Create computers that "learn" to organize photos with semantic criteria
Americans have developed a computer system that teaches computers to archive images with semantic criteria. Using a vocabulary of more than 300 words in English, associated with tens of thousands of images of reference, a computer can classify an image of the Internet and file with a description consistent in just 1.4 seconds. The successes are 98%. The system will archive and search for images with much greater precision and speed that current procedures and will have many applications, from the documentation of art collections up for satellite photos.
Researchers from Penn State University have managed to teach computers to interpret images using a vocabulary of more than 330 words in English, so they can describe a picture with any number of terms and criteria with archiving software.
According to a release from the university, this new system can identify collections of photos online when they are archived, so it is a huge time saver for the millions of Internet users to hand-labeled images that have a bearing on their computers. Also, this system can facilitate the search for photographs with the use of keywords.
The system was developed by James Wang, associate professor of the College of Information Sciences and Technology of Penn State University, which is dedicated to the study of the recovery semantic-sensitive imaging, security of graphics files and systems development cataloging photographs with indices language learned, among other research.
The system, described in the article Real-Time Computerized Annotation of Pictures by Wang and her partner Li Jia, the department of statistics of the same university, aims to solve the problem of automatic cataloging of digital images.
According to both authors, develop the capacity of computers to do this work can have multiple applications, such as image search on the Web at the sites of photo albums emulating a search engine, and even to facilitate the conduct of experiments scientists. It could also serve to catalog collections of art, satellite images or pictures of specific diseases.
Advanced statistical models
With the systematization of advanced statistical modeling and optimization techniques, the researchers have managed to teach computers to hundreds of semantic concepts from images that exemplify each of these concepts.
So the system has emerged alipro (Automatic Linguistic Indexing of Pictures-Real Time), fully automatic. Quickly, the system can identify images on-line in real time. The evidence that has been made in the image on Internet sites have proven to catalog with an optimum precision.
Most browsers are based on current labels to describe textual images, but not all sets of text are recorded. As a result, the images that are not documented can not be found by web browsers, while many textual descriptions are confusing respect to the image file. This tool can automatically change at present and many more images with much greater accuracy with respect to search criteria.
Analyzing the pixels
Alipro works by analyzing the pixels (dots of a graphic image) of the images and comparing them with data that is stored in a computer database belonging to tens of thousands of reference images. The computer suggests that file from 15 possible words to catalog the new image.
Thanks to the introduction of these tens of thousands of images, computers have learned to recognize certain objects and concepts, which automatically attach to the new images that have never "seen".
Scientists have succeeded in this way that, in 98% of the images analyzed, the system generates at least one correct annotation of the 15 selected words. These entries makes the system by an average of 1.4 seconds.
The system presents, however, difficulties in the recognition of blurred photos or low contrast or low resolution, where the objects are only partially or when the angle of the photograph shows an object in the file from your computer, but from any other angle. To address these problems, the researchers introduced increasingly images in the database, which will expand the possibilities of recognition.
Future Work
Bearing in mind that the images are the primary means of expression on the Internet, should therefore ensure that the search simple and effective throughout the amount of graphic information that is on the network, which is growing continuously. The automatic real-time and is therefore becoming increasingly necessary, say the researchers.
The next steps for improving the system and its accuracy alipro include the incorporation of three-dimensional information in the processes of learning, which will improve the functioning of the system. It also will improve and increase the amount of archival images that provide the semantic concepts that serve to recognize new images.
With regard to applications, the system will begin to alipro tested in certain domains, such as biomedicine. It could also be integrated into other systems for data recovery to improve their own efficiency. The development of alipro has been funded by the National Science Foundation.
You liked this article? Can leave a comment and continue the conversation, or you can subscribe to the feed and get articles like this automatically to your aggregator of content.
Comments
No comments yet.
Post a comment