Body Mass Estimation of Large Yellow Croaker Using Measurement-Oriented Instance Segmentation and Selective Stereo Measurement

Noncontact fish body mass estimation requires accurate body extraction and reliable physical scale under changing underwater observations. This study develops a measurement-oriented instance segmentation and selective stereo-measurement method for large yellow croaker (Larimichthys crocea), centered on fin-excluded projected body area. The model combines coordinate attention, area-aware consistency loss and area-bias-guided boundary loss within Mask2Former. The stereo workflow integrates visual measurability, geometry, depth and temporal checks before track-segment aggregation. Among three phenotype regressions fitted to 104 reference fish, the area model achieved the lowest leave-one-out mean absolute percentage error (MAPE) of 6.03%. On the same 74 measurable ground-truth test instances, mask-area MAPE decreased from 1.7417% to 1.3679%, a 21.46% reduction; segmentation average precision increased from 0.769 to 0.791. Across four videos, the complete model produced an area-based segment-averaged mass of 106.60 g against a reference cohort mean of 114.04 g. Absolute mean bias decreased from 8.14% for Mask2Former to 6.52%, with observation acceptance changing from 13.83% to 6.43%. The method connects measurement-oriented segmentation to quality-screened phenotype and mass estimates evaluated through reference regression, matched mask areas and video cohort-mean comparison.