The Research Of Image Retrieval Technology Based On Salient Region And Feature Fusion

With the rapid development of Internet technology, we have been in an era of informationexplosion. Image information is the most intuitive and the most abundant source ofinformation in our life, which prompts that how to manage and retrieval the huge amounts ofimages becomes a research hotspot in the field of computer vision. Content-based imageretrieval by using image visual characteristics to achieve the effective of image retrieval,which has been used at present widely, major scientific research institutions and Internetcompanies successively propose the service of images search by images. However, this paperstudies the two questions about content-based image retrieval. The first question is therehaven’t filtered out the redundant information of image in the feature extraction stage, butthen extract feature from the whole image features blindly. The redundancy of informationleads to some wrong matching, which affects the final retrieval results. The second question isthat using a single visual feature can only describe one aspect of the image information andrestrict the retrieval effect because that information contained in images is always rich andcolorful. For the two questions, this paper puts forward the corresponding solutions.Mechanism of visual attention (VAM) points out that the human visual system hasdifferent attention to different regions of the image. The region, which obtains more attentionfrom human eyes often contain important information of the image, is called salient region.The region not belong to salient region can be regarded as redundant information. Therefore,we can keep the salient region of image by extraction algorithm, and filter redundantinformation at the same time.Firstly, this paper proposes a new extraction algorithm, which based on the classic regionextraction algorithm, to extract the salient region. The new algorithm combines global andlocal saliency together. And by compared with classical salient region extraction algorithm,the experiment we made in this paper proves the new algorithm can extract the better salientregionSecondly, this paper proposes a strategy for feature fusion to solve the question that asingle visual characteristic can’t roundly describe the rich information of image. The main idea of the strategy is to improve the retrieval effect by integrating the advantage of a varietyof features. This paper summarizes some common strategies of feature fusion, includingweight method and multiple retrievals method. Although weight method is simple andconvenient for a variety of features fusion, but the weight distribution in different featuresusually require tedious manual distribution. And the multiple retrievals method can usuallymake the results in a rough positioning by the first retrieve, and get an accurate retrieval usingthe identification ability of other visual features. That is to say, multiple retrievals method hasstrong dependence on the first retrieval, therefore, when the effect of the first retrieval is notgood, then the multiple retrievals method will be discounted. In order to perform better infeature fusion, this paper proposes a new feature fusion method named post verify method.The idea of the post verify method is to make the color verify for the matching position ofSIFT feature, so as to make a rearrangement for the retrieval results. The experimental provesthat the posterior method is rational and practicable, and it has less artificial parameters toadjust. Therefore, it can improve the effect of retrieval effectively.Finally, this paper also implements a complete image retrieval system according to theextraction algorithm and feature fusion strategy proposed in this paper. Through largeexperiments results comparing, it is proved that the salient region and feature fusion canimprove the effect of image retrieval.