A Regression-Based Deep Learning Approach for Fish Fry Counting: Addressing Label Imbalance and Annotation Ambiguity

Manual counting of Nile tilapia fry is time-consuming and prone to human error, particularly under high-density conditions where visual overlap creates significant annotation ambiguity. Furthermore, real-world data collection often results in skewed label distributions. To address these challenges, this study proposes a regression-based fry counting system using EfficientNetV2S with transfer learning. Using a dataset collected from a real hatchery environment, targeted count-level balancing and rigorous annotation refinement were applied prior to model training to improve label consistency and data distribution. The novelty of this study lies in the integration of a count-level balancing strategy and a systematic annotation refinement process to address label imbalance and annotation ambiguity, two critical issues that are often overlooked in regression-based fish fry counting. The proposed model was evaluated using standard regression metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), coefficient of determination (R²), and Mean Absolute Percentage Error (MAPE). Experimental results showed that the best-performing model achieved an optimized MAE of 13.25 fry, RMSE of 16.93 fry, R² of 0.985, and MAPE of 7.07%. Comparative experiments further demonstrated that the proposed EfficientNetV2S architecture outperformed EfficientNetB0, ResNet50, and MobileNetV2. These findings indicate that systematically addressing annotation ambiguity and balancing training data significantly contributes to robust fry count estimation performance in practical aquaculture applications.