Adaptive Confidence and IoU Optimization for Accurate Real-Time Eel Counting Using YOLO-ONNX

Accurate fish counting is important in aquaculture hatcheries because it supports production planning, inventory monitoring, and fish dispersal management. Eels' small size, elongated body structure, transparency during the glass eel stage, and overlapping movement make it difficult to count them using image-based methods because, in contrast to other fish species, eels are prone to occlusion. Counting mistakes, duplicate detections, and missing detections are frequently caused by these traits. This study used adaptive confidence and Intersection over Union (IoU) threshold optimization to develop and evaluate an improved YOLO-ONNX-based eel counting framework. 1,541 eel images made up of glass eel and elver samples were used to train a YOLOv8n model, which was then exported to the Open Neural Network Exchange (ONNX) format for use. For threshold adjustment, a different set of 333 eel counting images arranged according to manual ground-truth counts was used. A confidence threshold of 0.25 and an Intersection over Union (IoU) threshold of 0.70 were utilized in the baseline YOLO-ONNX configuration. To find the optimal counting configuration, the adaptive optimization method tried several combinations of confidence and Intersection over Union. After selecting the optimized threshold configuration, a separate 200-image evaluation set was used to compare the baseline and optimized YOLO-ONNX configurations. Based on the results of the 200-image evaluation set, the optimized setup increased the mean counting accuracy by 11.29 percentage points, from 80.95% to 92.24%, using a confidence threshold of 0.15 and an IoU threshold of 0.40. The Mean Absolute Error decreased from 2.58 to 0.91, while the Root Mean Square Error decreased from 3.88 to 1.42. Based on the results, adaptive threshold optimization enhances detection-based eel counting by lowering overcounting and undercounting brought on by eel body overlap and occlusion. The proposed YOLO-ONNX framework provides a practical approach for improving eel counting accuracy through post-processing threshold optimization, without modifying the internal structure of the YOLO model.