A fish image segmentation methodology in aquaculture environment based on multi-feature fusion model.

Area of Science:

• Computer Vision
• Marine Biology
• Artificial Intelligence

Background:

• Underwater fish image processing is crucial for intelligent aquaculture but faces challenges like color cast, low contrast, and blur.
• Existing segmentation methods lack adaptive models, leading to suboptimal accuracy in complex marine environments.

Purpose of the Study:

• To develop an advanced convolutional neural network (CNN) model for accurate underwater fish image segmentation.
• To address limitations in current methods by proposing adaptive preprocessing and feature extraction modules.

Main Methods:

• A novel fish image preprocessing technique using pixel thresholding and minimum Euclidean distance for feature enhancement.
• Introduction of a multiscale attentional feature extraction module (MAFEM) integrating adaptive channel attention and dilated convolutional pyramid pooling.
• Training and validation of the model using a custom VOC-format dataset of underwater fish images.

Main Results:

• The proposed model achieved a mean intersection over union (MIoU) of 92.6% on the underwater fish image dataset.
• Demonstrated significant improvements over traditional models, with average increases of 1.84% in MIoU, 0.785% in mean pixel accuracy (MPA), and 1.18% in F1-score.
• The MAFEM module effectively strengthens the extraction of high-level semantic features.

Conclusions:

• The developed CNN model offers superior segmentation performance for underwater fish images compared to existing methods.
• Provides a robust foundation for intelligent monitoring systems, including fish body length measurement, weight estimation, and health status assessment.