Computer Vision

super resolution image

Members

  • Chayanont Eamwiwat 5830083121
  • Burin Naowarat 5831034631
  • Tharn Phongsavaree 5831029521

Our objectives

  • Enhance image to be more aesthetic and sharpness
  • Improve rocognization text in image
  • Can utilize image with higher performance

Dataset

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.

Dataset (con.)

  • training set (30k)
  • validation set (20k)
  • training set (20k)

download dataset

Model 1

(SRCNN)

Model 2

(CNN)

Model 3

(Skip pooling + Convolution and Deconvolution)

Evaluation

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

Conclusion

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)

Example of results (rgb image)

fig : left(blur) center(predict) right(original)

current limitations

  1. Currently, we are still working on generated dataset since we don't have real-world data.
  2. We have train and test on our environmnet as font family, font size, and backgroud were limited.
  3. There are other techniques and models that we haven't tried.
  4. Still dont't have the program for common user to use.