2020 Fiscal Year Final Research Report
An efficient learning sample generation in small sample problem for deep learning
Project/Area Number |
18K11495
|
Research Category |
Grant-in-Aid for Scientific Research (C)
|
Allocation Type | Multi-year Fund |
Section | 一般 |
Review Section |
Basic Section 61050:Intelligent robotics-related
|
Research Institution | Chukyo University |
Principal Investigator |
|
Co-Investigator(Kenkyū-buntansha) |
村瀬 洋 名古屋大学, 情報学研究科, 教授 (90362293)
道満 恵介 中京大学, 工学部, 准教授 (90645748)
|
Project Period (FY) |
2018-04-01 – 2021-03-31
|
Keywords | 深層学習 / 画像生成 / 物体検出 / 人の知覚 |
Outline of Final Research Achievements |
This research aims to improve the performance of deep learning for image recognition when the number of available training samples is limited. For this purpose, we have developed a method that maximizes the effectiveness of the training sample generation method. In lesion detection from medical images, it is often difficult to collect a sufficient number of samples. Generating images that account for the target object's location (e.g., its position in the organ, we have shown that the detection accuracy can be improved even with a small number of training data ). In addition, when estimating the characteristics of human perceptual functions, we found that large image deformations can degrade the estimation performance.
|
Free Research Field |
画像認識
|
Academic Significance and Societal Importance of the Research Achievements |
深層学習による画像認識や物体検出において,様々な理由から学習サンプルを多く集めることが困難な問題がある.そのような場合に画像合成や画像生成により学習サンプルを増加させる方法が提案されている.その際に,対象となる物体が存在する状況を想定した画像生成をおこなうことで,単純な画像生成手法に比べて認識精度が向上できることを示した.また画像を生成する場合に付与する画像変形の程度は,対象の画像や推定したい事柄によって変化させないといけないことを示した.
|