Image annotation by search

Abstract

Digital content have been gaining more attention in recent years, and digital images are becoming more popular as a means of personal recount of people’s life. To deal with the sudden influx of digital images, there is a need for automatic image annotation to automatically categorize the images via their semantic content. This will allow people to maintain their digital images library easily, providing a user friendly platform for storage and sharing of personal memories. In this report, a 4-phrase methodology for image annotation was used. Each of the 4 phrases can be easily replaced or modified without affecting other phrases, much like the modular approach in software engineering design. Each of the phrases will target a specific area of the image annotation process.To achieve the best results, experimentations were done using the various extraction techniques in combination with the classification techniques. This allowed us to discover the technique or combination of techniques that will obtain the highest accuracy in terms of image annotation. The results yielded from the experimentations indicated that the best combination of techniques is the use of Dense SIFT (Level 1) with Bag of keypoints and classified via Linear Support Vector Machine (SVM). The relatively faster speed of this combination as compared with other technique combinations enable this technique combination to have the best of both worlds, namely accuracy and processing speed.