Few-Shot Image Classification: Current Status and Research Trends
Overfitting
Transfer of learning
Contextual image classification
Sample (material)
DOI:
10.3390/electronics11111752
Publication Date:
2022-06-01T07:33:18Z
AUTHORS (6)
ABSTRACT
Conventional image classification methods usually require a large number of training samples for the model. However, in practical scenarios, amount available sample data is often insufficient, which easily leads to overfitting network construction. Few-shot learning provides an effective solution this problem and has been hot research topic. This paper intensive survey on state-of-the-art techniques based few-shot learning. According different deep mechanisms, existing algorithms are divided into four categories: transfer based, meta-learning augmentation multimodal methods. Transfer useful prior knowledge from source domain target domain. Meta-learning employ past guide new tasks. Data expand with auxiliary information. Multimodal use information modal facilitate implementation also summarizes datasets literature, experimental results tested by some representative provided compare their performance analyze pros cons. In addition, application outcomes fields discussed. Finally, few future directions identified.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (110)
CITATIONS (37)
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....