**Table**

**1.** *Number of true positives, false positives and false negatives for the 1st and 2nd degree burns shown in the right upper panel in Figure*

 *6.*


**Table 2.**

*Sensitivity and precision for the classification and detection of skin burns.*

In addition to higher performance, another advantage of untrained dictionaries over trained dictionaries is low computational complexity, since the collection of observations and the training stage are not required. In order to explain why untrained dictionaries provide better detection performance, there are two observations: (1) Atoms are normalized feature vectors, and most features (average value, variance, contrast, angular second moment, inverse difference moment) are positive. (2) After training, atoms are adjusted and most of their entries are no longer positive. Untrained atoms use raw features without adjustments introduced with training. Untrained features lie on a constrained sub-space given that they are non-negative. Thus, they are more densely concentrated within the same sub-space where the signal to be reconstructed is located. This is why the error introduced by the sparse reconstruction of a signal is smaller over untrained atoms.
