Neural network training method, image processing method, and apparatus
A neural network training method, includes: obtaining an input feature map of a training image; performing feature extraction processing on the input feature map by using a feature extraction core of a neural network to obtain a first candidate feature map; adding the first candidate feature map and a second candidate feature map to obtain an output feature map, where the second candidate feature map is a feature map obtained after a value corresponding to each element in the input feature map is increased by N times, and N is greater than 0; determining an image processing result of the training image based on the output feature map; and adjusting a parameter of the neural network based on the image processing result.
1 . A neural network training method comprising:
obtaining an input feature map of a training image;
performing feature extraction processing on the input feature map by using a feature extraction core of a neural network to obtain a first intermediate feature map, wherein the feature extraction processing enables each element in the first intermediate feature map to be an L 1 regular distance between the feature extraction core and data at a corresponding location in the input feature map;
adding the first intermediate feature map and a second intermediate feature map to obtain an output feature map, wherein the second intermediate feature map is a feature map obtained after a value corresponding to each element in the input feature map is increased by N times, and N is an integer that is greater than 0;
determining an image processing result of the training image based on the output feature map; and
adjusting a parameter of the neural network based on the image processing result;
wherein the adding of the first intermediate feature map and the second intermediate feature map to obtain the output feature map comprises:
processing the first intermediate feature map by using an activation function to obtain a processed first intermediate feature map; and
adding the processed first intermediate feature map and the second intermediate feature map to obtain the output feature map, and
wherein the processing of the first intermediate feature map by using the activation function comprises:
enhancing high-frequency texture information of the first intermediate feature map by using a power activation function, wherein the parameter of the neural network comprises a parameter of the power activation function.
2 . The method according to claim 1 , wherein the determining of the image processing result of the training image based on the output feature map comprises:
processing the output feature map by using another activation function to obtain a processed output feature map; and
determining the image processing result of the training image by using the processed output feature map.
3 . The method according to claim 2 , wherein the processing of the output feature map by using the another activation function comprises:
enhancing high-frequency texture information of the output feature map by using another power activation function, wherein the parameter of the neural network comprises a parameter of the another power activation function.
4 . The method according to claim 1 , wherein the power activation function is
( Y )=sign( Y )·| Y| α , wherein
Y is a feature map input into the power activation function, sign(⋅) is a symbolic function, |⋅| is an absolute value operation, α is a parameter of the power activation function, and α>0.
5 . An image processing method comprising:
obtaining an input feature map of a training image;
performing feature extraction processing on the input feature map by using a feature extraction core of a neural network to obtain a first intermediate feature map, wherein the feature extraction processing enables each element in the first intermediate feature map to be an L 1 regular distance between the feature extraction core and data at a corresponding location in the input feature map;
adding the first intermediate feature map and a second intermediate feature map to obtain an output feature map, wherein the second intermediate feature map is a feature map obtained after a value corresponding to each element in the input feature map is increased by N times, and N is an integer that is greater than 0;
determining an image processing result of the training image based on the output feature map;
adjusting a parameter of the neural network based on the image processing result to obtain a trained neural network;
obtaining an input feature map of a to-be-processed image; and
performing image processing on the input feature map of the to-be-processed image by using the trained neural network to obtain an image processing result of the to-be-processed image;
wherein the adding of the first intermediate feature map a and the second intermediate feature map to obtain the output feature map comprises:
processing the first intermediate feature map by using an activation function to obtain a processed first intermediate feature map, and
adding the processed first intermediate feature map and the second intermediate feature map to obtain the output feature map; and
serein the processing of the first intermediate feature map by using the activation function comprises:
enhancing high-frequency texture information of the first intermediate feature map by using a power activation function, wherein the parameter of the neural network comprises a parameter of the power activation function.
6 . The method according to claim 5 , wherein the image processing comprises at least one of image super-resolution processing, image denoising processing, image demosaicing processing, or image deblurring processing.
7 . The method according to claim 5 , wherein the determining of the image processing result of the training image based on the output feature map comprises:
processing the output feature map by using another activation function to obtain a processed output feature map; and
determining the image processing result of the training image by using the processed output feature map.
8 . The method according to claim 7 , wherein the processing of the output feature map by using the another activation function comprises:
enhancing high-frequency texture information of the output feature map by using another power activation function, wherein the parameter of the neural network comprises a parameter of the another power activation function.
9 . The method according to claim 5 , wherein the power activation function is
( Y )=sign( Y )·| Y| α , wherein
Y is a feature map input into the power activation function, sign(⋅) is a symbolic function, |⋅| is an absolute value operation, α is a parameter of the power activation function, and α>0.
10 . A neural network training apparatus comprising:
a processor and a non-transitory computer-readable memory storing instructions which upon execution by the processor cause the neural network training apparatus to at least be configured to:
obtain an input feature map of a training image;
perform feature extraction processing on the input feature map by using a feature extraction core of a neural network to obtain a first intermediate feature map, wherein the feature extraction processing enables each element in the first intermediate feature map to be an L 1 regular distance between the feature extraction core and data at a corresponding location in the input feature map;
perform processing including adding the first intermediate feature map and a second intermediate feature map to obtain an output feature map, wherein the second intermediate feature map is a feature map obtained after a value corresponding to each element in the input feature map is increased by N times, and N is an integer that is greater than 0;
perform image processing including determining an image processing result of the training image based on the output feature map; and
adjust a parameter of the neural network based on the image processing result,
wherein in order to perform the processing, the neural network training apparatus is further configured to:
process the first intermediate feature map by using an activation function to obtain a processed first intermediate feature map; and
add the processed first intermediate feature map and the second intermediate feature map to obtain the output feature map; and
wherein in order to perform the processing, the neutral network training apparatus is further configured to:
enhance high-frequency texture information of the first intermediate feature man by using a power activation function, wherein the parameter of the neural network comprises a para meter of the power activation function.
11 . The apparatus according to claim 10 , wherein in order to perform the image processing, the neural network training apparatus is further configured to:
process the output feature map by using another activation function to obtain a processed output feature map; and
determine the image processing result of the training image by using the processed output feature map.
12 . The apparatus according to claim 11 , wherein in order to perform the image processing, the neural network training apparatus is further configured to:
enhance high-frequency texture information of the output feature map by using another power activation function, wherein the parameter of the neural network comprises a parameter of the another power activation function.
13 . The apparatus according to claim 10 , wherein the power activation function is
( Y )=sign( Y )·| Y| α , wherein
Y is a feature map input into the power activation function, sign(⋅) is a symbolic function, |⋅| is an absolute value operation, α is a parameter of the power activation function, and α>0.
14 . The method according to claim 1 , wherein the parameter of the power activation function is learnable such that the power activation function is a learnable power activation function trained to help the neural network adapt to different tasks and/or scenarios.