I don't think anybody uses FPGAs in production for compute-intensive 
tasks. They're usually only used in the design stage. I don't really 
know about ASICs, large corporations might have tried to use them in 
private clouds but they're not widely available in public clouds.

CPUs and GPUs are definitely used, although I have never read about an 
image search company that uses GPUs in a public Cloud. That's probably 
because Amazon's GPU instances are expensive.

Google has tried to use quantum computers from D-Wave too, but I don't 
know if this has led them somewhere.

To conclude, I think most of the processing is still done on CPUs. 
That's definitely the case at Moodstocks anyway.

On a side note, features extraction is usually not the most 
computationally expensive step in an image search system.