KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
U-Net
U-shaped convolutional neural network architecture specifically designed for semantic segmentation of biomedical images, using skip connections between the encoder and decoder.
Semantic segmentation
Computer vision task of classifying each pixel of an image into a semantic category, allowing detailed pixel-level understanding.
Encoder-decoder architecture
Fundamental structure of segmentation networks where the encoder extracts hierarchical features while the decoder reconstructs the segmentation at full resolution.
Skip connections
Direct connections between encoder and decoder layers that preserve fine spatial information to improve the accuracy of segmentation boundaries.
Transposed convolution
Operation that performs the approximate inverse of a standard convolution, used to increase the spatial resolution of feature maps in the decoder.
Effective receptive field
Region of the input image that influences the activation of a particular neuron, crucial for understanding spatial context in segmentation networks.
IoU (Intersection over Union)
Evaluation metric measuring the overlap between predicted segmentation and ground truth, calculated as the ratio of intersection over union.
Dice coefficient
Metric similar to IoU but more sensitive to small regions, calculated as twice the intersection divided by the sum of pixels in both regions.
Multi-scale feature maps
Representations extracted at different spatial resolutions that allow the network to capture both fine details and global context.
Reflection padding
Image border padding technique that reflects existing pixels, better preserving structures than zero padding in segmentation networks.
Dice loss function
Loss function based on the Dice coefficient, optimized directly to maximize overlap between prediction and ground truth.
Deep supervision
Training technique that adds intermediate losses at multiple levels of the decoder to accelerate convergence and improve gradient flow.
Attention gates
Mechanisms that learn to selectively weight features from skip connections based on their relevance for the segmentation task.
Patch-based training
Training strategy that divides large images into smaller patches to manage memory constraints and increase sample diversity.
Test time augmentation
Inference technique that applies multiple augmentations to the input and averages predictions to improve robustness of segmentation results.
Bottleneck layer
Intermediate layer with minimal resolution in U-Net that contains the most abstract features before the decoding phase.
Instance normalization
Normalization technique applied individually to each sample of the batch, particularly effective for segmentation tasks.
Mixed precision training
Simultaneous use of float16 and float32 precisions to accelerate the training of segmentation networks while preserving numerical stability.