There is problem.

In just 4 minutes, manually, I was able to use Windows Snipping tool on each picture, import those pictures into Google's own image search, and use a Word Map, to discover single most common noun for each photo, then could easily correlate with the most common noun of the subject image.

This is a bit slower than simply running a captcha image through an advanced OCR... but fairly easy to write a script to take cropped screenshots of each image and run through Google's image search and then do a word correlation.

Google could try to lock down their Image Search API, limiting the number of searches per second to catch spammers, but that is easy to get around. They could try removing these images from their index, but that is make it hard to use good recognizable images.

Any attempt to obfuscate the images will result in pain for the user to recognize on mobile devices, just like captchas are getting harder.

My suggestion, is to overlap, blend, and blur borders of the images together, and analyze user input (press) coordinates on the server to determine the closest, most likely choice of the user. You have to eliminate a clean "cropping" of the image, while preserving human recognizability.