Deep Learning Based Algorithm Selection Approach for Underwater Image Processing
Uloženo v:
| Název: | Deep Learning Based Algorithm Selection Approach for Underwater Image Processing |
|---|---|
| Autoři: | Džaja, Barbara |
| Informace o vydavateli: | 2025. |
| Rok vydání: | 2025 |
| Témata: | Deep Learning, Algorithm Selection, Underwater Image Processing, Image Enhancement, Quality Metrics |
| Popis: | Underwater images are frequently affected by distortions caused by light properties such as refraction, absorption, and scattering. Although numerous image processing algorithms have been developed to target specific types of distortion, selecting the most appropriate algorithm for a given image remains a challenging task. This work proposes a deep learning-based approach for automatic selection of the most suitable underwater image enhancement algorithm. Six algorithms (Automatic Red Channel, Fusion, Hazeline, Backscatter Removal, Local Color Mapping & Color Transfer, AutoMSRCR) are evaluated using six quality metrics (UIQM, BRISQUE, NIQE, NORM, PSNR, SSIM). The neural network is trained using the EfficientNet-B7 architecture with images from public datasets (UIEB, EUVP) and additional images captured with a Chasing Dory device at the Faculty of Science, University of Split. Results indicate that ARC and Fusion algorithms most frequently yield the best quality outputs, while Hazeline performs worst on average. A dataset imbalance led to overfitting during training, highlighting the need for dataset expansion. Despite this limitation, the proposed approach demonstrates promising potential for automated algorithm selection in underwater image processing. |
| Druh dokumentu: | Conference object |
| Přístupové číslo: | edsair.dris...01492..1b8f59a50ff0d7179f8972445756de5e |
| Databáze: | OpenAIRE |
Nájsť tento článok vo Web of Science | Abstrakt: | Underwater images are frequently affected by distortions caused by light properties such as refraction, absorption, and scattering. Although numerous image processing algorithms have been developed to target specific types of distortion, selecting the most appropriate algorithm for a given image remains a challenging task. This work proposes a deep learning-based approach for automatic selection of the most suitable underwater image enhancement algorithm. Six algorithms (Automatic Red Channel, Fusion, Hazeline, Backscatter Removal, Local Color Mapping & Color Transfer, AutoMSRCR) are evaluated using six quality metrics (UIQM, BRISQUE, NIQE, NORM, PSNR, SSIM). The neural network is trained using the EfficientNet-B7 architecture with images from public datasets (UIEB, EUVP) and additional images captured with a Chasing Dory device at the Faculty of Science, University of Split. Results indicate that ARC and Fusion algorithms most frequently yield the best quality outputs, while Hazeline performs worst on average. A dataset imbalance led to overfitting during training, highlighting the need for dataset expansion. Despite this limitation, the proposed approach demonstrates promising potential for automated algorithm selection in underwater image processing. |
|---|