이미지 기반 AI 모델의 품질 고도화를 위한 이미지 증강 및 테스트 방법의 융복합 연구 = A Convergence Study on the Image Augmentation and Validation Method for Quality Improvement of Image-Based AI Model

다국어 초록 (Multilingual Abstract)

This study was initiated to solve the difficulty of constructing a test set without bias for each image feature to verify the quality of an image-based artificial intelligence model. In previous studies, it was difficult to use because a large amount of labeling data was required for model learning. The purpose of this study is to improve the performance of artificial intelligence models by organizing learning and test datasets without bias and testing them by cluster without manual classification or labeling. Two experiments were conducted on the artificial intelligence model DC-VTON of the virtual try-on and the results were compared. First, the test set was divided into five clusters each using two clustering methods and the model was tested. In the two methods, the deviation of the model quality indicator SSIM values of the clusters was compared. The deviation of convolutional neural network-based clustering was 0.022 larger than that of pixel value-based clustering, resulting in more differences in performance by cluster. Second, two experiments were conducted to add re-learning data for each cluster to reach the target value of the model's quality performance index. Comparing the SSIM average of clusters that had low performance with after re-learning, the differential addition method by cluster increased by 3.2% and the equal addition method by cluster increased by 0.64%. In clusters with low quality, complementing the learning data resulted in improved performance improvement. Based on the results of these studies, it is expected that the quality of image-based artificial intelligence models can be improved by clustering images, which are unstructured data, according to characteristics, to refine test sets and enhance cluster-specific test results.

참고문헌 (Reference)

• 1 서정민 ; 이우진, "패션과 인공지능을 융합한 2D 이미지 기반 가상 착의의 복합적 특성 개선 연구" 한국전시산업융합연구원 40 (40): 193-203, 2022
• 2 Minaee, Shervin., "Image Segmentation Using Deep Learning : A Survey" 44 (44): 2022
• 3 National Information Society Agency, "Guidelines for Data Quality Management for Artificial Intelligence Learning" Jinhan M&B 158-181, 2022
• 4 Zou, Xingxing., "FashionAI : A Hierarchical Dataset for Fashion Understanding" 2019
• 5 Ge, Yuying., "DeepFashion2 : A Versatile Benchmark for Detection, Pose Estimation, Segmentation and Re-Identification of Clothing Images" 2019
• 6 Breck, Eric., "Data Validation for Machine Learning" 2019
• 7 Redyuk, Sergey., "Automating Data Quality Validation for Dynamic Data Ingestion" 2021
• 8 Biessmann, Felix., "Automated Data Validation in Machine Learning Systems" 44 (44): 2021
• 9 Yadav, Sanjay., "Analysis of k-fold cross-validation over hold-out validation on colossal datasets for quality classification" 2016
• 10 Kohavi, Ron, "A study of cross-validation and bootstrap for accuracy estimation and model selection" 14 (14): 1995