Structural Noise Removal for Imaging Sonar: A Case Study for BlueView Imaging Sonar

In underwater acoustic vision based applications, a BlueView imaging sonar is advantageous for its long signal propagation distance and strong penetrating ability. However, the sonar images are generally suffering from structural noise due to interference and reverberation, which brings a target-like pattern and severely hinders the corresponding measurement and interpretation process. Most previous work assumes identical independent Gaussian noise model, which is potentially not well-suited to structure noise. Instead, we ultimately adopt pixel-wise adapted Gaussian mixture models to represent the noise distribution for each individual pixel based on the investigation of the noise characteristic. The parameters of each model are consequently learned from real image data via an expectation-maximization algorithm. Utilizing the noise models, the structure noise is primally suppressed through probabilistic thresholding and further removed via median filtering. Experimental results on both simulated and real data validate the effectiveness and efficiency of our method.