Systems and methods of optimized deep learning for image reconstruction
A deep learning network can have the following elements arranged in sequence: a convolutional block, a first residual block (RB), a first element-wise adder, a second RB, a second element-wise adder, and an upsampling unit. At least one processor can perform processing comprising generating a super resolution image by processing, with the deep learning network, an image having a resolution lower than a resolution of the super resolution image.
1 . A system comprising:
at least one processor; and
at least one non-transitory computer-readable medium storing:
a deep learning network having the following elements arranged in sequence:
a convolutional block,
a first residual block (RB),
a first element-wise adder,
a second RB,
a second element-wise adder, and
an upsampling layer; and
instructions that, when executed by the at least one processor, cause the at least one processor to perform processing comprising generating a super resolution image by processing, with the deep learning network, an image having a resolution lower than a resolution of the super resolution image,
wherein each of the first RB and the second RB respectively includes six serially connected dense residual Laplacian modules (DRLMs), six concatenation modules, and six convolutional units,
wherein each DRLM respectively includes one or more residual units (RUs), one or more concatenation units, one or more compression units, and one or more Laplacian attention units, and
wherein each RU respectively includes one and only one convolutional layer and one and only one rectified linear unit.
2 . The system of claim 1 , wherein each DRLM is configured to perform processing comprising:
obtaining a residual;
performing dense concatenation;
compressing parameters from the dense concatenation; and
performing Laplacian attention processing.
3 . The system of claim 1 , wherein the processing comprises:
performing feature extraction;
performing cascading over residual on a residual;
performing upsampling; and
performing reconstruction.
4 . The system of claim 3 , wherein the feature extraction comprises processing by the convolutional block.
5 . The system of claim 3 , wherein the cascading over residual on the residual comprises processing through the first RB and second RB with medium skip-connections.
6 . The system of claim 3 , wherein the upsampling comprises processing by the upsampling layer to increase the resolution of the image by including additional pixels based on interpolating using existing data.
7 . The system of claim 1 , wherein a number of feature maps in all the convolutional layers of the deep learning network is thirty two.
8 . The system of claim 1 , wherein a number of residual blocks in the deep learning network is two.
9 . The system of claim 1 , wherein each RU does not include an adder.
10 . A method comprising:
provisioning, by at least one processor, a deep learning network having the following elements arranged in sequence:
a convolutional block,
a first residual block (RB),
a first element-wise adder,
a second RB,
a second element-wise adder, and an upsampling layer; and
generating, by the at least one processor, a super resolution image by processing, with the deep learning network, an image having a resolution lower than a resolution of the super resolution image,
wherein each of the first RB and the second RB respectively includes six serially connected dense residual Laplacian modules (DRLMs), six concatenation modules, and six convolutional units,
wherein each DRLM respectively includes one or more residual units (RUs), one or more concatenation units, one or more compression units, and one or more Laplacian attention units, and
wherein each RU respectively includes one and only one convolutional layer and one and only one rectified linear unit.
11 . The method of claim 10 , wherein each DRLM performs processing comprising:
obtaining a residual;
performing dense concatenation;
compressing parameters from the dense concatenation; and
performing Laplacian attention processing.
12 . The method of claim 10 , wherein the processing comprises:
performing feature extraction;
performing cascading over residual on a residual;
performing upsampling; and
performing reconstruction.
13 . The method of claim 12 , wherein the feature extraction comprises processing by the convolutional block.
14 . The method of claim 12 , wherein the cascading over residual on the residual comprises processing through the first RB and second RB with medium skip-connections.
15 . The method of claim 12 , wherein the upsampling comprises processing by the upsampling layer to increase the resolution of the image by including additional pixels based on interpolating using existing data.