Photo-to-search: Using multimodal queries to search the web from mobile devices

Nowadays, mobile phones with the digital camera are getting more and more popular. With necessary technologies, they are possible to become a powerful tool to search the Web on the go. Most Web search engines only support text queries. Therefore, users have to convert their information needs into words. However, it is sometimes difficult to describe the needs in text and the text input is inconvenient on small devices. To solve the problem, we propose a system named Photo-to-Search which allows users to input multimodal queries. Particularly, we study queries with captured images and optional text messages in this paper. For example, the user can simply take a photo of the flower and input a few terms like "flower". Textually relevant Web images are retrieved according to the query terms. Afterwards, the snapped picture is compared with these images by the CBIR (Content Based Image Retrieval) method. According to the context of the visually similar images, related key phrases are extracted. Finally, the search results are returned in multiple forms. Our system can also search for very similar images on the Web, such as movie posters or photos of film stars, to find related information. Experimental results on the large scale data showed our system achieved satisfactory efficiency and performance. Copyright © 2005 ACM.