|
Advancements in Electro-Optical and Infrared Maritime Detection: A Survey |
|
Changyeol Jeong, Jeonghoon Park, Dayeong Kim, Seunghan Kim, Jeongseop Seo, Sungho Kim(Yeungnam University, Korea) |
|
|
Object Detection using Reinforcement Learning |
|
Keoung Hun Choi, Jong-Eun Ha(SeoulTech, Korea) |
| Through the construction of an object detection model using a deep neural network structure, faster and more accurate results were obtained than traditional models. After that, the transformer structure was applied to simplify the model structure, and it became possible to detect the size of the object more freely. However, due to the complex configuration of the data used for training, it is difficult to generate new data. This study presents a method of learning object detection models through reinforcement learning using simpler data. As a result of the training, the value of ap is 40.81. |
|
|
DETR with Additional Object Instance-Specific Features for Encoder |
|
Wang Yao, Jong-Eun Ha(SeoulTech, Korea) |
| Based on DETR-likely object detector, we observed that DETR and DETE-like models include backbone and encoder that have same effect on the image, that is, they both did the same feature extraction function. We propose to add additional embedding module, which represents the full class information, and establishes global attention between feature tokens to provide prior knowledge for the extractor. |
|
|
Adaptation of Synthetic SAR to Measured SAR using Complex Value Conditional GAN |
|
MinJun Kim, Sungho Kim(Yeungnam University, Korea) |
| There are studies that compensate for these disadvantages by acquiring synthetic SAR using CAD models and simulations, but the domain gap occurs due to problems such as differences in speckle strength and mean variance of data values. We propose a CGAN model based on complex network that can reduce effectively the domain gap by utilizing the complex number information of SAR data. |
|
|
Face Image-Based BVP Estimation Using a 2D Camera |
|
Seongryeong Lee, Sungho Kim(Yeungnam University, Korea) |
| Methods for estimating heart rate based on facial images suffer from high loss and difficulty correlating with other biomarkers. Therefore, a similar method is needed to estimate blood volume pulse (BVP), which can estimate oxygen saturation, blood pressure, etc. in addition to heart rate. In this paper, we estimated BVP using a 2D camera and checked the heart rate results for each scenario, and compared the features of the actual and predicted data using STFT. |
|
|
Analysis of technique characteristics based on YOLOv5 for real-time small object detection |
|
Yeonha Shin, Sungho Kim(Yeungnam University, Korea) |
| Small object detection using deep learning is required in various fields such as military and medical as well as security/surveillance using surveillance and reconnaissance drones. We apply some techniques to YOLOv5 to build a real-time small object detection model, compare and analyze the results. We compare the contribution of 1) sufficient feature maps fusion through a heavier neck than base model, 2) replacing up-sampling with Transposed Convolution to efficiently increase the feature map size. Each technique led to improved performance in detecting small objects than the existing models, and we will continue research to build a model with better performance by combining these techniques |
|