Neural network-based underwater image classification for Autonomous Underwater Vehicles

Abstract Image processing has been one of hot issues for real world robot applications such as navigation and visual servoing. In case of underwater robot application, however, conventional optical camera-based images have many limitations for real application due to visibility in turbid water, image saturation under underwater light in the deep water, and short visible range in the water. Thus, most of underwater image applications use high frequency sonar to get precise acoustic image. There have been some approaches to apply optical image processing methods to acoustic image, but performance is still not good enough for automatic classification/recognition. In this paper, a neural network-based image processing algorithm is proposed for acoustic image classification. Especially, shadow of an acoustic object is mainly used as a cue of the classification. The neural network classifies a pre-taught image from noisy and/or occlude object images. In order to get fast learning and retrieving, a Bidirectional Associative Memory (BAM) is used. It is remarked that the BAM doesn't need many learning trials, but just simple multiplication of two vectors for generating a correlation matrix. However, because of the simple calculation, it is not guaranteed to learn and recall all data set. Thus, it is needed to modify the BAM for improving its performance. In this paper, complement data set and weighted learning factor are used to improve the BAM performance. The test results show that the proposed method successfully classified 4 pre-taught object images from various underwater object images with up to 50% of B/W noise.