Two-stage high resolution remote sensing image retrieval combining semantic and visual features

Abstract

In this work, we put forward a two-stage image retrieval methodology by integrating high level image semantic features and low level visual features. At the first stage, we segment an image into parcels using a multiresolution remotely sensed image segmentation method combining rainfalling watershed algorithm and fast region merging. We then classify these parcels with Support Vector Machine (SVM), a famous non-linear classification scheme to connect the low-level visual features with high-level semantic features. These classes are then stored in semantic features databases for future use. When users carry out their rough semantic retrieval, they should choose and combine these semantic classes, and our method returns some image blocks which include the interested classes as the first rough retrieval results. At the second stage users should select an example from the results. We then construct and compare the similarity between the color and texture histograms for both the query example and each one in the semantic retrieval result. If the total similarity is higher than some threshold, the image will be returned as a suitable retrieval result. These images are sorted according their similarity as the final retrieval results. Experiments indicate our approach can get more effective and accurate results than content-based image retrieval only using visual features.