Yes, depending on the kind of data you want to work with you might want to explicitly get the blurry ones. That depends on the task of the ML model you train and requires domain expertise.