Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Transposed Convolution (Conv2DTranspose)
Learned operation that performs spatial upsampling of feature maps, enabling the reconstruction of high-resolution segmentation from low-resolution representations.
Skip Connection
Direct connection between layers of different hierarchical levels in an FCN, merging high-resolution semantic information with low-resolution contextual features.
Interpolation Pooling
Non-learnable upsampling technique using methods like bilinear or nearest neighbor to increase the spatial resolution of feature maps.
Pixel-wise Cross-Entropy Loss
Cost function calculating the cross-entropy between class predictions and ground truth labels for each individual pixel of the image.
Subsampling Factor
Ratio of spatial resolution reduction between the input and output of a pooling layer, typically 2 or 4, affecting the granularity of the final segmentation.
Effective Receptive Field
Area of the input image influencing the activation of an output pixel in an FCN, progressively widening with network depth to capture global context.
Feature Fusion
Process of combining feature maps from different spatial scales, typically through addition or concatenation, to improve segmentation accuracy.
Spatial Receptivity
Ability of an FCN to preserve precise localization information while capturing abstract semantic features through its hierarchical architecture.
Spatial Unpooling
Inverse operation of pooling that redistributes aggregated values to their original spatial positions, often combined with convolutions to refine segmentation.
Atrous Convolution
Modification of standard convolution by inserting spaces between kernel weights, increasing the receptive field without reducing the spatial resolution of feature maps.
Spatial Alignment
Technique ensuring precise correspondence between pixel positions in feature maps of different scales during skip connection fusion.
Segmentation Gradient
Backpropagation flow calculating how each weight in the FCN influences pixel-by-pixel segmentation error, essential for end-to-end optimization.
Spatial Regularization
Constraint added to the loss function to encourage spatial consistency of predictions, reducing noise and artifacts in segmentation maps.
Pixel Sampling
Strategy for selecting a subset of pixels for loss calculation, balancing rare classes and accelerating FCN training on high-resolution images.
Semantic Context
Relational information between pixels captured by deep layers of an FCN, enabling consistent predictions based on global scene understanding.
Classification Pretraining
Strategy for initializing FCN weights with a network pretrained on image classification, accelerating convergence and improving generalization.