Selective encryption reduces security by making it easier to find the data worth decoding. For example a check image would highlight a target of interest - even if part were encrypted. A clear text online order would provide valuable contextual information even if the credit card information were encrypted. With overall encryption, a lot of effort can be devoted to decoding what proves to be useless information.
Sounds like a good approach, but I wonder how much is saved (on average - I am sure it is media/image dependent) versus what additional power is used in the initial analysis? Looks like it works but getting major acceptance could prove a challenge.