http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
극소수 예시 오픈 세트 식별 문제에서의 비관측 범주 이미지 검출 성능 향상
구인용(Inyong Koo),정민기(Minki Jeong),김창익(Changick Kim) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.11
Few-shot open-set recognition is the task where a small set of support images is given, and the objective is to either classify the query images to seen classes or reject as unseen. Despite its practicality, the subject is yet less explored. In this paper, we adapt a few-shot image classification model to the open-set recognition task and suggest a better criterion for detecting unseen class images. Our method achieves better AUROC in unseen class detection without any trade-off in classification performance.
메트릭 학습 기반 오픈세트 극소수 학습에 대한 분류 성능 향상
박근철(Keunchul Park),정민기(Minki Jeong),김창익(Changick Kim) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.11
To conduct image classification with limited data is a key challenge for image recognition. Although several few-shot learning models already show good performance in closed-set classification, the actual environment for which deep learning is applied is open-set classification rather than closed-set classification. To achieve good open-set classification performance, we apply meta-learning in our model and suggest a layer normalization technique to make class decision boundaries more accurate. Our method shows better classification accuracy and unknown class sample detection capability, compared with previous few-shot learning methods.