Fig 3. QDCP images generated under different thresholds TO with 8-point unit-radius circular neighborhood.
Given a normal breast biopsy image (a) from the UCSB breast cancer dataset [41], a small threshold results in a QDCP image with large black area as most color changes are considered insignificant for image description (b); whereas with a large threshold, QDCP is sensitive to small color variations, and the major color change may be overwhelmed by small color variations (d). Since there are 36 rotation-invariant patterns under the setting, all QDCP images are normalized with respect to 36 for visualization.