Electronic device and method for training or applying neural network model

1. An electronic device for training or applying a neural network model, comprising:

a transceiver;

a storage medium configured to store multiple modules and the neural network model; and

a processor configured to couple to the storage medium and the transceiver, and configure to access and execute the modules, wherein the modules comprise:

a data collection module configured to receive an input data via the transceiver; and

a calculation module configured to perform convolution on the input data to generate a high-frequency feature map and a low-frequency feature map, and perform one of upsampling and downsampling to match a first size of the high-frequency feature map and a second size of the low-frequency feature map, concatenate the high-frequency feature map and the low-frequency feature map to generate a concatenated data, and input the concatenated data to an output layer of the neural network model.

2. The electronic device according to claim 1 , wherein the calculation module performs the downsampling on the high-frequency feature map to match the first size and the second size.

3. The electronic device according to claim 2 , wherein the calculation module concatenates the high-frequency feature map and the low-frequency feature map in response to matching between the first size the second size to generate the concatenated data.

4. The electronic device according to claim 1 , wherein the calculation module performs the upsampling on the low-frequency feature map to match the first size and the second size.

5. The electronic device according to claim 4 , wherein the calculation module concatenates the high-frequency feature map and the low-frequency feature map in response to matching between the first size and the second size to generate the concatenated data.

6. The electronic device according to claim 1 , wherein the calculation module divides the input data into a high-frequency input data and a low-frequency input data, and performs an XOR operation on a function of the high-frequency input data and a function of the upsampled low-frequency input data to generate the high-frequency feature map.

7. The electronic device according to claim 1 , wherein the calculation module divides the input data into a high-frequency input data and a low-frequency input data, and performs an XOR operation on a function of the downsampled high-frequency input data and a function of the low-frequency input data to generate the low-frequency feature map.

8. A method for training or applying a neural network model, comprising:

receiving an input data;

performing convolution on the input data to generate a high-frequency feature map and a low-frequency feature map;

performing one of upsampling and downsampling to match a first size of the high-frequency feature map and a second size of the low-frequency feature map;

concatenating the high-frequency feature map and the low-frequency feature map to generate a concatenated data; and

inputting the concatenated data to an output layer of the neural network model.

9. The method according to claim 8 , wherein the step of performing one of the upsampling and the downsampling to match the first size of the high-frequency feature map and the second size of the low-frequency feature map comprises:

performing the downsampling on the high-frequency feature map to match the first size and the second size.

10. The method according to claim 9 , wherein the step of concatenating the high-frequency feature map and the low-frequency feature map to generate the concatenated data comprises:

concatenating the high-frequency feature map and the low-frequency feature map in response to matching between the first size and the second size to generate the concatenated data.

11. The method according to claim 8 , wherein the step of performing one of the upsampling and the downsampling to match the first size of the high-frequency feature map and the second size of the low-frequency feature map comprises:

performing the upsampling on the low-frequency feature map to match the first size and the second size.

12. The method according to claim 11 , wherein the step of concatenating the high-frequency feature map and the low-frequency feature map to generate the concatenated data comprises:

concatenating the high-frequency feature map and the low-frequency feature map in response to matching between the first size and the second size to generate the concatenated data.

13. The method according to claim 8 , wherein the step of performing the convolution on the input data to generate the high-frequency feature map and the low-frequency feature map comprises:

dividing the input data into a high-frequency input data and a low-frequency input data; and

performing an XOR operation on a function of the high-frequency input data and a function of the upsampled low-frequency input data to generate the high-frequency feature map.

14. The method according to claim 8 , wherein the step of performing the convolution on the input data to generate the high-frequency feature map and the low-frequency feature map comprises:

dividing the input data into a high-frequency input data and a low-frequency input data; and

performing an XOR operation on a function of the downsampled high-frequency input data and a function of the low-frequency input data to generate the low-frequency feature map.