Abstract: Bridge the semantic gap between low-level visual feature and high-level semantic concepts has been the subject of intensive investigation on large scale remote sensing image management for years in order to improve the accuracy of automatic image annotation. An ocean remote sensing image auto-annotation method based on DBNMI model was proposed for contributions of semantic similarity about different regions of ocean remote sensing images. Initial remote sensing images were adaptively segmented, ocean remote sensing images were divided into background and the object region by means of a coarse-grained method, the relationship between low-level visual feature and high-level semantics label of the object region was modeled automatically, using DBN model, and the co-occurrence relations and adversarial relations between semantic concepts for improving image annotation results were calculated. The proposed approach is evaluated on a public remote sensing image dataset. The experimental results show a satisfactory improvement on accuracy.

Key words: deep belief networks, adaptive segmentation, ocean remote sensing image, image annotation