抄録
In deep learning, large and accurate supervised datasets are important. Datasets are created by organizing source images and annotating them with information according to the task to be trained using machine learning. However, if a dataset is not available, the creation of a supervised dataset is required in addition to the usual work. Software to assist annotation has been proposed, but it requires work in addition to the usual tasks. We propose a deep learning dataset generation cycle in which the annotation work is included in the image measurement work. In this paper, a deep learning dataset generation cycle was added to the image measurement support tool TouchDeMeasure and developed. The proposed system was evaluated by comparing the annotation time and annotation results of conventional and proposed methods for a sampling survey of clams. The results show that the proposed method reduces the annotation time to approximately 2 to 3 seconds, while the previous method takes approximately 10 to 19 seconds per annotation. Using the dataset obtained by the proposed method, we constructed and evaluated a learning model using Mask R-CNN and obtained 99.2% accuracy, demonstrating the effectiveness of the proposed method.
