Online sketch-based image retrieval using keyshape mining of geometrical objects

• Main Author: Abulaali, Huda
• Format: Thesis
• Language: English
• Published: 2017
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.utm.my/id/eprint/84054/1/HudaAbdulaaliAbdulbaqiPFC2017.pdf; http://eprints.utm.my/id/eprint/84054/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:126075

Online image retrieval has become an active information-sharing due to the massive use of the Internet. The key challenging problems are the semantic gap between the low-level visual features and high-semantic perception and interpretation, due to understating complexity of images and the hand-drawn query input representation which is not a regular input in addition to the huge amount of web images. Besides, the state-of-art research is highly desired to combine multiple types of different feature representations to close the semantic gap. This study developed a new schema to retrieve images directly from the web repository. It comprises three major phases. Firstly a new online input representation based on pixel mining to detect sketch shape features and correlate them with the semantic sketch objects meaning was designed. Secondly, training process was developed to obtain common templates using Singular Value Decomposition (SVD) technique to detect common sketch template. The outcome of this step is a sketch of variety templates dictionary. Lastly, the retrieval phase matched and compared the sketch with image repository using metadata annotation to retrieve the most relevant images. The sequence of processes in this schema converts the drawn input sketch to a string form which contains the sketch object elements. Then, the string is matched with the templates dictionary to specify the sketch metadata name. This selected name will be sent to a web repository to match and retrieve the relevant images. A series of experiments was conducted to evaluate the performance of the schema against the state of the art found in literature using the same datasets comprising one million images from FlickerIm and 0.2 million images from ImageNet. There was a significant retrieval in all cases of 100% precision for the first five retrieved images whereas the state of the art only achieved 88.8%. The schema has addressed many low features obstacles to retrieve more accurate images such as imperfect sketches, rotation, transpose and scaling. The schema has solved all these problems by using a high level semantic to retrieve accurate images from large databases and the web.