Brands Logo Recognition with Semi-Supervised Learning Using a Robust YOLOv8 Detection Model

Semi-supervised object recognition has emerged as a prominent area of research within computer vision. It offers the potential to significantly decrease the necessity for costly bounding-box annotations. Although there has been considerable success, the current advancements primarily concentrate on two-stage detection networks such as Faster RCNN, while study pertaining to single-stage detectors receives limited attention. This paper centers its attention on semi-supervised learning applied to the advanced and widely adopted single-stage detection network YOLOv8. Our method uses only a minimal amount of labeled data to successfully carry out the training, where we have explored various approaches like data augmentation, student-teacher network, pseudo labeling, and transfer learning. Furthermore, we have refined YOLOv8 implementation to optimize the advantages offered by semi-supervised learning. To validate our approach, we performed extensive experiments on the challenging OpenLogo Dataset, which contains 27,000 images across a total of 352 classes. The results were obtained with a limited amount of labeled data and a substantial amount of unlabeled data.