Targeted Data Augmentation and Hierarchical Classification with Deep Learning for Fish Species Identification in...

Area of Science:

• Marine Biology
• Computer Vision
• Artificial Intelligence

Background:

• Underwater video offers non-disruptive fish monitoring but presents image analysis challenges.
• Accurate fish species identification is vital for marine biodiversity studies.
• Efficient image processing techniques are needed for unconstrained underwater environments.

Purpose of the Study:

• To develop and evaluate novel Deep Convolutional Neural Network (CNN) approaches for fish species classification in unconstrained underwater video.
• To enhance the accuracy and efficiency of automated fish identification systems for marine monitoring.

Main Methods:

• Proposed two Deep Convolutional Neural Network (CNN) strategies for fish species classification.
• Implemented a traditional transfer learning framework with a novel data augmentation technique based on training/validation loss curves.
• Developed a hierarchical CNN classification model for multi-level fish categorization (family then species).

Main Results:

• Achieved high accuracy rates on benchmark datasets: 99.86% on the Fish Recognition Ground-Truth dataset and 81.53% on the LifeClef 2015 Fish dataset.
• Demonstrated the effectiveness of both proposed CNN approaches in unconstrained underwater environments.
• Validated the utility of targeted data augmentation and hierarchical classification for fish identification.

Conclusions:

• The proposed Deep Convolutional Neural Network (CNN) approaches are highly effective for fish species identification in challenging underwater video data.
• These methods significantly advance automated fish identification capabilities for marine biodiversity monitoring.
• The study highlights the potential of advanced machine learning techniques to overcome limitations in underwater ecological research.