Enhancing Underwater Image Segmentation Through Deep Learning-Based Image Enhancement Techniques: A Study on EUVP and SUIM Datasets

Main Article Content

Abhisheka Thumbesara Eshwara
Basavaraj Ningappa Jagadale
Madhuri Gurram Ramesh
Chandrakantha Tholalemane Satyanarayana

Abstract

Underwater image segmentation is critical for marine monitoring, archaeology and the navigation of autonomous underwater vehicles (AUVs). Yet, light scattering, low contrast and color distortion greatly reduce the quality of underwater images which affects the accuracy of segmentation. A pipeline is suggested, where two enhancement techniques, GANs and the CNN-based Ucolor method, are used before SAM performs the segmentation. The results are measured using PSNR and SSIM on a number of samples from the EUVP dataset. The SAM tool is used on segmentation images from the SUIM dataset in three versions: the raw image, the GAN-enhanced image and the Ucolor-enhanced image. The results of segmentation are checked numerically using mIoU, Dice Similarity Coefficient (F1 Score) and Pixel Accuracy and visually by superimposing the segmentation on the image. The results demonstrate that both GAN and Ucolor increase the accuracy of segmentation and Ucolor provides better improvement in quality. This shows that importance of preprocessing in underwater semantic segmentation. The method also proposes use of task-aware enhancement models and dynamic prompt tuning to improve real-time performance during AUV operations.

Article Details

Section
Articles