Image segmentation method, apparatus, and computer program to identify objects in images

An image segmentation apparatus comprising:

• an encoding module configured to obtain a plurality of feature maps having different resolutions with respect to an input image, using an artificial neural network including one or more residual blocks;
• and
• a decoding module configured to generate one prediction map using pairs of adjacent feature maps among the plurality of feature maps,
• wherein
• the decoding module performs an initial decoding round and one or more additional decoding rounds using an attention-embedded threefold fission (“ATP”) module,
• in the initial decoding round, one or more first combined feature maps are generated by combining an adjacent feature map pair selected from the plurality of feature mans, the adjacent feature map pair comprising a high-resolution feature map and an upsampled low-resolution feature map obtained by upsampling a low-resolution feature map having a lower resolution than the high-resolution feature map;
• each additional decoding round generates one or more additional combined feature maps by combining one or more combined feature map pairs generated in a previous round in a bracket-shaped manner;
• and
• the one or more additional decoding rounds are repeatedly performed until the one prediction map is generated.

An image segmentation program that is stored in a non-transitory computer-readable recording medium, wherein when the image segmentation program is executed by a processor, the processor is configured to:

• obtain a plurality of feature maps having different resolutions with respect to an input image, using a feed-forward artificial neural network;
• generate, in an initial round, one or more first combined feature maps, each first combined feature map generated by combining an adjacent feature map pair selected from the plurality of feature maps, the adjacent feature map pair comprising a high-resolution feature map and an upsampled low-resolution feature map obtained by upsampling a low-resolution feature map having a lower resolution than the high-resolution feature map;
• repeatedly perform, in one or more additional rounds, the generating of one or more additional combined feature maps by combining one or more combined feature map pairs generated in a previous round in a bracket-shaped manner until on prediction map is calculated as an output from the generation of the one or more first combined feature maps and the one or more additional combined feature maps;
• and
• classify one or more objects included in the input image using the prediction map.