Inverse kernel-based defocus deblurring method and apparatus
The present disclosure provides a method of effectively deblurring a defocus blur in an input image based on an inverse kernel. The defocus deblurring method includes: generating, by an encoder network, an input feature map by encoding the input image; filtering, by an atrous convolution network including a plurality of atrous convolutional layers arranged in parallel, the input feature map to generate an output feature map having reduced blur component; and generating, by a decoder network, an output image having reduced blur from the output feature map with the reduced blur component and the input image.
1 . A method of reducing defocus blur from an input image, comprising:
generating, by a processor executing at least one program instruction, using an encoder network, an input feature map by encoding the input image;
predicting, by the processor, a reference shape of a plurality of blurs shared with a plurality of pixels of the input image, based on the input feature map;
predicting, by the processor, a size of each blur for a position of each pixel without a shape of each blur for the position of each pixel;
generating, by the processor, a plurality of inverse kernels for deblurring the input image, based on predicted sizes of the plurality of blurs and the reference shape;
filtering, by the processor, using an atrous convolution network including a plurality of atrous convolutional layers, constructed based on the plurality of inverse kernels, arranged in parallel, the input feature map to generate an output feature map having reduced blur component; and
generating, by the processor, using a decoder network, an output image having reduced blur component from the output feature map with the reduced blur component and the input image,
wherein each of the plurality of atrous convolution layers being configured to share common convolution kernel weights corresponding to an inverse blur kernel having common shape based on the reference shape with respect to a dilation rate and having a size related to the dilation rate.
2 . The method of claim 1 , wherein filtering the input feature map to generate the output feature map having reduced blur component comprises:
filtering the input feature map consecutively by a plurality of atrous convolution network connected in series to repeatedly reduce the blur component.
3 . The method of claim 1 , wherein the plurality of atrous convolutional layers comprises respective convolutional kernels containing kernel weights shared with the other atrous convolutional layers.
4 . The method of claim 3 , wherein the convolutional kernels of the plurality of atrous convolutional layers are dilated by dilation rates different from each other.
5 . The method of claim 4 , wherein filtering the input feature map to generate the output feature map having reduced blur component comprises:
performing convolutions between the input feature map and each of the convolutional kernels of the plurality of atrous convolutional layers in parallel to acquire a plurality of convolution results; and
nonlinearly aggregating the plurality of convolution results by concatenating the plurality of convolution results and performing a convolution between a concatenation result and a network weight.
6 . The method of claim 5 , wherein filtering the input feature map to generate the output feature map having reduced blur component further comprises:
determining scale attention maps each comprising pixel-wise weights for respective one of the plurality of convolution results,
wherein nonlinearly aggregating the plurality of convolution results comprises:
multiplying each of the plurality of convolution results by a corresponding scale attention map;
concatenating a plurality of multiplication results; and
performing the convolution between the concatenation result and the network weight.
7 . The method of claim 6 , wherein filtering the input feature map to generate the output feature map having reduced blur component further comprises:
determining a channel-wise weight vector for compensating for a variation of a shape of the defocus blur in the input image according to a pixel position,
wherein nonlinearly aggregating the plurality of convolution results comprises:
multiplying each of the plurality of convolution results by a corresponding scale attention map;
multiplying each of a plurality of multiplication results by the channel-wise weight vector;
concatenating a plurality of multiplication results multiplied by the channel-wise weight vector; and
performing the convolution between the concatenation result and the network weight.
8 . An apparatus for reducing defocus blur from an input image, comprising:
a memory storing program instructions; and
a processor coupled to the memory and executing the program instructions stored in the memory,
wherein the processor, is configured to:
generate an input feature map by encoding the input image;
predict a reference shape of a plurality of blurs shared with a plurality of pixels of the input image, based on the input feature map;
predict a size of each blur for a position of each pixel without a shape of each blur for the position of each pixel;
generate a plurality of inverse kernels for deblurring the input image, based on predicted sizes of the plurality of blurs and the reference shape;
filter, using an atrous convolution network comprising a plurality of atrous convolutional layers, constructed based on the plurality of inverse kernels, arranged in parallel, the input feature map to generate an output feature map having reduced blur component; and
generate an output image having reduced blur component from the output feature map with the reduced blur component and the input image, and
wherein each of the plurality of atrous convolution layers being configured to share common convolution kernel weights corresponding to an inverse blur kernel having common shape based on the reference shape with respect to a dilation rate and having a size related to the dilation rate.
9 . The apparatus of claim 8 , wherein the processor, to filter the input feature map to generate the output feature map having reduced blur component, is further configured to:
filter the input feature map consecutively by a plurality of atrous convolution network connected in series to repeatedly reduce the blur component.
10 . The apparatus of claim 8 , wherein the plurality of atrous convolutional layers comprises respective convolutional kernels containing kernel weights shared with the other atrous convolutional layers.
11 . The apparatus of claim 10 , wherein the convolutional kernels of the plurality of atrous convolutional layers are dilated by dilation rates different from each other.
12 . The apparatus of claim 11 , wherein the processor is further configured to:
perform convolutions between the input feature map and respective convolutional kernels in parallel to acquire a plurality of convolution results, using the plurality of atrous convolutional layers in the atrous convolution network; and
nonlinearly aggregate the plurality of convolution results by concatenating the plurality of convolution results and performing a convolution between a concatenation result and a network weight.
13 . The apparatus of claim 12 , wherein the processor is further configured to:
determine scale attention maps each comprising pixel-wise weights for respective one of the plurality of convolution results; and
nonlinearly aggregate the plurality of convolution results by multiplying each of the plurality of convolution results by a corresponding scale attention map, concatenating a plurality of multiplication results, and performing the convolution between the concatenation result and the network weight.
14 . The apparatus of claim 13 , wherein the processor is further configured to:
determine a channel-wise weight vector for compensating for a variation of a shape of the defocus blur in the input image according to a pixel position; and
nonlinearly aggregate the plurality of convolution results by multiplying each of the plurality of convolution results by a corresponding scale attention map, multiplying each of a plurality of multiplication results by the channel-wise weight vector, concatenating a plurality of multiplication results multiplied by the channel-wise weight vector, and performing the convolution between the concatenation result and the network weight.