Underwater Object Detection via Structural Pruning of YOLOv7

Junjie Liu, Zhanying Li, Zijun Wang, Longhui Liu
Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, PMLR 278:289-297, 2025.

Abstract

This study addresses the challenge of deploying object detection models in resource-constrained underwater environments by optimizing YOLOv7 through pruning techniques.Underwater detection faces limitations due to low-light conditions, water turbidity, and mobile device constraints.The proposed method applies channel pruning to YOLOv7, strategically removing low-weight channels to reduce computational load and parameter count while maintaining accuracy.Comparative experiments evaluated pruning rates (0%, 20%, 40%, 50%, 60%, 80%) on the UPRC dataset, focusing on sea urchins, scallops, sea cucumbers, and starfish.Results showed that a 50% pruning rate achieved optimal balance: mAP increased by 2.3% (from 83.8% to 85.7%), while parameters and computations reduced to one-fourth of original values.

Cite this Paper

BibTeX
@InProceedings{pmlr-v278-liu25b, title = {Underwater Object Detection via Structural Pruning of YOLOv7}, author = {Liu, Junjie and Li, Zhanying and Wang, Zijun and Liu, Longhui}, booktitle = {Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing}, pages = {289--297}, year = {2025}, editor = {Zeng, Nianyin and Pachori, Ram Bilas and Wang, Dongshu}, volume = {278}, series = {Proceedings of Machine Learning Research}, month = {25--27 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v278/main/assets/liu25b/liu25b.pdf}, url = {https://proceedings.mlr.press/v278/liu25b.html}, abstract = {This study addresses the challenge of deploying object detection models in resource-constrained underwater environments by optimizing YOLOv7 through pruning techniques.Underwater detection faces limitations due to low-light conditions, water turbidity, and mobile device constraints.The proposed method applies channel pruning to YOLOv7, strategically removing low-weight channels to reduce computational load and parameter count while maintaining accuracy.Comparative experiments evaluated pruning rates (0%, 20%, 40%, 50%, 60%, 80%) on the UPRC dataset, focusing on sea urchins, scallops, sea cucumbers, and starfish.Results showed that a 50% pruning rate achieved optimal balance: mAP increased by 2.3% (from 83.8% to 85.7%), while parameters and computations reduced to one-fourth of original values.} }
Endnote
%0 Conference Paper %T Underwater Object Detection via Structural Pruning of YOLOv7 %A Junjie Liu %A Zhanying Li %A Zijun Wang %A Longhui Liu %B Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2025 %E Nianyin Zeng %E Ram Bilas Pachori %E Dongshu Wang %F pmlr-v278-liu25b %I PMLR %P 289--297 %U https://proceedings.mlr.press/v278/liu25b.html %V 278 %X This study addresses the challenge of deploying object detection models in resource-constrained underwater environments by optimizing YOLOv7 through pruning techniques.Underwater detection faces limitations due to low-light conditions, water turbidity, and mobile device constraints.The proposed method applies channel pruning to YOLOv7, strategically removing low-weight channels to reduce computational load and parameter count while maintaining accuracy.Comparative experiments evaluated pruning rates (0%, 20%, 40%, 50%, 60%, 80%) on the UPRC dataset, focusing on sea urchins, scallops, sea cucumbers, and starfish.Results showed that a 50% pruning rate achieved optimal balance: mAP increased by 2.3% (from 83.8% to 85.7%), while parameters and computations reduced to one-fourth of original values.
APA
Liu, J., Li, Z., Wang, Z. & Liu, L.. (2025). Underwater Object Detection via Structural Pruning of YOLOv7. Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 278:289-297 Available from https://proceedings.mlr.press/v278/liu25b.html.

Related Material