Image processing apparatus, imaging apparatus, image processing method, and storage medium

1. An image processing apparatus comprising one or more processors configured to:

convert an input image into information of a first feature amount by inputting the input image to a convolution layer of a first convolutional neural network including at least two layers;

convert the input image with a same resolution as the input image input to the convolution layer of the first convolutional neural network into information of a second feature amount by inputting the input image with the same resolution as the input image input to the convolution layer of the first convolutional neural network to a convolution layer of a second convolutional neural network; and

convert information of a feature amount including a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network.

2. The image processing apparatus according to claim 1 , wherein the one or more processors, in converting the input image into the information of the first feature amount, generate a multiscale feature map as the information of the first feature amount by repeatedly performing nonlinear calculation processing after convolving a plurality of filters with the input image.

3. The image processing apparatus according to claim 1 , wherein the one or more processors, in converting the input image with the same resolution as the input image input to the convolution layer of the first convolutional neural network, generate a multiscale feature map as the information of the second feature amount by convolving a plurality of filters of a plurality of types of sizes with the input image.

4. The image processing apparatus according to claim 1 , wherein the information of the first feature amount and the information of the second feature amount are added to each other by calculating a sum of the information of the first feature amount and the information of the second feature amount having the same scale as that of the information of the first feature amount for each element.

5. The image processing apparatus according to claim 1 , wherein the one or more processors generate the information of the third feature amount by adding the information of the first feature amount and the information of the second feature amount to each other.

6. The image processing apparatus according to claim 1 , wherein the one or more processors convert the information of the third feature amount into the output image by deconvolving a plurality of filters of a plurality of types of sizes with the information of the third feature amount and then by weighing a feature map obtained by performing a nonlinear calculation.

7. The image processing apparatus according to claim 6 , wherein the one or more processors, in converting the information of the feature amount including the third feature amount, use Leaky ReLU or Parametric ReLU for the nonlinear calculation.

8. The image processing apparatus according to claim 1 , wherein:

the first convolutional neural network uses a first filter and a first bias, and

the second convolutional neural network uses a second filter and a second bias, which are different from the first filter and the first bias, respectively.

9. The image processing apparatus according to claim 1 , wherein the one or more processors:

are further configured to acquire the input image,

convert the acquired input image into the information of the first feature amount, and

convert the acquired input image into the information of the second feature amount.

10. The image processing apparatus according to claim 9 , wherein the acquired input image is a deteriorated image.

11. The image processing apparatus according to claim 1 , wherein the one or more processors, in converting the information of the feature amount including the third feature amount, output a sharp image obtained by sharpening the input image as the output image.

12. An imaging apparatus comprising:

the image processing apparatus according to claim 1 ; and

an imaging apparatus that acquires the input image.

13. An image processing method comprising the steps of:

converting an input image into information of a first feature amount by inputting the input image to a convolution layer of a first convolutional neural network including at least two layers;

converting the input image with a same resolution as the input image input to the convolution layer of the first convolutional neural network into information of a second feature amount by inputting the input image with the same resolution as the input image input to the convolution layer of the first convolutional neural network to a convolution layer of a second convolutional neural network; and

converting information of a feature amount including a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network.

14. A non-transitory computer-readable storage medium storing a program executable by a computer to execute the image processing method according to claim 13 .

15. An image processing apparatus comprising one or more processors configured to:

convert an input image into information of a first feature amount using a first convolutional neural network including at least two layers;

convert the input image into information of a second feature amount using a second convolutional neural network; and

convert information of a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network,

wherein the one or more processors, in converting the input image into the information of the first feature amount, generate a multiscale feature map as the information of the first feature amount by repeatedly performing nonlinear calculation processing after convolving a plurality of filters with the input image.

16. An image processing apparatus comprising one or more processors configured to:

convert an input image into information of a first feature amount using a first convolutional neural network having at least two layers;

convert the input image into information of a second feature amount using a second convolutional neural network; and

convert information of a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network,

wherein the one or more processors convert the information of the third feature amount into the output image by deconvolving a plurality of filters of a plurality of types of sizes with the information of the third feature amount and then by weighing a feature map obtained by performing a nonlinear calculation.

17. An image processing apparatus comprising one or more processors configured to:

acquire an input image;

convert the acquired input image into information of a first feature amount using a first convolutional neural network including at least two layers;

convert the acquired input image into information of a second feature amount using a second convolutional neural network; and

convert information of a third feature amount generated by adding the information of the first feature amount and the information of the second feature amount to each other, into an output image using a third convolutional neural network.