Your sample guy doesn't look like he enjoys following instructions...
Don't use a random sample, use a dedicated picture with annotations (e.g use a generally bright photo, and circle the important elements in red, and label them):
- your username
Use a few smaller ones to show common mistakes, e.g. like that:
Rationale: People don't follow instructions, especially when they are... on the hunt. Use less words, more pictures. Positive feedback through a friendly sample image, with clearly labeled requirements. Negative feedback for unwanted behavior. In combination, the user can "guess" an acceptable amount of readability, etc.
I think seeing a random sample image at that point has some benefit - e.g. making the page more personal. You might even rotate / regulary change the sample picture, and use pictures from existign users that agreed. (Idea: "Award" this status to paying members)
You might try some server side OCR and face recognition. It would be an interesting project for sure, but I doubt it's easy to setup or maintain against breaking attempts. However, immediate automatic rejection would drive home the point that you take this verification serious.