Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Atrous Convolutions (or Dilated Convolutions)
Convolution operation that inserts spaces between kernel pixels, thereby increasing the receptive field without increasing the number of parameters, essential for capturing global context.
PSPNet (Pyramid Scene Parsing Network)
Model that applies pyramid pooling on different regions of the feature map to aggregate global context, then merges this information with local features for fine prediction.
Pyramid Pooling
Technique that applies pooling operations with different window sizes on the same feature map to capture contextual information at multiple scales.
Multi-branch Network
Architecture where multiple parallel paths process the input at different resolutions before their outputs are merged, enabling simultaneous analysis of detail and context.
Upsampling
Operation that increases the spatial resolution of a feature map, typically via deconvolution (transposed convolution) or bilinear interpolation, in the decoder part of the network.
Multi-scale Attention Map
Mechanism that learns to weight the importance of features from different scales, allowing the network to dynamically focus on the most relevant information for each pixel.
HRNet (High-Resolution Network)
Innovative architecture that maintains high-resolution representations throughout the network and repeatedly exchanges information with low-resolution representations for robust multi-scale segmentation.
Aggregated Context Module
Block designed to efficiently capture global context by aggregating information from large regions of the image, often used in complement to local paths for fine segmentation.