An unsupervised approach to automatic object extraction from a maritime video scene

In this paper, an efficient unsupervised approach for extracting objects from maritime background using solely still video images is proposed. Its main idea is that maritime background (sea) has the main particularity of absorbing only hot light frequencies (red and green), while an object has not this property. Therefore if a timely vector of class features is considered, then two distinct statistical classes can be easily obtained in one maritime image, using an appropriate unsupervised two-class classification algorithm, and thus one can efficiently extract object from background. This is achieved by firstly partitioning a maritime image into equally-sized blocks, and constructing a 13-dimensional block-wise feature vector (regardless of block size) based on some observed statistics of light absorption in a block. The latter capture at the same time the characteristics (higher-order geometric moments) of the Fourier spectrum of RGB intensities in an image block, and its Shannon entropies. Both types of RGB information allow together generally to characterize well background versus object regions. Furthermore, a mere statistical test makes it possible to detect situations where there is no object in a maritime video scene but only sea. We have tested the proposed approach using several maritime scene videos and it has shown good performances in terms of detection accuracy, robustness and computational time.