“A picture paints a thousand words”
but only if you know how to read it . .
Visual Recognition does for huge numbers of images what text analytics does for large collections of written text: it allows pictures to be processed automatically to find material of interest. This can provide the basis for open source intelligence or provide a means to process collected images to automatically find specified targets, such as military infrastructure.
The number of images from public sources, soldiers with (body) cameras, coalition partners and other sources is exploding. Clearly achieving timely, accurate classification of the images by “throwing people at the problem” does not scale and can be error prone.
Cognitive computing – in the form of unsupervised (machine) learning – will transform our ability to classify images.
Let’s consider recognising a military aircraft in an image using the Watson Visual Recognition (web) service. The service uses semantic classifiers built with machine-learning technology to recognize visual entities such as settings, objects, and events based on content such as color, texture, shape, and edges. Watson Visual Recognition allows us to build a classifier specific to our problem (“spot a military aircraft”) by learning from images we provide.
Figure 1 – Watson Visual Recognition Service (in IBM BlueMix)
To train the new classifier, we provide a set of positive (correct) and negative (incorrect) images as training sets, as shown in Figure 2. Careful examination of the negative images shows a combination of civil aircraft and birds.
Figure 2 – Training Sets: Positive & Negative
Once the classifier is trained, we can ask it to “recognize” images for us. Sample actual results are shown in Figure 3, but if the classifier misbehaves or the confidence is not high enough it can be re-trained with more complete and/ or comprehensive image libraries.
Figure 3 – Classification Results
For a real life problem, several classifiers would be built and applied to the same image, resulting in competing “image type” results with associated confidences. System accuracy could improve over time if the training sets are augmented with correct and incorrect classifications and the classifiers periodically re-trained.
Want to see this Watson Visual Recognition in action? We will be demonstrating it at or SPADE customer conference  in Berlin on the 12th & 13th April. Alternatively take a look at our BlueMix cloud integration platform.
References & Credits
- I’m grateful to my close colleague & friend Dr Doug Dykeman (@doug_dykeman) for showing me just how easy this can be!
- For insights into the underlying technology “Unsupervised One-Class Learning for Automatic Outlier Removal”
by John Smith et al
- If you are interested in attending our SPADE conference, please contact your local IBM account representative – or me if you don’t have one!