Method and apparatus for processing blurred image

1. A method for processing a blurred image which generates a deblurred image similar to a ground truth by removing a blur from an input blurred image, comprising steps of:

generating a first input feature map and a second input feature map with a feature distribution for blur removal from the blurred image;

generating a prediction feature map from the first input feature map by using a self-spatial feature transform (SSFT) module which transforms the feature distribution for blur removal into a feature distribution for face recognition without external information; and

generating a deblurred image based on the second input feature map and the prediction feature map.

2. The method for processing the blurred image of claim 1 , wherein the prediction feature map is generated by a prior information generator learned through a generative adversarial network (GAN).

3. The method for processing the blurred image of claim 2 , wherein the GAN includes the prior information generator;

a face recognition module that generates a facial feature map including texture information of the face from the ground truth through the pre-learned face recognition module; and

an identifier that inputs the prediction feature map or the facial feature map and identifies whether the input is the prediction feature map or the facial feature map.

4. The method for processing the blurred image of claim 3 , wherein the identifier identifies the input through steps of:

giving different weights for each channel to the prediction feature map and the facial feature map;

processing an internal feature by concatenating the channels according to the weight; and

classifying whether the input is the prediction feature map or the facial feature map according to a result of processing the internal feature.

5. The method for processing the blurred image of claim 4 , wherein the prior information generator is learned based on a pixel loss calculated by calculating a pixel unit distance, an adversarial loss generated by competitive learning of the prior information generator and the identifier, and a prior loss calculated by calculating a distance according to the weight of the prediction feature map.

6. The method for processing the blurred image of claim 1 , wherein the generating of the deblurred image includes steps of:

applying the prediction feature map used as prior information to the second input feature map using a spatial feature transform (SFT) module; and

generating the deblurred image by transforming the feature distribution for face recognition into the feature distribution for blur removal with respect to the second input feature map applied with the prediction feature map.

7. An apparatus for processing a blurred image which generates a deblurred image similar to a ground truth by removing a blur from an input blurred image, comprising:

an encoder that generates a first input feature map and a second input feature map with a feature distribution for blur removal from the blurred image;

a prior information generator that generates a prediction feature map from the first input feature map by using a self-spatial feature transform (SSFT) module which transforms the feature distribution for blur removal into a feature distribution for face recognition without external information; and

a decoder that generates a deblurred image based on the second input feature map and the prediction feature map.

8. The apparatus for processing the blurred image of claim 7 , wherein the prior information generator is learned through a generative adversarial network (GAN).

9. The apparatus for processing the blurred image of claim 8 , wherein the GAN includes

the prior information generator;

a face recognition module that extracts a facial feature map including texture information of the face from the ground truth through the pre-learned face recognition module; and

an identifier that inputs the prediction feature map or the facial feature map and identifies whether the input is the prediction feature map or the facial feature map.

10. The apparatus for processing the blurred image of claim 9 , wherein the identifier includes

a channel concentration module that gives different weights for each channel to the prediction feature map and the facial feature map;

a processing module that processes an internal feature by concatenating the channels according to the weight; and

a classification module that classifies whether the input is the prediction feature map or the facial feature map based on an output of the processing module.

11. The apparatus for processing the blurred image of claim 10 , wherein the prior information generator is learned based on a pixel loss calculated by calculating a pixel unit distance, an adversarial loss generated by competitive learning of the prior information generator and the identifier, and a prior loss calculated by calculating a distance according to the weight of the prediction feature map.

12. The apparatus for processing the blurred image of claim 7 , wherein the decoder generates the deblurred image by applying the prediction feature map used as prior information to the second input feature map using a spatial feature transform (SFT) module and transforming the feature distribution for face recognition into the feature distribution for blur removal with respect to the second input feature map applied with the prediction feature map.