Domain Adaptation for Agricultural Image Recognition and Segmentation Using Category Maps

Recognition accuracy obtained using deep learning drops precipitously when the training data are insufficient. This paper presents a data-expansion method for training of the transfer learning source domain. Using expanding images generated from weights on a category map as source data, we compared accuracies obtained from five derivative models and our previously reported method. Moreover, we obtained the result of domain adaptation between actual images and synthetic images using weights obtained during transfer learning. Based on those results, we verify whether the amount of training data can be expanded quantitatively and qualitatively. Experiment results obtained from two open benchmark datasets and our original benchmark dataset demonstrated that our proposed method outperforms the previous method under a guarantee of sufficient accuracy for the synthetic images.