Method for image coding using convolution neural network and apparatus thereof

1. An image decoding method performed by a decoding apparatus comprising:

obtaining residual information about a current block from a bitstream;

obtaining an index indicating one of a plurality of Convolution Neural Networks (CNNs) from the bitstream;

deriving a prediction sample for the current block;

deriving a residual sample for the current block based on the residual information;

deriving a reconstructed picture based on the prediction sample and the residual sample;

selecting a CNN for the reconstructed picture based on the index; and

performing a filtering on the reconstructed picture based on the selected CNN,

wherein performing the filtering on the reconstructed picture comprises:

splitting the reconstructed picture into input blocks having a specific size,

deriving features by performing a convolution operation for each of the input blocks based on a kernel for the selected CNN,

determining a filter set based on the features, and

performing a filtering on the reconstructed picture based on the filter set.

2. The image decoding method of claim 1 , wherein the sizes of the input blocks are 3×3 sizes or 5×5 sizes.

3. The image decoding method of claim 1 , comprising:

obtaining a flag indicating whether a post filtering is performed through the bitstream; and

obtaining a flag indicating whether the CNN-based filtering is performed through the bitstream, based on a value of the flag indicating whether the post filtering is performed being 1.

4. The image decoding method of claim 3 , comprising: obtaining a flag indicating whether a deblocking filtering is available, a flag indicating whether a sample adaptive offset (SAO) is available, and a flag indicating whether an adaptive loop filter (ALF) is available through the bitstream, based on the value of the flag indicating whether the CNN-based filtering is performed being 1.

5. The image decoding method of claim 1 , wherein performing the filtering on the reconstructed picture comprises:

performing a deblocking filtering, a sample adaptive offset (SAO), and an adaptive loop filter (ALF) on the reconstructed picture;

splitting the filtered reconstructed picture into input blocks having a specific size;

deriving features by performing a convolution operation for each of the input blocks based on a kernel for the selected CNN;

determining a filter set based on the features; and

performing a filtering on the reconstructed picture based on the filter set.

6. The image decoding method of claim 1 , wherein each CNN of the plurality of CNNs is different from other CNNs in at least one of a number of layers, a number of filter taps, or filter coefficients.

7. An image encoding method performed by an encoding apparatus, the method comprising:

deriving a prediction sample for a current block;

generating a residual sample for the current block based on the prediction sample and an original sample for the current block;

deriving a reconstructed picture based on the prediction sample and the residual sample;

selecting a convolution neural network (CNN) for the reconstructed picture from among a plurality of CNNs;

performing a filtering on the reconstructed picture based on the selected CNN; and

encoding information about a post filtering, residual information and an index indicating the selected CNN among the plurality of CNNs,

wherein performing the filtering on the reconstructed picture comprises:

splitting the reconstructed picture into input blocks having a specific size,

deriving features by performing a convolution operation for each of the input blocks based on a kernel for the selected CNN,

determining a filter set based on the features, and

performing a filtering on the reconstructed picture based on the filter set.

8. The image encoding method of claim 7 , wherein the sizes of the input blocks are 3×3 sizes or 5×5 sizes.

9. The image encoding method of claim 7 , wherein the information about the post filtering comprises: a flag indicating whether a post filtering is performed and a flag indicating whether the CNN-based filtering is performed.