Exploring Data Redundancy in Real-world Image Classification through Data Selection