Author(s): Shijun Pan; Keisuke Yoshida; Ryo Furutani
Linked Author(s): Keisuke Yoshida
Keywords: No keywords
Abstract: These years, Okayama Prefecture is advancing in releasing juvenile Ayus, building spawning grounds, and establishing a fishing ban to achieve the goal of increasing Ayu production. At present, in order to observe the activity status of the juvenile Ayus, the Fisheries Research Institute of the Okayama Prefecture Agriculture, Forestry and Fisheries Research Institute has installed an underwater camera at the Kamogoshi Weir of the Yoshii River. The purpose of setting under water camera is to measure the amount of juvenile Ayus running up the fish way through video for qualitative evaluation by visual counting. Nevertheless, visual counting is a labour-intensive task, and technicians with professional experience are required to distinguish Ayus from other fishes quickly and accurately. Therefore, in order to finish this kind of observation without intensive labour, automatic counting is imperative. Currently, the object detection technologies in the field of artificial intelligence has become matured. Based on the above situation, we created a juvenile Ayu dataset derived from Yoshii River camera and used the YOLOv5 model to automatically identify and calculate the number of juvenile Ayus. In this study, the videos taken by the underwater camera are segmented at 2-second intervals as a segmented image, and the segmented images are divided into two groups (training group and validation group). The dataset is marked by the professional personnel with relative experience. At last, the YOLOv5s model detected juvenile Ayus in the 86400 images derived from 4-days long video with the recognition accuracy of 0.74. Recognition accuracy is based on the F1 score with value from 0.0 (0%) ~1.0 (100%), that is a kind of index derived from confusion matrix.
Year: 2022