ارتقاي تشخيص لبه ي اشياء تصاوير زير آبي در داده‌هاي مصنوعي با استفاده از مدل بهبود يافته GAN

Underwater computer vision is crucial in marine exploration, environmental studies, an‎d the use of autonomous underwater vehicles (AUVs). Nevertheless, the low contrast, color distortion an‎d blurred edges are common in underwater images because of the light absorption an‎d scattering, which negatively impact the object detection an‎d pose estimation. To overcome these challenges, this research suggests an integrated underwater vision system that is used to improve images, detect objects, an‎d estimate their pose. To improve underwater images an‎d retain fine details an‎d edges, a Multi-Level Attention Generative Adversarial Network (MuLA-GAN) is used. The advanced YOLO-based detectors are then used to process enhanced images to identify marine objects more accurately during the different visibility conditions. Lastly, the system uses 6-degree-of-freedom (6-DoF) pose estimation to give accurate localization an‎d orientation data of the AUVs. The proposed framework is eva‎luated on both synthetic an‎d real underwater datasets. The results demonstrate that the model significantly improves visual quality an‎d enhances object detection an‎d pose estimation performance compared to conventional methods. The framework integrates an image enhancement module with detection an‎d pose estimation components within a unified pipeline. Experimental results demonstrate that the proposed MULA-EdgeGAN model, trained on the EUVP dataset, achieves significant improvements when eva‎luated on the UIEB benchmark. The model achieves an average PSNR of 17.54 dB an‎d an average SSIM of 0.8036. Furthermore, the enhanced images show superior edge preservation, evidenced by an Edge F1-score of 0.7153 an‎d Sobel strength of 53.82, leading to improved performance in downstream object detection an‎d 6-DoF pose estimation tasks.

محمدرضا احمدزاده , محمد داورپناه جزي

الهام محمودزاده , مينا اميري