Learning schemes for data imputation and object detection in data scarcity environments

Abstract

Recently, deep learning has seen extensive research across various fields, leveraging large datasets with labels. However, the efficiency of deep learning models based on large datasets diminishes when labeled data is scarce or absent. Consequently, there is a need for research on effective learning strategies to train deep learning models in data scarcity environments successfully. This study classifies data scarcity environments into semi-supervised, few-shot, and zero-shot learning. Appropriate learning schemes and network structures are proposed for each scenario to train deep learning models effectively in these environments. Given the potential occurrence of data scarcity in various domains, this research focuses on data imputation and object detection problems. In semi-supervised learning, a learning strategy called Random Drop Imputation with Self-training (RDIS) is proposed to address the issue of missing values in time-series data. RDIS aims to efficiently utilize missing data in time series and enhance the performance of missing value generation. The study explores the incremental few-shot object detection problem in the few-shot learning environment and introduces the incremental Two-stage Fine-Tuning Approach (iTFA). iTFA presents specialized learning schemes and network structures for the incremental few-shot object detection problem, demonstrating improved performance. Lastly, the study concentrates on the open-vocabulary object detection task in the zero-shot learning environment. Language-Aware RPN for Open-vocabulary Object Detection (LADet) is proposed to enhance object detection performance in a zero-shot environment, integrating RPN and a visual-language model. This comprehensive research contributes to advancing training strategies and network structures for AI models in diverse data scarcity environments.