本研究では、現実世界では十分に収集する事が難しい道路損傷画像において敵対的生成ネットワーク（GAN）により疑似的な道路損傷画像を生成し、新たな教師データとして利用することで精度指標の一つであるF値を5%向上することを示しました。また、2018年に同誌で発表した論文でリリースした9,053枚の画像データで構成されるRDD2018を拡張する形で、13,135枚のデータで構成されるRDD2019をリリースしました。
Machine learning can produce promising results when sufficient training data are available; however, infrastructure inspections typically do not provide sufficient training data for road damage. Combining a progressive growing GAN along with Poisson blending artificially generates road damage images that can be used as new training data to improve the accuracy of road damage detection. The addition of a synthesized road damage image to the training data improves the F-measure by 5% and 2% when the number of original images is small and relatively large, respectively. In addition, we released RDD 2019, which consists of 13,135 image data, as an extension of RDD 2018, which consists of 9,053 image data released in the paper published in the CACAIE in 2018. All of the results and the new Road Damage Dataset 2019 are publicly available (https://github.com/sekilab/RoadDamageDetector).
【論文情報】
Maeda, H., Sekimoto, Y., Seto, T., Kashiyama, T. and Omata, H., Generative adversarial network for road damage detection, Computer-Aided Civil and Infrastructure Engineering, Wiley, Vo.36, pp.47-60, Available online 2 June 2020. DOI (Impact factor in 2019: 8.552)
