I am looking for a search engine, or only the technology, that can comprehend natural language semantics (to some level, of course) in order to find correspond images.
The other way around can work as well.
For example: a software that can distinguish between two men hugging and two men wrestling roman style. Or - input: a photo of a desk with a laptop, a cup of coffee, Rubik's cube and a star wars toy that would yield the output: programmer desk.
I came across Metamind.io for my work, which uses some deep Neural Network wizardry to recognize and label images the right way.
You might get good results from their trained classifier, or train your own.