This work is an expansion of Regression_Blur. For requirements and installing anaconda environment click here
To train your network with your own data use VGG16_Regressor_V2.py. The model used throughout the results and blur field prediction is under models.py using VGG_14. Results_V2.py uses a test dataset and accepts multiple patch sizes as a parameter to test results on different image patch sizes.
For pre-trained weights contact the author at varelal@nmsu.edu.
Nub_Generator.ipynb has the function blur_field_gen which generates a blur field of size M, N. This size should be the same size of the image desired to for blur. The specified row and column will divide the blur field by row patches and column patches. The values array will sweep the row patches and assign the value in the array element. For a unique values of patches there should be row * col ammount of values within the array.
The non_uniform_blur function will accept the sharp image img, and two blur field maps one for length L and another for angle A. This function will do a per pixel blur.
Blur_Field_Prediction.py uses the trained model and loops through the image at a certain patch size to predict the length and angle blur parameters for the center pixel of the patch. The script does predict a per pixel blur field skipping a boarder size of PATCH_SIZE / 2.