System and method for estimating optimal parameters
A method and system of generating an adjustment parameter value for a control parameter to enhance a new image, which includes configuring a neural network, trained to restore image quality for a derivative image, to that of an earlier version of the derivative image, to generate as an output the adjustment parameter value, for the control parameter in response to input of data derived from the new image, and changing a control parameter of the new image, by generating the adjustment parameter value by calculating an inverse of the output value, and applying the adjustment parameter value to the control parameter of the new image so as to generate an enhanced image.
1. A method of training a neural network, the method comprising:
providing the neural network, the neural network comprising:
a layer;
a first head downstream from the layer, the first head outputting a first output value associated with a first control parameter; and
a second head downstream from the layer and in parallel with the first head, the second head outputting a second output value associated with a second control parameter;
inputting first control parameter training images into the layer of the neural network;
receiving the first output value from the first head;
comparing the first output value and a first modified parameter value associated with at least one of the first control parameter training images;
generating, based at least in part on the comparing, a first control parameter error value;
adjusting the layer and the first head based at least in part on the first control parameter error value;
inputting second control parameter training images into the neural network;
receiving the second output value from the second head;
comparing the second output value and a second modified parameter value associated with at least one of the second control parameter training images;
generating, based at least in part on the comparing, a second control parameter error value; and
adjusting the second head based at least in part on the second control parameter error value.
2. The method according to claim 1 , further comprising:
determining a first learning rate associated with the first head; and
determining a second learning rate associated with the second head, the second learning rate different from the first learning rate,
wherein the layer or the first head is adjusted based at least in part on the first learning rate and the second head is adjusted based at least in part on the second learning rate.
3. The method according to claim 1 , wherein adjusting the second head comprises adjusting the second head while not adjusting at least a portion of the layer or the first head.
4. The method according to claim 1 , wherein the first control parameter is an inverse of the first modified parameter value.
5. The method according to claim 1 , wherein the first control parameter or the second control parameter comprises at least one of: brightness, contrast, saturation, sharpness, blur, denoise strength, tint, color temperature, parametric transformation, linear piece-wise transformation, tone compression, or a face beautification parameter.
6. The method according to claim 1 , further comprising:
providing a trained neural network based at least in part on adjusting the layer, adjusting the first head, or adjusting the second head; and
enhancing an image input into the trained neural network.
7. The method according to claim 1 , wherein the first head or the second head represents a convolutional neural network.
8. The method according to claim 1 , further comprising:
training the neural network based at least in part on adjusting the first head and adjusting the second head; and
providing the trained neural network as a service over a network.
9. A system comprising:
one or more processors;
memory storing instructions that, when executed by the one or more processors, causes the system to perform operations comprising:
providing a neural network, the neural network comprising:
a layer;
a first head downstream from the layer, the first head outputting a first output value associated with a first control parameter; and
a second head downstream from the layer and in parallel with the first head, the second head outputting a second output value associated with a second control parameter;
inputting first control parameter training images into the layer of the neural network;
receiving the first output value from the first head;
comparing the first output value and a first modified parameter value associated with at least one of the first control parameter training images;
generating, based at least in part on the comparing, a first control parameter error value;
adjusting the layer and the first head based at least in part on the first control parameter error value;
inputting second control parameter training images into the neural network;
receiving the second output value from the second head;
comparing the second output value and a second modified parameter value associated with at least one of the second control parameter training images;
generating, based at least in part on the comparing, a second control parameter error value; and
adjusting the second head based at least in part on the second control parameter error value.
10. The system according to claim 9 , the operations further comprising:
determining a first learning rate associated with the first head; and
determining a second learning rate associated with the second head, the second learning rate different from the first learning rate,
wherein the layer or the first head is adjusted based at least in part on the first learning rate and the second head is adjusted based at least in part on the second learning rate.
11. The system according to claim 9 , wherein adjusting the second head comprises adjusting the second head while not adjusting at least a portion of the layer or the first head.
12. The system according to claim 9 , wherein the first control parameter or the second control parameter comprises at least one of: brightness, contrast, saturation, sharpness, blur, denoise strength, tint, color temperature, parametric transformation, linear piece-wise transformation, tone compression, or a face beautification parameter.
13. The system according to claim 9 , the operations further comprising:
providing a trained neural network based at least in part on adjusting the layer, adjusting the first head, or adjusting the second head; and
enhancing an image input into the trained neural network.
14. The method according to claim 1 , wherein the at least one of the first control parameter training images is based at least in part on an image defined by a user.
15. The method according to claim 1 , wherein the at least one of the first control parameter training images is based at least in part on an image comprising an artistic style.
16. The method according to claim 1 , wherein the layer is a first layer, the neural network comprises a second layer, and further comprising inputting an output of the second layer into the first head.
17. The system according to claim 9 , wherein the first control parameter is an inverse of the first modified parameter value.
18. The system according to claim 9 , wherein the at least one of the first control parameter training images is based at least in part on an image defined by a user.
19. The system according to claim 9 , wherein the at least one of the first control parameter training images is based at least in part on an image comprising an artistic style.
20. The system according to claim 9 , wherein the layer is a first layer, the neural network comprises a second layer, and further comprising inputting an output of the second layer into the first head.
21. A device comprising: one or more processors; memory storing device instructions that, when executed by the one or more processors, causes the device to perform operations comprising:
providing a neural network, the neural network comprising:
a layer;
a first head downstream from the layer, the first head outputting a first output value associated with a first control parameter; and
a second head downstream from the layer and in parallel with the first head, the second head outputting a second output value associated with a second control parameter;
inputting first control parameter training images into the layer of the neural network;
receiving the first output value from the first head;
comparing the first output value and a first modified parameter value associated with at least one of the first control parameter training images;
generating, based at least in part on the comparing, a first control parameter error value;
adjusting the layer and the first head based at least in part on the first control parameter error value;
inputting second control parameter training images into the neural network;
receiving the second output value from the second head;
comparing the second output value and a second modified parameter value associated with at least one of the second control parameter training images;
generating, based at least in part on the comparing, a second control parameter error value; and
adjusting the second head based at least in part on the second control parameter error value.
22. The device according to claim 21 , the operations further comprising:
determining a first learning rate associated with the first head; and
determining a second learning rate associated with the second head, the second learning rate different from the first learning rate,
wherein the layer or the first head is adjusted based at least in part on the first learning rate and the second head is adjusted based at least in part on the second learning rate.
23. The device according to claim 21 , wherein adjusting the second head comprises adjusting the second head while not adjusting at least a portion of the layer or the first head.
24. The device according to claim 21 , wherein the first control parameter is an inverse of the first modified parameter value.
25. The device according to claim 21 , wherein the first control parameter or the second control parameter comprises at least one of: brightness, contrast, saturation, sharpness, blur, denoise strength, tint, color temperature, parametric transformation, linear piece-wise transformation, tone compression, or a face beautification parameter.
26. The device according to claim 21 , the operations further comprising:
providing a trained neural network based at least in part on adjusting the layer, adjusting the first head, or adjusting the second head; and
enhancing an image input into the trained neural network.
27. The device according to claim 21 , the operations further comprising:
training the neural network based at least in part on adjusting the first head and adjusting the second head; and
providing the trained neural network as a service over a network.