Weakly supervised underwater fish segmentation using affinity LCFCN.

Area of Science:

• Computer Vision
• Marine Biology
• Aquaculture Technology

Background:

• Accurate fish body measurements are crucial for marine and aquaculture productivity.
• Current manual and fully-supervised methods for fish measurement are time-consuming and labor-intensive.
• Per-pixel segmentation labels require up to 2 minutes per fish, hindering scalability.

Purpose of the Study:

• To develop an efficient fish segmentation model using point-level supervision.
• To reduce the time and effort required for annotating fish images for measurement.
• To improve the accuracy and efficiency of automated fish body measurement estimation.

Main Methods:

• Proposed Affinity-LCFCN (A-LCFCN), a fully convolutional neural network model.
• Employed point-level supervision, requiring only a single click per fish (average 1 second annotation time).
• Integrated per-pixel scores and affinity matrix outputs, refined using a random walk and trained with localization-based counting fully convolutional neural network (LCFCN) loss.

Main Results:

• A-LCFCN demonstrated superior performance compared to fully-supervised models under fixed annotation budgets.
• The model achieved better segmentation results than the standard LCFCN and other baseline methods.
• Point-level supervision significantly reduced annotation time, making fish measurement more efficient.

Conclusions:

• Affinity-LCFCN offers an efficient and effective solution for fish segmentation using minimal annotation effort.
• The proposed method holds significant potential for enhancing productivity in marine and aquaculture applications.
• Point-level supervision is a viable and advantageous alternative to traditional per-pixel labeling for fish measurement tasks.