mdbloice/Augmentor
Augmentor
Augmentor est une bibliothèque et un framework Python d'augmentation d'images conçus pour étendre les jeux de données de machine learning. Il fonctionne comme un outil de prétraitement qui génère des variations d'images synthétiques pour accroître la diversité des données et comme un streamer de données d'entraînement qui alimente les boucles de réseaux de neurones avec des images et étiquettes augmentées sans nécessiter de stockage disque intermédiaire.
Le framework maintient l'alignement spatial entre les images et leurs masques correspondants, ce qui est requis pour l'entraînement à la segmentation sémantique. Il prend en charge diverses transformations géométriques et au niveau des pixels, incluant les distorsions élastiques, les changements de perspective via inclinaison et déformation, la rotation, le cisaillement et l'effacement aléatoire de régions.
Le système inclut des capacités de stratégies de traitement par classe pour remédier au déséquilibre des données et utilise le multi-threading pour accélérer la génération parallèle de jeux de données augmentés. Il fournit également des utilitaires pour nettoyer et standardiser les fichiers images bruts lors de l'étape de prétraitement.
Features
- Computer Vision Dataset Expansion - Increases training set size and diversity by generating multiple synthetic variations of images.
- Image Augmentation - Applies geometric distortions and random erasures to images to improve vision model robustness.
- Transform Sequence Assemblers - Implements sequences of transformations to generate augmented data from a single input image.
- Image Data Preprocessing - Provides utilities to clean and standardize raw image files for use in machine learning processing pipelines.
- Image Dataset Expansion Frameworks - Provides a framework to expand image datasets using synchronized transformations for images and masks.
- ML Training Data Streamers - Streams augmented images and labels directly into training loops to avoid storing intermediate files.
- On-the-Fly Training Transformations - Streams augmented images and labels directly into training loops to avoid writing intermediate files to disk.
- Synchronized Label Augmentations - Applies identical geometric transformations to images and labels to maintain spatial alignment.
- Image Augmentations - A Python library for generating synthetic image variations to increase training dataset diversity.
- Training Sample Streaming - Streams augmented image batches and labels directly into neural network training loops.
- Geometric Mask Synchronization - Ensures that spatial coordinates and rotation parameters are applied identically to both images and their label masks.
- Class-Specific Augmentation Strategies - Allows defining unique processing pipelines for different object categories to handle data imbalances.
- Per-Class Processing Configurations - Provides a configuration system for unique processing pipelines tailored to different object categories.
- Computer Vision Preprocessing - Provides utilities for cleaning and standardizing raw image files to prepare them for training.
- Region Erasers - Implements random region erasing to obscure portions of images, improving model robustness against occlusions.
- Semantic Segmentation Training - Maintains spatial alignment between images and masks during transformations for pixel-level labeling.
- Parallel Image Generation - Uses multi-threading to distribute image processing tasks across CPU cores for faster dataset generation.
- Rotation and Shear transforms - Provides capabilities to rotate and shear images while cropping results to maintain consistent dimensions.
- Coordinate-Based Spatial Distortions - Provides spatial distortions like tilt, skew, and elastic warping to simulate various camera perspectives.
- Elastic Distortion Augmentations - Implements elastic distortions to create realistic image variations while preserving labels and aspect ratios.
- Perspective Warping - Provides geometric transformations to alter image viewpoints and simulate different camera angles through tilt and skew.
- Parallel Image Transformers - Uses multi-threading to accelerate the parallel generation and saving of augmented image datasets across CPU cores.
- Vision par ordinateur - Image augmentation library.
- Data Wrangling - User-friendly tool for image augmentation in classification tasks.
- Outils de développement - Python library for image augmentation in machine learning pipelines.
