Glossary of Terms
A
Accuracy
Percentage of correct predictions over the total predictions made. Main metric to evaluate the performance of classification models.
AI (Artificial Intelligence)
Technology that enables machines to simulate human cognitive capabilities such as learning and decision making.
Anomaly Detection
Type of model that identifies anomalous or unusual patterns by comparing them with "normal" examples learned during training.
Annotation
Process of marking or labeling images with information that allows the AI model to learn. Includes classification, bounding boxes, segmentation, etc.
API (Application Programming Interface)
Interface that allows communication between different software. Rosepetal exposes APIs to make predictions and integrate with other systems.
Augmentation (Data Augmentation)
Technique that generates variations of original images (rotation, zoom, noise) to artificially increase the dataset size.
B
Backbone
Base architecture of a neural network (like ResNet, EfficientNet) that is used as foundation to build specific models.
Batch Size
Number of images processed simultaneously during training. Higher values require more memory but can accelerate the process.
Bounding Box
Rectangle that delimits an object in an image, defined by x,y coordinates of the top-left point and width/height.
C
Classification
Type of model that assigns a category or class to a complete image. Example: "good" vs "defective".
Confidence
Numerical value (0-100%) that indicates how sure the model is of its prediction. Higher confidence indicates greater certainty.
Controller
Rosepetal module that manages communication with devices, system monitoring and diagnostic tools.
Convolutional Neural Network (CNN)
Type of neural network specialized in image processing, that uses convolutional filters to detect features.
D
Dataset
Organized collection of labeled images used to train artificial intelligence models.
Data Augmentation
See "Augmentation". Techniques to create variations of existing data and improve model robustness.
Deploy
Process of putting a trained model or workflow into operation for production use.
Detection
Type of model that locates and classifies multiple objects within an image, providing location coordinates.
E
Epoch
One complete pass through the entire dataset during training. More epochs allow more learning but also more risk of overfitting.
Export
Function that allows downloading datasets, models or configurations in standard formats for external use.
F
F1-Score
Metric that combines precision and recall, useful for evaluating models with imbalanced classes. Value between 0 and 1, where 1 is perfect.
False Negative
Error where the model doesn't detect something that is present (e.g.: not detecting a defect that exists).
False Positive
Error where the model detects something that isn't present (e.g.: detecting a defect where there is none).
Flow
Automated sequence of tasks created with Node-RED that can include image capture, AI analysis and actions based on results.
G
GPU (Graphics Processing Unit)
Specialized processor that significantly accelerates AI model training by processing multiple operations in parallel.
H
Hyperparameters
Model configurations that are set before training, such as learning rate, batch size, number of epochs, etc.
I
Inference
Process of using a trained model to make predictions on new images not seen during training.
IoU (Intersection over Union)
Metric that measures the overlap between model prediction and ground truth, used in detection and segmentation.
J
JSON
Lightweight and easy-to-read data exchange format, used for configurations and prediction results.
L
Labeling
Manual process of annotating images with information that the model will use to learn. It's fundamental for supervised training.
Learning Rate
Hyperparameter that controls how fast the model learns. Very high values can cause instability, very low ones make learning slow.
Loss Function
Mathematical function that measures how far model predictions are from correct values. Training seeks to minimize this function.
M
Machine Learning (ML)
Subfield of AI that enables systems to automatically learn and improve from data without being explicitly programmed.
mAP (mean Average Precision)
Standard metric for evaluating object detection models, which averages precision across different classes and IoU thresholds.
Model
Trained algorithm capable of making specific predictions (classification, detection, segmentation) on new images.
MQTT
Lightweight communication protocol used in IoT and industrial systems to exchange messages between devices.
N
Node-RED
Visual programming tool that allows creating workflows by connecting nodes that represent different functionalities.
Neural Network
Computational model inspired by human brain functioning, composed of interconnected layers of artificial neurons.
O
Object Detection
Technique that identifies and locates multiple objects of interest within an image, providing both classification and location.
ONNX
Open format for representing machine learning models, which allows interoperability between different frameworks.
Overfitting
Problem where the model memorizes training data but doesn't generalize well to new data. Detected when validation accuracy decreases.
P
Precision
Metric that measures what percentage of positive predictions were correct. Important when false positives are costly.
Prediction
Result that a trained model produces when processing a new image. Includes predicted class and confidence level.
Preprocessing
Transformations applied to images before training or prediction, such as resizing, normalization, etc.
Q
Quality Control
Application of AI models to inspect products and determine if they meet established quality standards.
Quantization
Optimization technique that reduces numerical precision of models to make them faster and use less memory.
R
Recall
Metric that measures what percentage of actual positive cases were correctly identified. Important when false negatives are costly.
ROI (Region of Interest)
Specific area of an image that is of interest for analysis, allows focusing processing on relevant zones.
Rosepetal
Visual inspection platform with artificial intelligence developed for industrial quality control applications.
S
Segmentation
Technique that delimits precise areas of interest pixel by pixel, creating masks that identify exactly which pixels belong to each class.
Supervised Learning
Type of machine learning that uses labeled data to train models that can make predictions on new data.
T
Tag
Classification or category assigned to an image during the annotation process. Fundamental basis for supervised training.
Transfer Learning
Technique that leverages a model pre-trained on a large dataset and adapts it for a specific task with less data.
Training
Process by which a machine learning algorithm learns patterns from labeled data to make future predictions.
U
Underfitting
Problem where the model is too simple to capture underlying patterns in the data, resulting in poor performance.
Unsupervised Learning
Type of machine learning that finds patterns in unlabeled data, used in anomaly detection.
V
Validation
Process of evaluating model performance during training using a separate dataset that the model hasn't seen.
Vision AI
Subfield of artificial intelligence focused on enabling machines to interpret and understand visual content.
VLM (Vision-Language Model)
AI model capable of understanding both images and text, used to generate automatic descriptions or answer questions about images.
W
Webhook
Communication method that allows sending automatic notifications to external systems when specific events occur.
Workflow
Automated sequence of tasks that defines how data is processed from capture to decision making.
X
XML
Markup format used to structure data, commonly used to export annotations and configurations.
Y
YOLO (You Only Look Once)
Popular architecture for real-time object detection that processes the complete image in a single pass.
Z
Zero-shot Learning
Ability of a model to recognize classes it didn't see during training, based on descriptions or shared characteristics.
Zoom
Functionality to zoom in/out in visualization and labeling tools to work with greater precision on image details.