The Application Research On Maximally Stable Extremal Region In The Field Of Image Retrieval

With the development of internet, lots of images and videos are being produced quickly. How to retrieve the information we need has become a research focus in the field of image processing. Traditional image retrieval technology can't satisfy the current requirement any more, because it needs a great deal of manual work to label the data and has a low precision. Therefore, Content-Based Image Retrieval (CBIR) has become more and more important to retrieve and recognize the information.The key point of CBIR is how to extract and describe features from an image. This paper chooses Maximally Stable Extremal Region (MSER) to segment and extract image content, after comparing some affine invariant regions. The paper summarizes the definition and properties of MSER, and then presents how to detect the region using an effective algorithm in quasi-linear time.This paper proposes an image retrieval method based on object region using a text retrieval approach. First, it extracts MSERs from the images in the database, and then uses local entropy to filtrate the redundant regions which have little information. After using SIFT to describe the region, an improved k-means algorithm has been used to quantize the descriptors. Visual vocabulary is formed by the cluster centers. The paper regards an image as a document which contains some visual keywords, and then realizes an image retrieval method based on global features using inverted file systems and document rankings. Besides, the paper demonstrates an image retrieval technology based on object region using an effective voting function. To further improve query performance, the paper adds an efficient spatial verification stage to re-rank the original results, which greatly enhances the retrieval precision.In addition, the paper introduces a novel video copy detection technology using MSER, which regards video as a series of consistent key frames and uses image retrieval methods to realize video search. It uses gray histogram to detect shot boundary and extract key frames. After matching the key frames, it uses an effective temporal voting function to organize the correct candidate frames and delete the wrong ones. After sorting by the similarity score, the video copies can be detected. This paper has built an image retrieval system and a video copy detection system. After a series of experiments using TREVID data, it demonstrates that MSER is an excellent affine invariant detector, and the retrieval system using MSER can get a high recall and precision score.