See also what's "In the Labs" !
Technology breaktrough
Understanding visual content has so far been a capability exclusive to the human eye and brain: pixels never made any sense to computers. Today, LTU Technologies has developed and deployed effective solutions that enable computers to see, understand, and natively handle any organization’s images
LTU’s core technology is the result of years of internationally recognized research performed by LTU's founders at MIT, Oxford and INRIA. It is now the exclusive property of LTU Technologies. LTU's technology is patented.
The LTU core technology engine is integrated as a modular image recognition and retrieval platform. This platform provides a highly scalable and redundant architecture: it is deployed within organizations that handle several million visual assets, and require a very high level of systems and data security. In addition, the platform is designed to be adapted to suit the specificities of your organizations workflow, environment or visual assets.
The LTU Engine
LTU’s core technology, the “LTU Engine”, is an image analysis system that indexes, recognizes, and describes images according to their visual content.
The LTU Engine’s input can be any type of visual data: personal or professional digital photographs, web crawled images, graphic designer images, scanned documents, trademarks, videos, vector images… The engine is able to analyze the arrays of pixels contained in these visual data, and produce a description of their graphical content: the image DNA.
Duplicate, cloned and similar images
LTU’s core technology allows to distinguish between duplicate, cloned and similar images.
LTU DNAs are robust towards a very large number of "cloning" graphical transformations. A test of 200 has been tested so far. Basically, whatever you can do with Photoshop.
Classification, clustering, categorization, keywording
Technology details
In a nutshell
LTU did nothing more than copying the human visual system. Indeed, the human retina is made of millions of light sensitive cells, that produce what could be considered as “pixels”. This information is not directly sent to the brain for analysis or recognition: a complex network of synapses transforms the millions of cell information into a reduced set of data that is transported over the visual nerve. And the number of connections in the visual nerve is several thousands: just like the dimension of LTU’s DNAs!
In the LTU core technology, the retina and the visual nerves would be the DNA computation module, and the brain would be the retrieval and recognition modules.
From Pixels to Content DNA
In the image analysis process, the first step is image segmentation: LTU's technology breaks down an image into relevant, visually-stable segments (Fig. 2), using a nonparametric, multiscale approach.
The second step is image indexing: the system assigns a unique identifier (or index) to the segmented image, called the signature (or "content DNA"). The content DNA is an optimized combination of unique visual features such as color, texture, shape, spatial configuration. The DNA also has extended invariance properties specific to image quality, image size, image brightness, contrast, distortion, object translation, object rotation and scale.
At the end of this process, the image is represented by a compact numerical vector (the content DNA) which efficiently encodes all the details of its content. In its dual representation, the image may be viewed as a point in a high-dimensional feature space. Note that the feature space has been extensively tested and optimized in order to maximize the discriminance of the description process.
Fig. 2: Breaking a complex image into its visually-relevant regions (a procedure known as image segmentation) is a prerequisite before indexing its content.
From Content DNA to Semantic Description
In the description process, the content DNA will feed several expert modules in order to be recognized with respect to its knowledge base (Fig. 3). The knowledge base may either be an internal LTU database enabling absolute content description, or an external database to which the input query has to be linked, enabling relative content description. The experts use state-of-the-art pattern recognition techniques, namely Neural Networks, Radial Basis Functions, Bayesian Estimation, and Support Vector Machines. Note that the recognition procedure has been designed for statistically behaving like human subjects, on whom extensive tests have been conducted by LTU.
The analyzer outperforms existing image classification systems because of its flexibility and learning abilities. The analyzer learns object profiles, refines its sense of what an object "looks like" and, therefore, continuously enriches its internal knowledge base. LTU's technology is also able to learn from user actions in an interactive context. Last but not least, the whole process, from image indexing to semantic content description is completed in real-time.
Fig. 3: Inferring semantic description from the Content DNA is a complex, high-dimensional pattern recognition problem in feature space.
