Salient Region Based And Step By Step Clustering For Target Search

Content based object search is a hot issue in the recent years. It contains two aspects: object matching and object localization. In the object searching framework, feature bundling (image partition) and object localization are two important techniques that are different from content based similar image retrieval. After feature bundling, most of the methods of content based similar image retrieval can be used to match object. For the object localization, most of the object localization methods use the location of the maximum similar image patch of the query object as the localization result. However, the accuracy of object matching and the time consuming of object retrieval are always contradicted each other. Thus, it is necessary to develop new methods to effectively improve object matching accuracy and time cost for the object search. In this paper, we investigate the target search based on the VLAD (Vector of Locally Aggregated Descriptors) image representation model, which includes content based similar image retrieval and object search. Corresponding algorithms are proposed for codebook construction, image representation and target localization. The main contributions are as follows:(1) For the codebook construction, a method of dimensionality reduction based on the feature binarization and a two-step clustering algorithm are proposed. According to the binarization transform, the original SURF feature is transformed into an 8 dimensional feature vector, which can significantly reduce the computational cost and memory usage. Then the feature vectors obtained by the dimensionality reduction are randomly partitioned into multiple subsets. Cluster algorithm is performed respectively within the subsets and among the subsets to obtain the final clustering results.(2) For the image representation, a weighted feature early fusion based image representation method is proposed. In addition, a weighted salient region based VLAD representation method is proposed. Since the dimensionality reduction will cause information loss, feature early fusion method is proposed to compensate this information loss. Also, different weights are used for the salient regions and non-salient regions for the VLAD image representation model to improve the image retrieval accuracy.(3) For the target localization in the object search, a salient region based scalable overlapping partition method is proposed. Scalable overlapping partition partitions an image into 65 different patches with different sizes. Then image salient region detection is used to solve the problems that an object region is partitioned into several patches or an image patche may include more than one object.Experimental results show that the clustering speed of the two-step clustering algorithm is much faster than the one-step clustering algorithm based on the k-means clustering algorithm for the codebook construction, while the cluster accuracy can be ensured. For the image representation, weighted feature early fusion method can greatly improve the image retrieval accuracy according to the experimental results on the Holidays and UKB databases. On the other hand, by the weighted salient region based VLAD representation method, image retrieval accuracy can be further improved. For the target localization, experiments on the Oxford5K shows that salient region based scalable overlapping partition method can obtain more accuracy object localization results and improve the object search speed.