Research And Implementation Of Sketch-based Image Retrieval

With the explosion of increasing amount of multimedia information in local databases and also on the web, image management and retrieval have become an important challenge. However, searching for a desired image is always a hard task and has attracted many new researches.Content-based image retrieval (CBIR) is one of the most important applications that are working on this issue. Its aim is to facilitate the search in image databases based on the image content rather than text retrieval techniques.Most of these modes require one or several available images or additional specific text descriptions. However, the requirement is sometimes not feasible, as there is no existed image to be set as the example or the text description cannot be accurate enough. In these cases, the query result usually does not meet the user's expectation.The appearance of query by sketch provides a possible approach for solving the problem above. It meets the user's requirement of describing the semantic features of their target images, and the system is more intuitive and easy to use. However, there is great difference between sketch and image. Image is an artifact, which has a similar appearance to some subject. Sketch is not a'real', but a brief description of the image in his/her mind, which leads to losses of the luminance information and many details.As to our knowledge, little has been done to get the spatial structure information of the query sketch. Most query by sketch systems only treat the user's sketch as an image example for searching, but have not pay attention to its own spatial information distribution. In this paper, we develop a new approach for computing image spatial weight distribution. The method works directly with the edge pixels in the image sectors defined by angular radial partitioning.On the bases of a great amount of test data and statistical analyses, this paper discusses the characteristic of sketch's edge points distribution and estimates the spatial weight map of user-drawn sketch. From the experiment results, we can see the weight map of user's drawn sketch works well.In this paper, we apply one of popular color, texture, and shape features to represent both the user's sketches and images in dataset. Then try to retrieve the target image by using each of the features and analyze experimental results. At last, to prove the effectiveness of our proposed approach, we realize the search by sketch system, under the platform of the open source software'Digikam'. Digikam is an image organizer and editor using KDE Platform, and we rework the image search part.