Convolutional neural network, and processing method, processing device, processing system and medium for the same

1. A processing method for a convolutional neural network, comprising steps of:

using an activation recorder layer as an activation function layer in the convolutional neural network, wherein in response to that a probe image with contents is inputted to the convolutional neural network, the activation recorder layer performs an activation operation the same as the activation function layer does and records an activation result of the activation operation;

modifying the convolutional neural network, wherein step of modifying comprises replacing the activation recorder layer with a hidden layer that uses the recorded activation result; and

inputting an analysis image to the modified convolutional neural network as an input image, so as to output an output image of the modified convolutional neural network, thereby analyzing a forward effect or a backward effect between the input image and the output image, the analysis image being a pixel-level binary image.

2. The processing method of claim 1 , wherein step of inputting an analysis image to the modified convolutional neural network as an input image, so as to obtain an analysis result, thereby analyzing a forward effect between the input image and the output image of the convolutional neural network before modification comprises:

inputting an all-zero image to the modified convolutional neural network so as to output a first output value as a bias coefficient;

inputting one or more analysis images to the modified convolutional neural network so as to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

analyzing correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a forward effect.

3. The processing method of claim 1 , wherein the convolutional neural network comprises an upsampling layer.

4. The processing method of claim 1 , wherein step of inputting an analysis image to the modified convolutional neural network as an input image, so as to obtain an analysis result, thereby analyzing a backward effect between the input image and the output image of the convolutional neural network before modification comprises:

inputting an all-zero image to the modified convolutional neural network to output a first output value as a bias coefficient;

inputting all possible different analysis images to the modified convolutional neural network to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels, and positions of the pixel that is 1 being different in different analysis images; and

analyzing correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the output image of the convolutional neural network on respective pixels of the input image as a backward effect.

5. The processing method of claim 1 , wherein

step of modifying the convolutional neural network comprises replacing the activation recorder layer with a hidden layer that uses the recorded activation result comprises:

configuring a deconvolutional network as the modified convolutional neural network, the deconvolutional network being an inverse network of the convolutional neural network,

step of inputting an analysis image to the modified convolutional neural network as an input image, so as to obtain an analysis result, thereby analyzing a backward effect between the input image and the output image of the convolutional neural network before modification comprises:

inputting an all-zero image to the modified convolutional neural network so as to output a first output value as a bias coefficient;

inputting one or more analysis images to the modified convolutional neural network so as to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

analyzing correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a backward effect.

6. The processing method of claim 5 , wherein the convolutional neural network comprises an upsampling layer, and configuring a deconvolutional network as the modified convolutional neural network comprises:

replacing the upsampling layer with a downsampling layer corresponding to the upsampling layer.

7. A processing device for a convolutional neural network, comprising:

a recorder, configured to use an activation recorder layer as an activation function layer in the convolutional neural network, wherein in response to that a probe image with contents is inputted to the convolutional neural network, the recorder enables the activation recorder layer to perform an activation operation the same as the activation function layer does and record an activation result of the activation operation;

a modifier, configured to modify the convolutional neural network, wherein step of modifying comprises replacing the activation recorder layer with a hidden layer that uses the recorded activation result; and

an analyzer, configured to input an analysis image to the modified convolutional neural network as an input image, so as to output an output image of the modified convolutional neural network, thereby analyzing a forward effect or a backward effect between the input image and the output image, the analysis image being a pixel-level binary image.

8. The processing device of claim 7 , wherein the analyzer is configured to:

input an all-zero image to the modified convolutional neural network so as to output a first output value as a bias coefficient;

input one or more analysis images to the modified convolutional neural network so as to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

analyze correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a forward effect.

9. The processing device of claim 7 , wherein the convolutional neural network comprises an upsampling layer.

10. The processing device of claim 7 , wherein the analyzer is configured to:

input an all-zero image to the modified convolutional neural network to output a first output value as a bias coefficient;

input all possible different analysis images to the modified convolutional neural network to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels, and positions of the pixel that is 1 being different in different analysis images; and

analyze correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the output image of the convolutional neural network on respective pixels of the input image as a backward effect.

11. The processing device of claim 7 , wherein

the modifier is configured to:

configure a deconvolutional network as the modified convolutional neural network, the deconvolutional network being an inverse network of the convolutional neural network,

the analyzer is configured to:

input an all-zero image to the modified convolutional neural network so as to output a first output value as a bias coefficient;

input one or more analysis images to the modified convolutional neural network so as to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

analyze correspondence between the analysis image and the output image at a pixel level, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a backward effect.

12. The processing device of claim 11 , wherein the convolutional neural network comprises an upsampling layer, and the analyzer is configured to configure a deconvolutional network as the modified convolutional neural network through the following step:

replacing the upsampling layer with a downsampling layer corresponding to the upsampling layer.

13. A processing system for a convolutional neural network, comprising:

one or more processors;

one or more memories in which computer-readable codes are stored, the computer-readable codes being capable of causing, when executed by the one or more processors, the processing method of claim 1 to be executed.

14. A convolutional neural network, comprising:

one or more convolutional layers;

one or more hidden layers corresponding to the one or more convolutional layers, the one or more hidden layers being used to replace corresponding one or more activation recorder layers, the one or more activation recorder layers serving as an activation function layer in the convolutional neural network, wherein in response to that a probe image with contents is inputted to the convolutional neural network, the one or more activation recorder layers perform an activation operation the same as the activation function layer does and record an activation result of the activation operation, the one or more hidden layers use the recorded activation result;

an input terminal that receives one or more analysis images; and

on output terminal that outputs an output image of the modified convolutional neural network, thereby analyzing a forward effect or a backward effect between the input image and the output image, the analysis image being a pixel-level binary image.

15. The convolutional neural network of claim 14 , wherein before receiving the analysis images, the input terminal receives an all-zero image so as to output a first output value from the output terminal as a bias coefficient;

the input terminal is configured to receive the one or more analysis images so as to output an output image of the modified convolutional neural network from the output terminal based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

correspondence between the analysis image and the output image at a pixel level is analyzed, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a forward effect.

16. The convolutional neural network of claim 14 , further comprising an upsampling layer.

17. The convolutional neural network of claim 14 , wherein before receiving the analysis images, the input terminal receives an all-zero image to output a first output value from the output terminal as a bias coefficient;

the input terminal receives all possible different analysis images to output an output image of the modified convolutional neural network from the output terminal based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels, and positions of the pixel that is 1 being different in different analysis images; and

correspondence between the analysis image and the output image at a pixel level is analyzed, so as to obtain the effect of respective pixels of the output image of the modified convolutional neural network on respective pixels of the input image as a backward effect.

18. The convolutional neural network of claim 14 , wherein the convolutional neural network may be replaced with a deconvolutional network as the modified convolutional neural network, the deconvolutional network is an inverse network of the convolutional neural network, the input terminal is replaced with an output terminal of the modified convolutional neural network, and the output terminal is replaced with an input terminal of the modified convolutional neural network;

the input terminal of the modified convolutional neural network receives an all-zero image so as to output a first output value from the output terminal of the modified convolutional neural network as a bias coefficient;

the input terminal of the modified convolutional neural network receives one or more analysis images so as to output an output image of the modified convolutional neural network based on the bias coefficient, the analysis image being a binary image that is 1 at a certain pixel and 0 at other pixels; and

correspondence between the analysis image and the output image at a pixel level is analyzed, so as to obtain the effect of respective pixels of the input image of the modified convolutional neural network on respective pixels of the output image as a backward effect.

19. The convolutional neural network of claim 18 , further comprising an upsampling layer, wherein the upsampling layer is replaced with a downsampling layer corresponding to the upsampling layer in the deconvolutional network.

20. A non-transitory computer storage medium for storing computer-readable codes capable of causing, when executed by one or more processors, the processing method of claim 1 to be executed.