Optimized Detection of Red Devil Fish in Low-Quality Underwater Images from Lake Toba Using a Hybrid CNN and Transfer Learning Approach

The detection of freshwater fish in turbid underwater environments presents significant challenges due to poor image quality caused by low lighting, suspended particles, and visual noise. This study proposes an optimized detection model for Amphilophus labiatus (Red Devil fish) in the murky waters of Lake Toba, Indonesia, using a hybrid Convolutional Neural Network (CNN) integrated with transfer learning and visual enhancement techniques. The proposed architecture combines MobileNetV2 and ResNet50 backbones with CLAHE (Contrast Limited Adaptive Histogram Equalization) and median filtering to improve image clarity and feature extraction. A custom dataset comprising 3,500 annotated underwater images was used to train and evaluate the model. The hybrid model achieved a detection accuracy of 96.1%, a precision of 95.6%, a recall of 94.8%, and a mean Average Precision ([email protected]) of 0.941—outperforming baseline models such as YOLOv5 and Faster R-CNN. Visual diagnostics and Grad-CAM attention maps confirm the model's ability to focus on key anatomical features under varying image conditions. The architecture is optimized for real-time deployment on edge-AI devices, supporting conservation efforts and biodiversity monitoring in freshwater ecosystems