@techreport{oai:ipsj.ixsq.nii.ac.jp:00240444,
 author = {長谷川, 雄史 and 劉, 継康 and 伊藤, 大心 and 宮本, 健 and 木谷, 哲 and 井手, 美優 and 仁平, 雅也 and Takefumi, Hasegawa and Jikang, Liu and Daishin, Ito and Ken, Miyamoto and Tetsu, Kitani and Miyu, Ide and Masaya, Nidaira},
 issue = {19},
 month = {Oct},
 note = {道路インフラの点検管理では，経年劣化による路面損傷部分を検出して修繕する．路面損傷検出に深層学習の画像認識モデルを適用する際には多量の学習画像が必要となるが，路面損傷部分は少なく学習画像が少量になる傾向がある．このため画像認識モデル学習時に使用した学習画像のバイアスを軽減することで転移学習効率を向上させ，少量の学習画像であっても路面損傷部分を検出する画像認識手法を提案する．本手法では画像認識モデルの転移学習時に因果推論の介入処理を用いた Few-shot Learning を適用する．本手法の適用前後の画像認識モデルで路面損傷部分の検出性能を比較評価し，検出性能の向上を確認した．, Inspection and management of road infrastructure involves detecting and repairing road surface damage due to age-related deterioration. Image recognition models for detecting road surface damage require a large amount of training images, but the number of training images tends to be small because there are only a few damaged areas of the road surface. We propose an image recognition method that can detect new types of road surface damage with a small number of training images. This is achieved by improving transfer learning by reducing the bias in training images of image recognition model. This method reduces bias in training images by applying few-shot learning, incorporating causal inference during the transfer learning process of image recognition models. We evaluated the detection precision of the image recognition model before and after applying this method. The results showed an improvement in detection precision after applying this method.},
 title = {学習画像のバイアス軽減手法を用いた路面損傷検出},
 year = {2024}
}