A Feature Learning and Object Recognition Framework for Underwater Fish Images.

Area of Science:

• Marine Biology
• Computer Vision
• Artificial Intelligence

Background:

• Accurate live fish recognition is vital for fisheries surveys, but underwater image analysis faces challenges like poor quality, uncontrolled environments, and limited data.
• Existing feature extraction methods often require human supervision, hindering automation in fisheries applications.

Purpose of the Study:

• To develop a fully unsupervised framework for underwater fish recognition.
• To address challenges in automated fish identification in fisheries survey applications.

Main Methods:

• Proposed a framework with unsupervised feature learning and an error-resilient classifier.
• Utilized saliency and relaxation labeling for object part initialization and learned a non-rigid part model.
• Implemented an unsupervised clustering approach for a binary class hierarchy and introduced partial classification for ambiguous images.

Main Results:

• Achieved high accuracy in underwater fish recognition on public and self-collected datasets.
• Demonstrated effectiveness in handling high uncertainty and class imbalance in underwater images.
• The proposed framework successfully automates feature learning and classification.

Conclusions:

• The developed unsupervised framework provides an effective solution for automated underwater fish recognition.
• The approach overcomes significant challenges in fisheries survey image analysis, improving data acquisition and processing.