Denoising and Segmentation of SONAR Images for Rescue Operations

Abstract:
Floods cause environmental damage and pose challenges for post-flood rescue operations due to sand particles in the water. Forward-looking sonar is used to detect obstacles; however, it can suffer from speckle noise, low object-to-background contrast, and low signal-to-noise ratio. This paper proposes an image denoising and segmentation methodology that improves the detection rate for subsequent object detectors. The denoising methodology is compared to several state-of-the-art algorithms and has outperformed others in four baseline benchmarks, such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Mean-Squared Error (MSE) and Multiscale Structural Similarity Index (MSSIM). The segmentation uses two-level thresholding and morphological operations, and the sonar detection accuracies are computed on multiple obstacles. The proposed approach has shown higher accuracy in terms of true positives and false positives.