Detecting anomalies in graph-structured data is critical for identifying unusual patterns within complex systems, with applications spanning cybersecurity, fraud detection, and risk assessment. In this work, we present VQ-VGAE, a novel architecture that combines Vector Quantization (VQ) with the Variational Graph Auto-Encoder (VGAE) for unsupervised anomaly detection. By incorporating discrete latent variables, VQ-VGAE offers a new way to model intricate graph anomalies, which are often missed by traditional methods. The framework leverages the expressive power of Graph Convolutional Networks (GCNs) to learn graph structure while integrating the probabilistic approach of Variational Auto-Encoders (VAEs). Extensive experiments on both transactional and business process graph datasets reveal that VQ-VGAE consistently surpasses conventional Graph Auto-Encoder (GAE) and VGAE models, showing marked improvements in AUC-ROC and Average Precision (AP). Beyond performance gains, leveraging a discrete latent space facilitates interpretable anomaly representations, offering deeper insights into the nature of deviations within graphs. This interpretable approach to anomaly detection enhances our ability to understand and manage complex networks across diverse domains.

Lung cancer remains a significant global cause of mortality, affecting populations worldwide. Deep Learning (DL) systems show promise in early detection using clinical data to reduce mortality rates. However, these systems heavily rely on large amounts of annotated anomalous data, and heavily depends on selecting appropriate hyperparameters that define the network’s structure and learning process.In this study, we propose an optimization scheme based on Tree Parzen Estimator (TPE) and Bayesian optimization (BO) algorithms for hyperparameter optimization in unsupervised lung cancer detection. First, we used a fast residual attention GAN-based model. Then, we employed the Mixup consistency regularization technique to encourage the discriminator to attend the pixel-level details of the input data. Furthermore, a new cost function for the discriminator is defined based on the mixup to enhance the output. This study is evaluated in the context of lung cancer detection. When compared to empirical optimization, both TPE and Bayesian optimization demonstrate significant improvements in the precision of the fast Residual Attention GAN model used in this study. Various metrics, including precision, f1-score, and the area under the curve (AUC) are employed to assess the system’s efficiency.

In this paper, we address the issue of designing a smoke detector robust to illumination variations. Our contribution consists in resorting to color invariants as salient smoke features. More precisely, the proposed detector employs consecutively of an illumination invariant color representation, a photometric gain based background subtraction, a chrominance detection and a smoke identification based on two invariant color descriptors. The experimental results show that the proposed method can effectively detect smoke with robustness to illumination changes and noises, frequently encountered in wildfire video-surveillance environments.

In this paper, we present a probabilistic framework for edge and region grouping using conditional random field. Our model is built on a hybrid adjacency graph of atomic region and contour primitives. Unary and pairwise potentials that capture similarity, proximity and curvilinear continuity are defined. Similarity, for both region and edge cues, is measured by likelihood ratios learned from a human labeled ground truth. We use a stochastic graph partition algorithm, Swendsen-Wang Cut, to perform inference on this model. Experimental results are shown on gray-scale natural images.

We propose a non-homogeneous conditional random field (CRF) built over an adjacency graph of superpixels for contextual region grouping. Our model includes spatially dependent potentials that capture contextual interactions of the data as well as the labels. Both superpixels and segments are described with local statistics which take into account their contexts in the image. This results the non-homogeneity of the fields which improves the region grouping process of natural images. In our energy formulation, the similarity is measured by a likelihood ratio learned from a human labeled ground truth. The inference is performed using a cluster sampling method, the Swendsen-Wang cut algorithm. Results are shown on various natural images.

Motion-based segmentation is an important task in object-oriented video applications. Different approaches to motion segmentation with level sets. have been proposed. The key features of level sets representation are its ability to handle variations in the topology of the segmentation and its numerical stability. These approaches rely on certain prior information to perform motion segmentation. We present a new algorithm for segmenting an image into distinct regions of homogeneous motion. The two problems of motion estimation and motion segmentation are jointly solved without need of prior information. We provide experimental results on image sequences with synthetic and natural motion.