Image processing method and apparatus using convolutional neural network
An apparatus is provided. The apparatus includes an input/output interface configured to receive an image and output a result, a memory storing one or more instructions for processing the image by using a convolutional neural network, and a processor configured to process the image by executing the one or more instructions, wherein the convolutional neural network (CNN) may include one or more spatial transformation modules, and the spatial transformation module may include a spatial transformer configured to apply a spatial transform to first input data that is the image or an output of a previous spatial transformation module, by using a spatial transformation function, a first convolutional layer configured to perform a convolution operation between the first input data to which the spatial transform is applied and a first filter, and a spatial inverse transformer configured to apply a spatial inverse transform to an output of the first convolutional layer.
1 . An image processing apparatus comprising:
an input/output interface configured to receive an image and output a processing result for the image;
a memory storing one or more instructions for processing the image by using a convolutional neural network; and
a processor configured to process the image by executing the one or more instructions,
wherein the convolutional neural network includes a plurality of spatial transformation modules,
wherein the one or more instructions, when executed by the processor, cause the image processing apparatus to perform, by using a first spatial transformation module among the plurality of spatial transformation modules, the following operations:
apply a spatial transform to first input data that is an image or an output of a previous spatial transformation module, by using a first spatial transformation function;
perform a convolution operation between the first input data to which the spatial transform is applied and a first filter; and
apply a spatial inverse transform to a result of the convolution operation by using a first spatial inverse transformation function, and
wherein the one or more instructions, when executed by the processor, cause the image processing apparatus to perform, by using a second spatial transformation module among the plurality of spatial transformation modules, the following operations:
apply a spatial transform to second input data that is an output of a previous spatial transformation module, by using a second spatial transformation function;
perform a convolution operation between the second input data to which the spatial transform is applied and a second filter; and
apply a spatial inverse transform to a result of the convolution operation by using a second spatial inverse transformation function,
wherein the second spatial transformation function is different from the first spatial transformation function.
2 . The image processing apparatus of claim 1 ,
wherein the first spatial transformation function and the second spatial transformation function are reversible functions.
3 . The image processing apparatus of claim 2 ,
wherein whether the first spatial transformation function is reversible is determined according to a form of the first input data, and
wherein whether the second spatial transformation function is reversible is determined according to a form of the second input data.
4 . The image processing apparatus of claim 1 , wherein the first spatial transformation function and the second spatial transformation function are each one of a permutation function, a rotation function, a flip function, or a scale function.
5 . The image processing apparatus of claim 1 , wherein the one or more instructions, when executed by the processor, cause the image processing apparatus to perform, by using a convolutional layer, a convolution operation between the second filter and the second input data that is the image or an output of a previous spatial transformation module.
6 . The image processing apparatus of claim 1 , wherein the one or more instructions, when executed by the processor, cause the image processing apparatus to:
perform, by using a pooling layer, sub sampling on an output of a previous layer or an output of a previous spatial transformation module; and
calculate, by using an output layer, a processing result for the image by using an output of a previous layer.
7 . The image processing apparatus of claim 6 , wherein the output layer calculates a location of an object included in the image.
8 . The image processing apparatus of claim 6 , wherein the output layer calculates a type of an object included in the image.
9 . An image processing method using a convolutional neural network including a plurality of spatial transformation modules, the image processing method comprising:
by a first spatial transformation module among the plurality of spatial transformation modules:
applying a spatial transform to first input data that is an image or an output of a previous spatial transformation module, by using a first spatial transformation function;
performing a convolution operation between the first input data to which the spatial transform is applied and a first filter; and
applying a spatial inverse transform to a result of the convolution operation by using a first spatial inverse transformation function, and
by a second spatial transformation module among the plurality of spatial transformation modules:
applying a spatial transform to second input data that is an output of a previous spatial transformation module, by using a second spatial transformation function;
performing a convolution operation between the second input data to which the spatial transform is applied and a second filter; and
applying a spatial inverse transform to a result of the convolution operation by using a second spatial inverse transformation function,
wherein the second spatial transformation function is different from the first spatial transformation function.
10 . The image processing method of claim 2 , wherein the first spatial transformation function and the second spatial transformation function are reversible functions.
11 . The image processing method of claim 10 , wherein whether the first spatial transformation function is reversible is determined according to a form of the first input data, and
wherein whether the second spatial transformation function is reversible is determined according to a form of the second input data.
12 . The image processing method of claim 9 , wherein the first spatial transformation function and the second spatial transformation function are each one of a permutation function, a rotation function, a flip function, or a scale function.
13 . The image processing method of claim 2 , further comprising:
by a convolutional layer, performing a convolution operation between the second filter and the second input data that is the image or an output of a previous spatial transformation module.
14 . The image processing method of claim 9 , further comprising:
performing, by a pooling layer, sub sampling on an output of a previous layer or an output of a previous spatial transformation module; and
calculating, by an output layer, a processing result for the image by using an output of a previous layer.
15 . The image processing method of claim 14 , wherein the output layer calculates a location of an object included in the image.
16 . The image processing method of claim 14 , wherein the output layer calculates a type of an object included in the image.
17 . The image processing method of claim 9 , wherein the processing comprises at least one of classification, detection, segmentation, or depth estimation performed on the image.
18 . A non-transitory computer-readable recording medium having recorded thereon a computer program for performing image processing operations, the operations comprising:
by a first spatial transformation module among a plurality of spatial transformation modules:
applying a spatial transform to first input data that is an image or an output of a previous spatial transformation module, by using a first spatial transformation function;
performing a convolution operation between the first input data to which the spatial transform is applied and a first filter; and
applying a spatial inverse transform to a result of the convolution operation by using a first spatial inverse transformation function, and
by a second spatial transformation module among the plurality of spatial transformation modules:
applying a spatial transform to second input data that is an output of a previous spatial transformation module, by using a second spatial transformation function;
performing a convolution operation between the second input data to which the spatial transform is applied and a second filter; and
applying a spatial inverse transform to a result of the convolution operation by using a second spatial inverse transformation function,
wherein the second spatial transformation function is different from the first spatial transformation function.