RESEARCH ON A RANDOM FOREST-BASED UNDERWATER TARGET IMAGE RECOGNITION METHOD

Min Chen, Da Song, Weibo Liu, Xiaomei Xie

ABSTRACT: Aimed to solve the problem of the low recognition rate of underwater images in bad underwater environment, an underwater target images’ gray level pixel matrix has been employed as the feature matrices to train random forest classifier and get the final classification result from the votes of multi decision trees in this paper. The Gaussian noise with different variances are added into the underwater acoustic image to simulate the different levels of complexity in underwater environment, by which the recognition rate and false alarm rate of diver, fish, propeller, bubble and pipeline are obtained. Furthermore, the performance of this method by ROC curve and AUC are analyzed, and the simulation results show that if the SNR (Signal-to-noise ratio) of underwater target is not less than 2.62dB or the PSNR (Peak signal-to-noise ratio) is not less than 12.7dB, the method proposed in this paper can detect underwater target with weak signal, and achieve more than 90% of the target recognition rate.

Keywords: Underwater target, image recognition, random forest, ROC curve