Research On Sketch Based Image Retrieval

With the development of internet technology, the number of digital images is increasing sharply and content-based image retrieval technique has gained extensive attention from domestic and foreign scholars. In recent years, with the popularity of touch screen devices such as tablet computer and smart phone, researchers begin to focus on sketch-based image retrieval technique. With the help of touch screen devices, people of different ages and with various painting skills can easily draw a sketch that arises in their mind and then images with similar shape can be found from a large image dataset based on the lines of the sketch.The concept of sketch-based image retrieval is first brought up in1980s. However, due to the variability and uncertainty of the lines in hand-drawn sketches, researchers face a lot of difficulties and challenges in feature representation, matching and the index structure building procedures. Therefore, it makes slow progress since1980s. In2010, Microsoft Asia Research Institute proposed a real time sketch-based image retrieval system-MindFinder, which can realize image retrieval from a large dataset only by matching the lines in the hand-drawn sketch and the contours in the images from dataset without using any keywords. This technique once again triggered the research enthusiasm on sketch-based image retrieval.In this paper, we propose a sketch-based image retrieval system based on bag of words framework. Each hand-drawn sketch can be represented by a histogram that is related to the visual words. When extracting features, we take the adapted HoG feature-Gradient based HoG as our local descriptor, which can efficiently represent the sketch images. Hierarchical K-means clustering is used to build the visual vocabulary, which performs better than the traditional K-means clustering algorithm. Finally, the similarities between query sketch and images from the dataset are measured by cosine distance function, which can realize image retrieval from training dataset. We can also use multi-SVM classification to get the keywords of the query sketches, which are then sent to the search engines to realize online image retrieval. Experiments on the sketch dataset released by Eitz and the shape dataset provided by Microsoft Office show that our algorithm can achieve more satisfactory results than MindFinder and Eitz’s sketch retrieval algorithm.