super resolution image
We used our generated images. Images contains a single line of text in various font family,font size, words, background colors, words colors, and position.
download dataset
(SRCNN)
(CNN)
(Skip pooling + Convolution and Deconvolution)
model | mse | mae | psnr |
do nothing | 1330.621 | 12.0985 | 17.926 |
SRCNN | 583.965 | 8.952 | 21.784 |
model2 | 334.135 | 6.053 | 24.239 |
model3 | 100.618 | 3.096 | 29.728 |
Table : All blur types in gray scale
The best model is model3 which has PSNR=29.728. This model outperforms because it contains more and complex layers than other model since we pass the information from lower layer to the top layer. Hence, the model3 can capture more information than other model.
Example of results (gray-scale image)
fig : left(blur) center(predict) right(original)
fig : left(blur) center(predict) right(original)