Enhanced deep learning models for automatic fish species identification in underwater imagery.

Area of Science:

• Marine Ecology
• Computer Vision
• Deep Learning

Background:

• Underwater cameras are vital for marine ecology research.
• Managing vast amounts of underwater imagery requires automated species identification.
• Current methods face challenges in accurately identifying fish species from complex visual data.

Purpose of the Study:

• To develop an efficient and accurate deep learning framework for automated fish species identification.
• To enhance image preprocessing and feature extraction for improved detection and classification accuracy.
• To optimize model hyperparameters using an advanced metaheuristic algorithm.

Main Methods:

• A two-stage deep learning approach combining image preprocessing with advanced convolutional neural networks.
• Stage 1: Unsharp Mask Filter (UMF) preprocessing followed by an enhanced region-based fully convolutional network (R-FCN) with precise region of interest (PS-Pr-RoI) pooling.
• Stage 2: Integration of ShuffleNetV2 with the Squeeze and Excitation (SE) module (Improved ShuffleNetV2) for enhanced classification, with hyperparameters optimized by the Enhanced Northern Goshawk Optimization Algorithm (ENGO).

Main Results:

• The enhanced R-FCN model achieved 99.94% accuracy, 99.58% precision and recall, and 99.27% F-measure on the Fish4knowledge dataset.
• The ENGO-optimized ShuffleNetV2 model demonstrated 99.93% accuracy, 99.19% precision, 98.29% recall, and 98.71% F-measure.
• Both models exhibited superior performance in fish detection and classification.

Conclusions:

• The proposed two-stage deep learning framework significantly improves automated fish species identification accuracy.
• The integration of UMF, enhanced R-FCN, Improved ShuffleNetV2, and ENGO offers a robust solution for marine image data management.
• This technology has the potential to advance marine ecological studies through efficient and precise data analysis.