http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
CNN을 활용한 램프 점등 이미지 균일도 정량화 방안 연구
이수현(Soohyeon Lee),이수봉(Soobong Lee) 한국자동차공학회 2022 한국자동차공학회 학술대회 및 전시회 Vol.2022 No.11
The lighting image of a car lamp is one of the important factors influencing the exterior design of a car, but it is difficult to find an evaluation method for quantitatively evaluating the lighting image. If there is a method of quantitatively evaluating how uniform the lighting image of the lamp is, it will be easy to manage the quality of the lighting image. This study attempted to determine whether it is possible to quantitatively evaluate the lighting image data of a car lamp using a deep learning algorithm that learned a good lighting image and a poor lighting image of a car lamp using deep learning of image data. CNN algorithms were used for deep learning of image data, and learning was conducted using images of various automobile position and tail lamp lighting. The learned algorithm classified good lighting images and poor lighting image data with a probability of more than 90% for the new lighting image data, and was able to provide lighting uniformity scores for the image data. The method of learning deep learning algorithms could help develop evaluation methods that are difficult to develop in conventional ways or quantitative evaluation methods for complex data.
가상의 학습데이터 생성 및 Pix2pix GAN 학습을 이용한 얼굴 그림자 제거
이수현(Soohyeon Lee),최영우(Yeongwoo Choi) 한국디지털콘텐츠학회 2021 한국디지털콘텐츠학회논문지 Vol.22 No.2
Various lighting changes are one of the factors that degrade the recognition performance of face images, in particular, when a shadow is formed on a face image due to lighting or surrounding environment, it is a general tendency that recognition performance is greatly degraded. Therefore, if the face image in which the shadow has occurred can be restored to its original state, improvement in face recognition performance can be expected. In this study, we propose a method of mitigating and removing shadows using Pix2pix, one of the representative models of GAN, a hostile generating neural network. Since the Pix2pix GAN model requires a pair of images corresponding to training, for this, we propose an idea to create a virtual training image corresponding to it from a normal face image using various image blending methods and use it as a training pair. Results of testing the model trained using the data generated by the proposed method, it can be seen that the shadows of the face image are naturally reduced and that the facial recognition performance is improved.