Методи підвищення точності та ефективності систем розпізнавання образів на основі згорткових нейронних мереж

Abstract

This work overviews the evolution of modern computer vision from handcrafted feature descriptors to deep hierarchical models that automatically learn complex visual representations. Convolutional neural networks construct multilevel feature hierarchies, complicating small object detection. Feature pyramid networks address scale variance by combining semantic richness with multiresolution maps. Fully convolutional networks with skip connections fuse deep semantics and details in segmentation, evaluated by mean Intersection over Union.