AI 용어집
인공지능 완전 사전
Instance segmentation
Advanced technique that segments each object individually in an image, distinguishing different instances of the same semantic class.
Panoptic segmentation
Unified approach combining semantic segmentation and instance segmentation to provide a complete understanding of all pixels in the image.
U-Net
Convolutional neural network architecture with encoder-decoder structure and skip connections, optimized for medical and biological segmentation.
Superpixels
Groups of connected pixels sharing similar characteristics (color, texture), used as primitives to reduce computational complexity.
Watershed
Segmentation algorithm based on topography that treats the image as a topographic relief to identify watersheds and dividing lines.
DeepLab
Family of semantic segmentation models using atrous convolutions to capture multi-scale context without losing spatial resolution.
FCN (Fully Convolutional Network)
First fully convolutional deep learning architecture enabling pixel-by-pixel segmentation by replacing fully-connected layers with convolutions.
Active Contours
Parametric segmentation method using deformable curves that evolve under the influence of internal and external forces to delimit objects.
Region Growing
Iterative segmentation algorithm that grows regions from seed points by adding neighboring pixels that meet homogeneity criteria.
Graph Cut
Segmentation approach based on graph theory that formulates the problem as a cut energy minimization in a graph constructed from the image.
Thresholding
Binary segmentation technique that classifies pixels into two categories by comparing their intensity to one or more predefined thresholds.
Mean Shift
Non-parametric clustering algorithm that identifies modes in the distribution of pixels in color and spatial space for segmentation.
GrabCut
Interactive segmentation algorithm based on graph cut using Gaussian mixture models to model foreground and background.
CRF (Conditional Random Field)
Discriminative probabilistic model used to refine segmentation results by modeling spatial dependencies between neighboring pixels.
Atrous Convolution
Convolution operation with holes that increases the receptive field without increasing parameters, essential for semantic segmentation.
IoU (Intersection over Union)
Standard evaluation metric measuring the overlap between predicted segmentation masks and ground truths.