Image identification device, method for performing semantic segmentation, and storage medium

An image identification device comprising:

• an image acquisition unit configured to acquire an image;
• a feature value extraction unit configured to extract a plurality of feature values of the acquired image;
• a creation unit configured to create a feature map by performing a convolution on each of the plurality of feature values;
• and
• a weighted feature value generation unit configured to calculate a modified feature value by performing convolution on the feature map, aggregate contexts by performing global average pooling on the calculated modified feature value, generate an attention that is a weighting factor of each channel, multiply the generated attention by the feature map, perform augmentation and attenuation weighting on the plurality of feature maps, and thus generate a weighted feature value.

A method for performing semantic segmentation of an image using a neural network system, the method comprising:

• a process of inputting the image;
• a process of extracting a plurality of feature values of the acquired image;
• a process of creating a feature map for each of the plurality of feature values of the image;
• and
• a process of multiplying each of the feature maps by a weighting factor that is an arbitrary positive value indicating a degree of importance of a feature.

A computer-readable non-transitory storage medium storing a program causing a computer to:

• acquire an image;
• extract a plurality of feature values of the acquired image;
• create a feature map for each of the plurality of feature values of the image;
• and
• multiply each of the feature maps by a weighting factor that is an arbitrary positive value indicating a degree of importance of a feature.