Search

This paper presents a signal generation architecture for simulating arbitrarily intercepted low-probability-of-intercept (LPI) signals in a battlefield environment. Additionally, the performances of deep learning models in classifying the generated signals are compared in terms of the classification time, GPU memory usage, and accuracy. Previous studies have utilized ensemble learning to minimize signal classification time and enhance accuracy; however, an explicit criteria for the adoption of deep learning models have been lacking. This paper presents the analysis of the performances of 11 deep learning models on the basis of simulation results and proposes an ensemble model that utilizes MobileNet V3 Small as the main model and Densenet-169 as the sub-model. [ABSTRACT FROM AUTHOR]

In recent years, leveraging deep learning models trained on image data has become prevalent for addressing various challenges, such as anomaly detection. Adequate data acquisition is crucial for effectively training these models. However, in many domains, while normal data is abundant, abnormal data with variations is sporadically produced, posing challenges for its collection. Training deep learning models with such imbalanced darasets inevitably leads to class imbalance issues. To mitigate data scarcity, data augmentation techniques like image rotation and zoom are commonly employed. Excessive use of these techniques, however, risks overfitting to the training data. This paper proposes a novel approach to address this issue by generating abnormal data through image translation using CycleGAN and anomaly image synthesis using SinGAN. The study demonstrates the positive impact of integrating generated abnormal data into the training dataset on model performance, as evidenced by comparing the performance before and after the addition of synthesized images in a classification model [ABSTRACT FROM AUTHOR]

The efficiency of container loading in logistics is a critical factor directly affecting a company’s profitability. Since improper loading planning can cause space inefficiency, product damage, additional costs, and work requirements, the importance of skilled workers who establish and carry out appropriate loading plans is emphasized. This 3D Bin Packing Problem (3D BPP) is an important task in various industries such as logistics, warehouse management, and cargo transportation. In this study, the latest reinforcement learning methodology is applied and effective solutions to 3D BPP are presented. In particular, this study uses the Proximal Policy Optimization (PPO) algorithm of reinforcement learning for stable and effective learning, learns Multimodal models and Graph Attention Networks (GAT) models, and evaluates loading capacity in 3D environments. In this paper, we experiment with two models through 3D loading environment and data under the same conditions and find the best model to apply practically to the field environment by deriving the optimal model. [ABSTRACT FROM AUTHOR]

Real-world information exists in various modalities. To integrate this, research on multi-modal image matching has been conducted in the field of computer vision and deep learning. However, computer vision algorithms like SIFT and ORB face challenges in producing good results due to domain differences between EO and IR images. There's also a limitation in deep learning models due to data acquisition challenges. This paper addresses these issues by generating synthetic images resembling IR images from EO images and trained the deep learning-based feature point extraction model R2D2 in multi-modality. Experiments with EO and IR image pairs in various environments showed that the image matching success rate of R2D2 in multi-modal scenarios is only about 4.6% different compared to ORB in single modality, demonstrating successful matching. This confirms the potential of using synthetic data for multi-modal training in R2D2. [ABSTRACT FROM AUTHOR]

Pancreatic Ductal Adenocarcinoma (PDAC) is the most common and deadly form of pancreatic cancer. Currently, histopathological diagnosis and prognosis of PDAC are time-consuming and labor-intensive for pathologists. Recent advances in pathological AI research aim to alleviate this. We accumulated training data, distinguishing PDAC areas in Whole Slide Images (WSIs) based on medical findings. Using this data, we trained a deep convolutional neural network for supervised learning to automatically interpret PDAC areas. The AI model achieved high Dice scores and, by visualizing the segmentation results of the predicted histological images, validated that PDAC diagnosis and identification of associated regions are automatically possible, similar to pathologists. Additionally, the AI model, which showed high specificity, suggests its potential as a co-pilot for pathological diagnosis and annotation. [ABSTRACT FROM AUTHOR]