| Summary: | 碩士 === 國立中央大學 === 資訊管理研究所 === 99 === With the advent of the Internet and an increase in web images, manual image annotation becomes a difficult task and more time-consuming than automatic image annotation. Most research proposed algorithms for matching the keywords and the images accurately. However, those methods annotated images in original resolution, and it might cost more time and storage. In addition, different feature representation approach can cause various performance of annotation .We aimed to annotate images with different resolution and different feature representation approach and discussed the effect of these two factors.
We chose Corel, PASCAL VOC2008 and Corel 5000 to be our experiment data sets, and selected Bicubic Interpolation to scale these data sets into 256x256 resolution, 128x128 resolution, 64x64 resolution, 32x32 resolution and 16x16 resolution. Furthermore, local feature representation and Bag-of-Words feature representation were used in our experiment. In annotation step, we used support vector machine and K nearest neighbor algorithms.
Finally, the experimental results indicated that the accuracy of annotation didn’t decrease but the time of annotation was reduced rapidly when the image resolution was diminished. Besides, we also compared two feature representation approaches, the performance of local feature representation was better than Bag-of-Words feature representation, especially in support vector machine. Meanwhile, in different resolution, the performance of Bag-of-Words feature representation was more stable than local feature representation.
|
|---|
