AI-ordlista
Den kompletta ordlistan över AI
Skip Connections
Direct links that bypass one or more network layers, allowing the preservation of fine spatial information in the encoder and transferring it to the decoder.
Mirror Padding
Border padding technique using pixel reflection, used in U-Net to prevent information loss at edges during convolutions without padding.
Dice Loss
Cost function based on the Dice coefficient, optimized for imbalanced segmentation problems by measuring the overlap between prediction and ground truth.
Multi-class Segmentation
Extension of U-Net to simultaneously predict multiple object classes, using an output layer with softmax activation and one-hot encoding of labels.
U-Net++ (Nested U-Net)
Improved variant of U-Net introducing nested skip connections and dimension reduction blocks, reducing vanishing gradient and improving feature fusion.
Attention U-Net
U-Net architecture augmented with attention mechanisms that weight skip connection features, allowing the model to focus on the most relevant structures.
Intersection over Union (IoU) Score
Evaluation metric measuring the ratio between the intersection and union of predicted and actual segmentation masks, standard for evaluating U-Net performance.
Convolutional Block
Basic unit of U-Net composed of two successive 3x3 convolutions with ReLU activation and batch normalization, followed by a pooling operation in the encoder.