Method and device of inverse tone mapping and electronic device

1. A method of inverse tone mapping, comprising steps of:

acquiring one or more low dynamic range images;

performing a decomposition operation to the one or more low dynamic range images to acquire a detail layer and a basic layer of the low dynamic range image;

restoring the detail layer and the basic layer by using a predetermined first restoration network and a predetermined second restoration network to acquire restored detail layer and basic layer, respectively; and

adjusting the restored detail layer and basic layer by using a predetermined fusion network to acquire an adjusted high dynamic range image;

wherein the first restoration network is a residual network and the second first restoration network is a U-Net network;

wherein the residual network comprises one or more convolution layers on both sides and multiple residual blocks in the middle, and each of the residual blocks contains a first convolution layer, a second activation layer, and a third convolution layer and a fourth activation layer arranged in sequence; and before the fourth activation layer further comprises: performing an addition operation on an input image of the residual block and an output image of the third convolution layer; and

wherein the U-Net network comprises multiple convolution blocks and deconvolution blocks, the multiple convolution blocks are located in front of the multiple deconvolution blocks, and each of the multiple convolution block comprises a convolution layer, an activation layer, and a convolution layer and an activation layer arranged in sequence, and in each of the multiple deconvolution blocks, first up-sampling to expand the resolution of the a feature map, and then performing the a convolution operation; each of the multiple deconvolution blocks contains an up- sampling layer and a convolution layer and an activation layer arranged in sequence.

2. The method according to claim 1 , wherein the step of acquiring one or more low dynamic range images comprises step of:

compressing an original image to acquire a compressed low dynamic range image.

3. The method according to claim 2 , wherein the step of performing a decomposition operation to the low dynamic range image to acquire a detail layer and a basic layer of the low dynamic range image comprises step of:

performing a decomposition operation to the low dynamic range image based on the Retinex theory to acquire the detail layer and the basic layer of the low dynamic range image.

4. The method according to claim 3 , wherein the step of performing a decomposition operation to the low dynamic range image based on the Retinex theory to acquire the detail layer and the basic layer of the low dynamic range image specifically comprises step of:

performing edge-preserving filtering to the low dynamic range image, and using the image acquired after the edge-preserving filtering as the basic layer of the low dynamic range image; and

calculating a difference between the low dynamic range image and the basic layer image, and the image acquired after the difference being used as the detail layer of the low dynamic range image.

5. The method according to claim 1 , wherein the step of restoring the detail layer and the basic layer by using a predetermined first restoration network and a predetermined second restoration network comprises step of:

restoring the detail layer by using the residual network and restoring the basic layer by using the U-Net network.

6. The method according to claim 5 , wherein the detail layer contains high frequency components and compression artifacts of the low dynamic range image, and the basic layer contains low frequency components of the low dynamic range image; the step of restoring the detail layer by using the residual network and restoring the basic layer by using the U-Net network comprises step of:

restoring the high frequency components by using the residual network to remove the compression artifacts, and restoring the low frequency components by using U-Net network.

7. The method according to claim 6 , wherein the high frequency components comprise edges and contours, and the low frequency components comprise color information and structural information.

8. The method according to claim 1 , wherein the fusion network uses a residual network.

9. An electronic device, comprising: a storage device, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program the processor implements steps as following:

acquiring one or more low dynamic range images;

performing a decomposition operation to the one or more low dynamic range images to acquire a detail layer and a basic layer of the low dynamic range image;

restoring the detail layer and the basic layer by using a predetermined first restoration network and a predetermined second restoration network to acquire restored detail layer and basic layer, respectively; and

adjusting the restored detail layer and basic layer by using a predetermined fusion network to acquire an adjusted high dynamic range image;

wherein the first restoration network is a residual network and the second first restoration network is a U-Net network;

wherein the residual network comprises one or more convolution layers on both sides and multiple residual blocks in the middle, and each of the residual blocks contains a first convolution layer, a second activation layer, and a third convolution layer and a fourth activation layer arranged in sequence; and before the fourth activation layer further comprises: performing an addition operation on an input image of the residual block and an output image of the third convolution layer; and

wherein the U-Net network comprises multiple convolution blocks and deconvolution blocks, the multiple convolution blocks are located in front of the multiple deconvolution blocks, and each of the multiple convolution block comprises a convolution layer, an activation layer, and a convolution layer and an activation layer arranged in sequence, and in each of the multiple deconvolution blocks, first up-sampling to expand the resolution of a feature map, and then performing a convolution operation; each of the multiple deconvolution blocks contains an up-sampling layer and a convolution layer and an activation layer arranged in sequence.