An Improved Image Retrieval Method Based On Feature Fusion

Traditional content-based image retrieval method is used to extract feature vector based on color,texture or shape feature,and then getting the k-nearest vectors from the image database as candidates.Image retrieval method based on feature fusion is more accurate and more stable than single feature retrieval algorithm,this method is focusd by more and more researchers.How to improve the retrieval precision and avoid affecting the retrieval speed is a hotspot of science research.On the basis of further study of image retrieval,systematic research has been done around the image feature extraction and an improved feature fusion algorithm for image retrieval is proposed by this thesis.The performance of single feature extraction method for image retrieval is always inadequate when acrossing image datasets,an improved FCTH-Bo VW feature fusion algorithm is proposed in this thesis and k-reciprocal nearest neighbors based on graph is used to fuse FCTH global feature and Bo VW local feature.Traditional feature fusion method is mainly fused by liner weighting,but it can be easily affected by the weight value,and it is difficult to find an optimal weight.In most cases,we should set different weight in different image datasets,which is unapplicable.We use k-reciprocal nearest neighbors feature fusion method to get results by searching for the maximum dense sub-graph,which can avoid setting the optimal weight value.For the FCTH descriptor,an annual weight method is proposed.For the Bo VW,a soft assignment method is adopted to quantize the local feature descriptor.Then combining this new quantification method and VLAD feature to improve the retrieval speed.The experiment result shows that the improved fusion method gets a higher retrieval precision result,and can be adapted to varieties of image datasets.