Method and apparatus for recognizing image
An image recognition method using a region-based convolutional neural network (R-CNN) includes generating a feature map from an input image, detecting one or more regions of interest (ROIs) in the feature map, classifying the ROIs into groups based on setting information, performing pooling on the ROIs classified into the groups independently for each of the groups, and performing a regression operation on a result of the pooling and applying an image classifier to a result of the regression operation.
1. An image recognition method using a region-based convolutional neural network (R-CNN), the image recognition method comprising:
generating a feature map from an input image;
detecting one or more regions of interest (ROIs) in the feature map;
classifying the one or more ROIs into groups based on aspect ratios of the one or more ROIs;
performing pooling on the one or more ROIs classified into the groups for each of the groups; and
performing a regression operation on a result of the pooling, and applying an image classifier to a result of the regression operation,
wherein the classifying comprises:
classifying an ROI of the one or more ROIs into a group having a representative aspect ratio closest to an aspect ratio of the one or more ROIs.
2. The image recognition method of claim 1 , wherein the performing of the pooling comprises:
setting divided regions in each of the one or more ROIs classified into the groups,
wherein a ratio between a number of divided regions in an ROI of the one or more ROIs in a horizontal direction and a number of divided regions in the ROI of the one or more ROIs in a vertical direction is equal to a representative aspect ratio of a group comprising the ROI.
3. The image recognition method of claim 1 , wherein the performing of the pooling comprises:
setting divided regions in each of the one or more ROIs classified into the groups,
wherein a difference between a number of divided region in any two ROI of the one or more ROIs is less than a threshold value.
4. The image recognition method of claim 1 , wherein the aspect ratio comprises a ratio between a horizontal length and a vertical length of each of the one or more ROIs.
5. The image recognition method of claim 1 , wherein the detecting of the one or more ROIs comprises detecting the one or more ROIs based on performing region proposal.
6. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the image recognition method of claim 1 .
7. An image recognition apparatus comprising:
a processor configured to:
generate a feature map from an input image;
detect one or more regions of interest (ROIs) in the feature map;
classify the one or more ROIs into groups based on aspect ratios of the one or more ROIs;
perform pooling on the one or more ROIs classified into the groups for each of the groups; and
perform a regression operation on a result of the pooling, and apply an image classifier to a result of the regression operation,
wherein the processor classifying the one or more ROIs is configured to:
classify an ROI of the one or more ROIs into a group having a representative aspect ratio closest to an aspect ratio of the one or more ROIs.
8. The image recognition apparatus of claim 7 , wherein the processor is further configured to:
set divided regions in each of the one or more ROIs classified into the groups,
wherein a ratio between a number of divided regions in an ROI of the one or more ROIs in a horizontal direction and a number of divided regions in the ROI of the one or more ROIs in a vertical direction is equal to a representative aspect ratio of a group comprising the ROI.
9. The image recognition apparatus of claim 7 , wherein the processor is further configured to:
set divided regions in each of the one or more ROIs classified into the groups,
wherein a difference between a number of divided region in any two ROI of the one or more ROIs is less than a threshold value.