![]() | Only 14 pages are availabe for public view |
Abstract Recently, large collections of digital images are created and grown rapidly. Digital images are produced each day by individuals and companies in many areas. Because of this rapid growth and huge production of digital images, there is an urgent demand to facilitate the searching of images. Image retrieval is a technique of searching for images in an image dataset. Currently, image retrieval is considered as an area of extensive active research in the computer vision field, especially in content based image retrieval (CBIR). It retrieves similar images of image query from large image dataset based on the content (features) of the image query.In the CBIR research community, image features can be either global features or local features. The global features describe the visual content of the entire image by single vector. They have speed in extraction of features and computing similarity between images. However, they still very rigid to represent an image (i.e., they ignore the image details). Consequently, they fail to detect significant visual features. In contrast, local features are based only on extract local surfaces around specific keypoints. They succeed greatly to detect important visual features. Accordingly, they are better in dealing with rotation, scaling changes, and illumination changes than the global features. Thus, local feature descriptors provide better retrieval effectiveness than global features descriptors. However, the number of local descriptors that extracted from the images may be huge, especially in the large image datasets.In this thesis, we proposed a system for CBIR that uses local feature descriptors with Bag of Visual Word (BoVW) model and Support Vector Machine (SVM) algorithm to efficiently retrieve similar images from standard datasets. The system uses Scale Invariant Feature Transform (SIFT) or Speeded Up Robust Features (SURF) as local descriptors to produce image signatures. On the other hand, to solve the problem of the large number of local descriptors that are extract from images, BoVW model is used to quantize local descriptors into ”visual words”. The BoVW model uses K-Means as a clustering algorithm to build visual dictionary. SVM algorithm is used to accelerate building of BoVW descriptor of images in order to efficiently retrieve more relevant images.We conducted experiments to test the performance of our system and compare performance of the local descriptors (SIFT, SURF) with BoVW. The performance of the system is evaluated by calculating precision, recall, and time. The system was evaluated using two different standard datasets. Based on the experimental results, we can conclude that SIFT or SURF more suitable according to the type of the used dataset. The experimental results revealed that our system performs better in terms of precision with a very short query time on different standard datasets. |