Kibae Lee

Chong Hyun Lee, Kibae Lee

The unlabeled data are generally assumed to be normal data in detecting abnormal data via semisupervised learning. This assumption, however, causes inevitable detection error when distribution of unlabeled data is different from distribution of labeled normal dataset. To deal the problem caused by distribution gap between labeled and unlabeled data, we propose a semi-supervised anomaly detection algorithm using KL divergence (SAD-KL). The proposed SAD-KL is composed of two steps: (1) estimating KL divergence of probability density functions (PDFs) of the local outlier factors (LOFs) of the labeled normal data and the unlabeled data (2) estimating detection probability and threshold for detecting normal data in unlabeled data by using the KL divergence. We show that the PDFs of the LOFs follow Burr distribution and use them for detection. Once the threshold is computed, the SAD-KL runs iteratively until the labeling change rate is lower than the predefined threshold. Experiments results show that the SAD-KL shows superior detection probability over the existing algorithms even though it takes less learning time.

Chong Hyun Lee, Kibae Lee, Hyun Hee Yim

We propose a tensor-based domain alignment (DA) algorithm designed to align source and target tensors within an invariant subspace through the use of alignment matrices. These matrices along with the subspace undergo iterative optimization of which constraint is on oblique manifold, which offers greater flexibility and adaptability compared to the traditional Stiefel manifold. Moreover, regularization terms defined to preserve the variance of both source and target tensors, ensures robust performance. Our framework is versatile, effectively generalizing existing tensor-based DA methods as special cases. Through extensive experiments, we demonstrate that our approach not only enhances DA conversion speed but also significantly boosts classification accuracy. This positions our method as superior to current state-of-the-art techniques, making it a preferable choice for complex domain adaptation tasks.

Chong Hyun Lee, Kibae Lee

We propose DaigNet, a new approach to object detection with which we can detect an object bounding box using diagonal constraints on adjacency matrix of a graph convolutional network (GCN). We propose two diagonalization algorithms based on hard and soft constraints on adjacency matrix and two loss functions using diagonal constraint and complementary constraint. The DaigNet eliminates the need for designing a set of anchor boxes commonly used. To prove feasibility of our novel detector, we adopt detection head in YOLO models. Experiments show that the DiagNet achieves 7.5% higher mAP50 on Pascal VOC than YOLOv1. The DiagNet also shows 5.1% higher mAP on MS COCO than YOLOv3u, 3.7% higher mAP than YOLOv5u, and 2.9% higher mAP than YOLOv8.