AI 용어집
인공지능 완전 사전
Dilated Convolution (or Atrous Convolution)
Convolution operation that inserts spaces between kernel elements, allowing for an increased receptive field without increasing the number of parameters or computational complexity.
Dilation Rate
Parameter of dilated convolutions that defines the spacing between pixels in the convolution kernel, directly controlling the increase in receptive field.
Encoder-Decoder with Atrous Convolution
DeepLabv3+ architecture that combines a powerful encoder (e.g., ResNet) with a simple and efficient decoder to refine segmentation predictions while preserving edge details.
Output Stride
Ratio between the input image resolution and that of the final feature maps; DeepLab often uses an output stride of 16 or 8 to balance accuracy and speed.
Image-Level Features
Global features extracted from the entire image, often via global pooling, and integrated into DeepLab's ASPP module to improve context class classification.
DeepLabv1
First version of DeepLab that introduced the use of dilated convolutions in fully convolutional networks (FCN) for semantic segmentation, increasing the receptive field without loss of resolution.
DeepLabv2
DeepLab version that integrated the Atrous Spatial Pyramid Pooling (ASPP) module to effectively capture multi-scale context, becoming a benchmark for semantic segmentation.
DeepLabv3
Iteration that improved the ASPP module by adding global pooling and 1x1 convolutions, and explored the application of cascaded dilated convolutions for better performance.
DeepLabv3+
Latest major version that adds a simple yet effective decoder to the DeepLabv3 architecture, refining segmentation predictions by merging low-level and high-level features.
Fully Connected Conditional Random Field (CRF)
Post-processing method used in early versions of DeepLab to refine the edges of segmentation predictions by modeling relationships between all pixels in the image.
Bilinear Interpolation Upsampling
Technique used in DeepLab to restore the spatial resolution of feature maps after dilated convolutions, in order to produce a segmentation map of the same size as the input image.
Contextual Aggregation
Fundamental principle of DeepLab aimed at aggregating contextual information from different spatial scopes through dilated convolutions and the ASPP module for more robust prediction.