Fish Segmentation in Sonar Images by Mask R-CNN on Feature Maps of Conditional Random Fields.

Area of Science:

• Computer Vision
• Marine Biology
• Machine Learning

Background:

• Imaging sonar systems are crucial for monitoring fish behavior in challenging aquatic environments.
• Accurate fish segmentation in sonar images is essential for behavior analysis.
• Standard Mask R-CNN models struggle with variations in sonar image quality across different fish farms.

Purpose of the Study:

• To develop a robust fish segmentation method for sonar images.
• To enhance the transferability of trained Mask R-CNN models to new environments.
• To improve the accuracy and efficiency of fish behavior monitoring systems.

Main Methods:

• A novel preprocessing convolutional neural network (PreCNN) was developed for semantic segmentation.
• PreCNN integrates conditional random fields and utilizes successive sonar images.
• Semi-supervised learning was employed to leverage unlabeled sonar image data.

Main Results:

• Mask R-CNN performance significantly improved when applied to PreCNN's standardized feature maps.
• The combined PreCNN-Mask R-CNN approach demonstrated enhanced accuracy over direct Mask R-CNN application.
• Models trained on one fish farm were more effectively applied to new fish farms using the PreCNN preprocessing.

Conclusions:

• The proposed PreCNN effectively standardizes feature maps, decoupling environmental learning from instance learning.
• This approach enhances the generalization capability of Mask R-CNN for fish segmentation in diverse sonar imaging conditions.
• The PreCNN-Mask R-CNN system offers a more accurate and transferable solution for automated fish behavior analysis.