Many attempts have been made to identify the region of interest in an image. In this paper, we have provided a new approach for ROI detection using the output of image annotation. Our claim is that because ROI is a subjective concept, a method should be used to diagnosis human mental models and for this purpose, we have used KNN base annotation in our method. Because many people in pictures that are similar to test image, have chosen a particular labels, it is our belief that the content is probably more focused and so by recognizing areas of the test image that create this content, ROI can be determined with a good approximation.