during the production process. Therefore, the detection of product defects becomes crucial. Due to the complex texture of the Mini LED substrate and the diverse defect forms, with significant differences in shape, angle, and scale, traditional manual detection is inefficient and has limited accuracy, while conventional classification or object detection networks are difficult to effectively deal with such geometric deformations, and the object detection level box is difficult to fit the complete shape of the defect. In response to these, we propose the SD-YOLOv8 rotated object detection method based on YOLOv8, which is specifically designed for the defect detection task of Mini LED. To enhance the detection capability of small object defects, we add a small object detection head to the shallow layerularly rotating objects, and in order to enhance the feature extraction capability for defects of irregin the neck part, we integrate deformable convolutional on the Cross Stage Partial (CSP) bottleneck with 2 Convolutions (C2f) module in the neck part. Adaptive sampling of the rotation and deformation regions is achieved by using learnable offset and modulation factors. In addition, a oriented bounding box (OBB) detection head is adopted to predict the object position and rotation angle more accurately. The experimental results based on the self-built Mini LED defect dataset show that our model's mAP50 reaches 96.7%, which is 2.3% higher than the baseline YOLOv8-OBB. The results verify that SD-YOLOv8 has superior performance of high precision and high efficiency in the industrial inspection scenarios of Mini LED.