http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Wonho Jung(정원호),Sungjin Cheong(정성진),Jae Woong Bae(배재웅),Yong-Hwa Park(박용화) 대한기계학회 2021 대한기계학회 춘추학술대회 Vol.2021 No.11
It is important to recognize the drowning person as soon as possible in maritime accidents. In real maritime accidents, it is difficult to identify the drowning person because of their small size compared to the marine environment. To solve this problem, this paper presents a methodology to detect small target using commercial games with 3D graphical engines. Proposed methodology combines as following four steps: (1) divide high-resolution original image into several small patches, (2) image processing using CLAHE and Canny edge detection, (3) detecting small targets using convolutional neural networks (4) restore patches into original image. To detect small target in the high-resolution original image, small patches and image processing techniques are considered to raise the signal-to-noise ratio of the small target. The small patches are uses as test data of convolutional neural networks (CNN), the softmax values of each patch are displayed on the reconstructed image. To enhance the accuracy of CNN, virtual image data acquired from the commercial game using the 3D graphical engine are used as training data. In order to verify the performance of the proposed methodology, a case study of real maritime accident situation was conducted. The performance of the proposed methodology outperforms original deep convolutional neural networks.