Recognition of Visible Disease Symptoms in Paralichthys olivaceus Using Progressive Candidate Correction

Recognizing visible disease symptoms in olive flounder (Paralichthys olivaceus) from red, green and blue (RGB) images requires distinguishing tissue changes from camouflage, reflections, and normal variations in body shape. To address this challenge, this study proposes progressive candidate correction, which retains class-specific symptom candidates and adjusts their responses using symptom location and visual characteristics as either supportive or inhibitory cues. Evaluation used 1461 RGB images split by individual. Original training and validation input preparation used ground-truth symptom annotations with fish masks to determine crops and background retention. Under this preparation, with training-calibrated thresholds fixed for validation, the method captured 71.9% of annotated symptom regions, compared with 49.3–63.1% for direct semantic segmentation. At separately calibrated operating points, capture before and after correction remained similar (72.0 and 71.9%), outside-annotation candidate area decreased from 93.1 to 21.7% of the fish region, and allowed-part agreement across class-specific candidates outside annotations increased from 42.0 to 92.2%. Supplementary full-model retraining using fish-mask-only input preparation yielded 71.2% capture, but lower allowed-part agreement, auxiliary segmentation overlap, and capture for some symptom classes. The proposed method could help prioritize visible symptoms for follow-up in aquaculture, serving as a screening aid rather than a substitute for diagnostic examination.