An Efficient Recognition Technique for Mine-like Objects Using Nearest-neighbor Classification

Broadband active sonars (15-150 kHz) capture morphological characteristics of underwater objects and are used by dolphins to recognize targets in cluttered environments, motivating their use in underwater mine countermeasure applications. However, the data from broadband sonars is very high-dimensional (typically 1400), requiring classification algorithms that can operate in these spaces with limited training data. Standard statistical approaches such as probability density estimation are often ill-suited to this task. This paper presents a new algorithm for mine-like object recognition that has shown promising results during in-water tests. The technique employs a nearest-neighbor classifier in conjunction with a non-metric similarity function and synthetic augmentation of the training data. Experimental results are presented comparing this method to a standard algorithm (LDA/PCA) and indicating that the nearest-neighbor approach addresses some deficiencies in existing techniques for the mine-like object recognition problem.