What question did this study set out to answer?

The study aims to develop a framework for accurately estimating fish length and pose from underwater images, addressing environmental challenges.

May 30, 2026Open Access

MorphoCal: a multi-stage deep learning framework for fish length estimation in challenging underwater pond environments

Q: What does this research mean for the field?

The MorphoCal deep learning framework, utilizing a coordinate-attention-enhanced YOLO network and single-shot calibration, achieves highly accurate fish pose estimation and centimetre-level length reconstruction in challenging underwater environments. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Key Points

The study aims to develop a framework for accurately estimating fish length and pose from underwater images, addressing environmental challenges.
Developed MorphoCal, a multi-stage deep learning framework combining YOLO-based detection and metric reconstruction.
Integrated AquaYOLO-PoseCA for pose estimation and utilized checkerboard calibration for metric accuracy.
Evaluated on the DePondFi’24 dataset under real-world pond conditions.
Achieved bounding box mAP of 0.959 and pose mAP of 0.848, indicating high detection accuracy.
Demonstrated mAP 50−95 of 0.712 while maintaining computational efficiency with 29.4 GFLOPs.
Reconstructed fish lengths showed low deviation from ruler-based measurements.

Abstract

Introduction Underwater imaging plays an important role in monitoring aquatic ecosystems and aquaculture environments. Accurate estimation of fish pose and body length from underwater imagery is essential for analysing fish behaviour, growth patterns, and biomass dynamics. However, underwater scenes are often affected by turbidity, uneven illumination, occlusion, and suspended particles, which reduce the accuracy of keypoint detection and metric measurements. Methods To address these challenges, this study proposes MorphoCal , a multi-stage deep learning framework for fish pose estimation and geometric length reconstruction under real-world pond conditions. The framework integrates AquaYOLO-PoseC A , a coordinateattention–enhanced YOLO-based keypoint detection network, with a single-shot checkerboard calibration and ray–plane projection module for centimetre-level metric reconstruction. Coordinate Attention is incorporated into the YOLO backbone to encode direction-aware spatial features and preserve positional information, improving anatomical keypoint localization for elongated fish structures. The architecture performs joint fish detection and keypoint estimation in a single-stage forward pass using PAN–FPN feature fusion. Results Experiments on the DePondFi’24 dataset, consisting of multi-species fish captured under natural pond conditions, show that AquaYOLO-PoseCA achieves a bounding box mAP of 0.959, pose mAP of 0.848, and mAP 50− 95 of 0.712, while maintaining computational efficiency with 29.4 GFLOPs and 11.4M parameters. The reconstructed fish lengths show low deviation from ruler-based ground truth measurements. Discussion The proposed MorphoCal framework enables reliable fish pose estimation and centimetre-level length reconstruction under challenging underwater conditions, supporting noninvasive fish monitoring, growth assessment, and biomass estimation in intelligent aquaculture systems.

Ask AI

Helpful

Bookmark

View Full Paper