Discrimination Enhancement For Image Retrieval

With the rapid development of Internet and multimedia technology, the number of digital images has grown explosively. Although the large number of images bring much rich information and the convenient in daily life. It is difficult to retrieval the target images effectively from large scale image database. Under this background, image retrieval have developed rapidly in recent years and draw much attention from academia and commercial world. At present, traditional Text-based image retrieval technology is commonly used in search engines, while this method needs too much efforts to annotate textual labels for large scale database. As a result, this method will consume much time and the retrieval results are always produced subjectively. So it is hard to meet the retrieval requirements from large scale database. Therefore, content-based image retrieval turned out and became the mainstream in image retrieval field. It is proposed which aims to search the similar images based on the the inherent properties of the query image. This technology tries to describe the inherent properties of images from the similar perspective of human visual perception and then search the images which have similar properties with the source image. Now, most existing large scale image retrieval systems are enhanced by the Bag-of-Visual-words model and robust global image features, thus this method has the scalability suitable for large image database. However, the traditional model “Bag-of-Visual-words” does not capture the spatial relationship among local features. Meanwhile, it suffers from visual word ambiguity and quantization error, therefore the discrimination ability of the local features is reduced. These unavoidable disadvantages greatly affect the retrieval performance.To solve this problem, the research is beginning to geometric verification methods and global descriptor generation in image retrieval. According to the research of the spatial relationship with the saliency mechanism and the distribution entropy boosting aggregated descriptors, to enhace the discrimination of features in images. For the geometric verification stage, building the geometric encoding between query image and initial retrieval images, then verify the geometric consistency of matched features for optimizing the initial retrieval results. In this research, we focus on large scale image retrieval based on the geometric verification method, which is fused with the saliency mechanism and described by local region. For generating global descriptors stage, we use distribution entropy as a complementation information to residual for enhancing the searching accuracy. Summarized as follows:1. We propose Spatial Encoding Based On Hierarchical Salient Information, it mainly explores the geometric context of all visual words in images. It could increase the discriminative power of the features and reduces the computational time in geometric verification step based on hierarchical salient information. We represent the spatial layout of every three-point by angle encoding and location encoding. Meanwhile, in order to reduce the computing time, the hierarchical multi point encoding is applied, the matched features are classified based on the matched features are whether locate the hierarchical salient regions or not. Next, we generate the final ranking list by summing all spatial matching scores with weights based on hierarchical salient information. Experimental results prove that our scheme improves the retrieval accuracy significantly with low time consuming during the geometric verification step.2. We propose Hierarchical Geometry Verification via Maximum Entropy Saliency in Image Retrieval. This method aims at filtering the redundant matches and remaining the full information of retrieval target in images, meanwhile it fully explores spatial relationship of all visual words in images. First of all, hierarchical salient regions of query image based on maximum entropy principle are obtained and visual features with salient tags are labeled. Then we compute the saliency matching score by the tags which added to the feature descriptors, and regard the score as the weight information in geometry verification step. Second we define a spatial pattern as a triangle composed of three matched features by integrating angle encoding and location encoding for evaluating the similarity between every two spatial patterns based on the principle of similar traingles. Finally, we generate the final ranking list by summing all spatial matching scores with weight. In the experiment, we compare the hierarchical saliency’s effect in image retrieval schemes and hierarchical saliency’s effect in other geometric verification methods. Experiment results prove our method can not only improves the retrieval accuracy significantly but also reduce time consuming of the full retrieval.3. We propose Spatial Verification Method Based On Local Region Constraint, most geometric verification methods usually estimate whether all the matches of the entire image plane follow consistent geometric transformation, and in such way, some irrespective matches which are involved in post-processing step affect the improvement of the retrieval performance. So, we propose a Spatial Verification Method Based On Local Region Constraint method to remove false positive matches. First, we put forward the concepts of relevant features and irrelevant features, and then select a pair of matched features as reference matched features. Next, we compute the local region range of a center match pair. Further, we define the constraint value of the center matches according to the number of other matches in the local regions and judge whether the center match pairs are satisfy the condition of local region constraint or not. If the condition is met, we adopted the rotational-circle constraint method to estimate whether all the matches in the local regions follow consistent geometric transformation. We quarter the circles to measure the geometric similarity of local matches and generate the spatial encoding strictly. And further we rotate the segmented circle to control the strictness of geometric constraints. Finally we judge whether the reference matched features are correct or not, based on verifying the all spatial encoding are whether satisfied the geometric consistency or not. In the experiment, we compare the several typical geoemtrci verification methods. It demonstrates that our method could improve the retrieval accuracy significantly with low computational cost.4. We present a novel image presentation called Distribution Entropy Boosting aggregated descriptor, which extend the original vector of locally aggregated descriptors. Since aggregated descriptor only adopts residuals to depict the distribution information of SIFT descriptors around the cluster and neglect other statistical clues, the original aggregated descriptors might not distinctive enough. To address this issue, this paper proposes to use distribution entropy as a complementation information to residual for enhancing the searching accuracy. In order to fuse two descriptor sources organically, new fusion normalization stage meeting power-law is also investigated. Experiments on two-step image search benchmarks prove that our proposed method achieves the state-of-art performance.