Visual saliency based maritime target detection

We present an unsupervised maritime target detection method based on visual saliency. From the perspective of maritime target detection as a salient object detection problem, we first extract visual features using momentum contrast (MoCo) learning and saliency cues from the visual features using the low-rank and sparse matrix factorization (LSMF) algorithm. Then, a saliency map is constructed based on the saliency cues. Finally, maritime target detection results are obtained by binarizing the saliency map. The flowchart of the proposed maritime target detection method is illustrated in Fig. 1. There are three major advantages of the proposed maritime target detection method. First, the method does not depend on the large-scale hand-annotated image set for training. Second, the method has a good generalization ability since convolutional neural networks are used to extract high-level semantic features from images. Third, the proposed method does not depend on any top-down priors. The effectiveness of the proposed maritime target detection method is evaluated on a set of marine images with various baseline methods compared with our method. The experimental results indicate that our method outperforms the baseline methods in terms of a set of quantitative metrics.