Depending on the image recognition system it's easy to evade detection. In neural nets trained to recognize a specific image, it s possible to make an image that will destroy the networks ability to reconize the selected image/text.
Adversarial Examples is data in which a small amount of noise that is artificially created to misrecognize a sample.The example used for the image is shown below.
https://medium.com/analytics-vidhya/introduction-of-adversarial-examples-improve-image-recognition-imagenet-sota-method-using-1fe981b303e#: