EXPERIMENTAL RESULTS
The test images cover both generic images whose high-frequency bands are concentrated in luminance channel and full-color images which present significant edges in all color channels.
In all the experiments, the magnification factor is set as 3 and the size of the quaternion dictionary of size 256 is used. We use the training image set in the Yang’s package which is available at http://www.ifp.illinois.edu/˜jyang29/. Some commonly-used parameters of our method in the experiments are listed as follows. We choose patch size as 5 × 5 considering the ef-fectiveness and the computation efficiency. Empirically, we choose the error tolerance ε of QOMP in the training phase as 0.05. We prune the training samples from 150,000 to about 100,000 by re-moving those samples with relatively small variance. The iteration number T in Algorithm 1 is set as 100. The sparsity parameter K is set as 15 in algorithm 2 for QOMP during the SR phase.
We compare our results with the typical interpolation method of bicubic and two typical SR methods using image sparse model,