Abstract: Objective retrieval plays an important role in daily life, however the model "Bag of words" exists some problems for image retrieval. For example, the quantified visual words reduce the discriminative power of the local features and the model does not capture the spatial relationship among local features, thus affecting the retrieval performance. In order to realize objective retrieval in complex environment, a feature saliency constraint based image retrieval method is proposed. By locating the salient object, this method can enhance the discriminative power of the local features and capture the spatial relationship among local features. First, the salient regions are distilled and the features inside these regions are marked as salient features. Second, the image retrieval is performed with spatial pyramid model using the salient features. Finally, the distance ratio of any two pairs of the features in each image is computed using the pair of salient features of the query image and retrieval image; then, the ratios which meet the threshold are recorded and their average is calculated to obtain the final results. Experimental results show that the proposed method improves the accuracy of the retrieval significantly and reduces the computing time during the geometric verification.

Key words: image retrieval, Bag of Words model, spatial pyramid, geometric verification, saliency