Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Geometric rotation
Spatial transformation applied to images that performs an angular rotation around a reference point, generally the center of the image, to generate variations in object orientation.
Spatial translation
Horizontal or vertical movement of pixels in an image without altering their orientation, allowing to simulate different positions of objects in the frame.
Homothetic scaling
Proportional modification of image dimensions by a uniform scale factor, preserving aspect ratios while changing the apparent size of objects.
Shearing transformation
Geometric deformation that tilts the coordinate axes, creating a shearing effect that simulates oblique perspectives of objects in the image.
Mirror flipping
Horizontal or vertical inversion of the image creating a mirror symmetry, essential for eliminating directional biases in machine learning.
Random cropping
Random extraction of sub-regions of an image at different positions and scales, increasing cropping diversity and improving model robustness.
Elastic deformation
Complex non-linear transformation applying variable local displacements to pixels, simulating organic and natural deformations of image structures.
Perspective transformation
Geometric projection modifying the image perspective by simulating different viewpoints, crucial for learning perspective invariance.
Affine transformation
Linear combination of transformations preserving parallelism and distance ratios, including rotation, translation, scaling, and shearing.
Homography
Projective transformation between two planes preserving straight lines but not necessarily angles, used for complex viewpoint changes.
Elastic warping
Fluid and continuous deformation of image structure based on displacement fields, simulating natural morphological variations of objects.
Random zoom
Variable enlargement or reduction of image portions simulating different observation distances, essential for multi-scale learning.
Geometric padding
Extension of image borders through specific methods (reflect, wrap, constant) before transformation to avoid artifacts and information loss.
Bilinear interpolation
Pixel resampling method using the weighted average of the four nearest neighbors to estimate values after geometric transformation.
Spatial Transformer Network
Learnable neural module dynamically applying geometric transformations to feature maps to normalize spatial representations.
Geometric invariance
Property of a model to produce stable predictions in response to geometric transformations of input data, fundamental for generalization.
Coordinate transformation matrix
2D/3D mathematical representation encoding geometric transformation parameters to systematically apply spatial deformations.
Bounding box adaptation
Automatic adjustment of annotation bounding boxes concurrent with geometric transformations to maintain label consistency.
Grid sampling
Regular grid sampling technique to apply variable local transformations, enabling complex and controlled geometric deformations.
Rigid transformation
Combination of rotation and translation preserving distances and angles between points, used to simulate position changes without deformation.