Presentation Information
[2Yin-A-56]Size Measurement of Farmed Flounder using Monocular Image Segmentation
〇Ryusei Tsukuda1, Yuta Kayatani1, Kazuki Kobayashi1, Shuhei Tarashima1 (1. NTT DOCOMO BUSINESS, Inc.)
Keywords:
computer vision,land-based aquaculture,part-aware panoptic segmentation,body length measurement
Body length measurement is critical for flounder management in land-based aquaculture; however, manual methods are labor-intensive and induce physiological stress in fish. We propose an automated approach using monocular camera images captured from above the tank. Since flounder inhabit the tank bottom, pixel length in the image plane can be converted to actual size. Our method employs part-aware panoptic segmentation to divide the flounder into anatomical parts such as the head and tail, then estimates body length from the pixel length of each part and camera parameters. By learning joint object-part representations, this approach ensures robust measurement across varying fish postures. Experimental results using images from actual aquaculture facilities confirm the effectiveness of this approach; a relative error of 2.65% compared to the ground-truth length was achieved.