Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
YOLO (You Only Look Once)
Real-time detection system processing the image in a single pass for optimized speed.
R-CNN (Region-based Convolutional Neural Networks)
Two-stage architecture first proposing regions of interest then classifying them.
SSD (Single Shot MultiBox Detector)
Single-shot detector using multi-scale feature layers to detect objects of different sizes.
Faster R-CNN
Improvement of R-CNN integrating a region proposal network for faster detection.
Mask R-CNN
Faster R-CNN extension adding a segmentation branch for precise object masks.
RetinaNet
Architecture introducing focal loss to balance classes and improve detection of rare objects.
DETR (DEtection TRansformer)
Transformer-based approach eliminating manual components such as anchors and NMS.
EfficientDet
Family of models optimizing the accuracy/efficiency trade-off via compound scaling.
Cascade R-CNN
Cascade architecture training detectors with increasing IoUs for improved accuracy.
CornerNet
Keypoint-based detector localizing the top-left and bottom-right corners of objects.
FCOS (Fully Convolutional One-Stage)
Anchor-free detector directly predicting bounding boxes from pixels on feature maps.
CenterNet
Keypoint-based detector identifying object centers and their dimensions.
YOLOv5 and variants
Modern evolution of YOLO with optimized architecture and pre-training on large datasets.
Vision Transformers for Detection
Application of transformers to vision for detection without traditional convolutions.