IP3 2022 – Machine Learning (LOT 13727)

This lot is generally related to a semantic image segmentation system to combine local information and global context in a balanced manner to accurately identify an object in an image. Disclosed is an image segmentation device consisting of an encoding module for obtaining multiple feature maps having different resolutions of an input image. The system uses an artificial neural network with residual blocks and a decoding module for generating a prediction map using pairs of adjacent feature maps. Also disclosed is a decoding module for performing decoding rounds using attention-embedded threefold fusion (ATF) modules to generate a combined feature map using a high-resolution feature map. The technology may be implemented in autonomous driving systems, augmented reality, 3D image segmentation systems, etc.