An adversarial approach in detecting inappropriate text content in images. An expression from a listing of expressions may be selected. The listing of expressions may include words, phrases, or other textual content indicative of a particular type of message. Using the selected expression as a reference, the image is searched for a section that could be similar to the selected expression. The similarity between the selected expression and the section of the image may be in terms of shape. The section may be scored against the selected expression to determine how well the selected expression matches the section. The score may be used to determine whether or not the selected expression is present in the image.