Semi-automatic recognition of juvenile scallops reared in lantern nets from time-lapse images using a deep learning technique

Abstract

Image analyses based on deep learning techniques have been expanding rapidly in various fields, and the development environment has become more accessible. In our previous study, numerous time-lapse images (approximately 9000 images) of juvenile Yesso scallops reared in lantern nets were acquired, but only a few have been used for data analysis owing to the difficulty associated with the automatic processing of such vast numbers of images. In this study, an algorithm was developed for the automatic detection of scallops from training and test images selected from the time-lapse images (195 images) based on the use of a deep learning technique referred to as “semantic segmentation.” The developed algorithm recognized the juvenile scallops with high accuracy in the test images. The algorithm was also applied to the other time-lapse images, and high accuracy was confirmed by visual inspection with the exception of a few cases. Data analysis was conducted within the automatically recognized areas of the scallops to explain their growth and behavioral changes owing to stormy weather. Thus, numerical analyses conducted based on semi-automatic processing of massive time-lapse image data may be useful for ecological studies of cage-cultured scallops.