Over 150 terms used in machine vision, computer vision, and automated defect detection. Use this reference to understand technical specifications, compare systems, and communicate with vendors.
Aberration
Optical imperfection causing image distortion. Common types include chromatic aberration (colour fringing) and spherical aberration (softness at edges). Higher-quality lenses minimise aberrations.
Accuracy
How close a measurement is to the true value. In defect detection, accuracy refers to the percentage of correct classifications (true positives + true negatives) divided by total inspections.
AI (Artificial Intelligence)
Computer systems that perform tasks requiring human-like intelligence. In machine vision, AI typically refers to deep learning models that learn to recognise defects from training images.
Algorithm
A step-by-step procedure for solving a problem. Vision algorithms include edge detection, blob analysis, pattern matching, and neural network inference.
Annotation
The process of labelling training images to teach AI models. Includes drawing bounding boxes around defects, marking regions, or assigning class labels to images.
Anomaly Detection
The process of training a model only on good parts (unsupervised learning) so it flags any deviation as a defect. Crucial when defective parts are too rare to train a standard classification model. The model learns what "normal" looks like, then identifies anything that doesn't match.
Aperture
The opening in a lens that controls how much light reaches the sensor. Measured in f-stops (f/1.4, f/2.8, f/8, etc.). Smaller f-number = larger opening = more light but shallower depth of field.
Area Scan Camera
Camera with a 2D sensor that captures complete frames. Standard for most machine vision. Compare with line scan cameras which capture one row at a time.
Aspect Ratio
The ratio of image width to height. Common ratios: 4:3 (standard), 16:9 (wide), 1:1 (square). Must match between sensor and display for correct proportions.
Backlight
Illumination placed behind the object, creating a silhouette. Ideal for dimensional measurement, edge detection, and inspecting transparent materials. Creates maximum contrast for size/shape analysis.
Barrel Distortion
Lens distortion where straight lines curve outward near image edges, making the image look "barrel-shaped". Common in wide-angle lenses. Can be corrected in software.
Bayer Pattern
The colour filter arrangement used on most colour sensors. Alternating pattern of red, green, and blue filters over pixels. Raw Bayer data must be interpolated (demosaiced) to produce colour images.
Binary Image
An image with only two values: black (0) and white (255). Created by thresholding a grayscale image. Simplifies analysis for presence/absence detection and blob analysis.
Bit Depth
The number of bits used to represent each pixel's intensity. 8-bit = 256 grey levels, 10-bit = 1024 levels, 12-bit = 4096 levels. Higher bit depth captures subtle variations but increases data volume.
Blob Analysis
A technique for finding and measuring connected regions of similar pixels. Used to detect and characterise defects, count objects, and measure areas. Also called connected component analysis.
Bounding Box
A rectangle that encloses an object or defect in an image. Defined by coordinates (x, y, width, height). Used in object detection to locate and classify items.
Calibration
The process of mapping pixel coordinates to real-world physical units (mm, inches) and correcting for lens distortion. Essential for accurate dimensional measurement of defects. Typically done using a calibration target with known dimensions.
C-Mount
A lens mounting standard with 1" diameter, 32 threads per inch. Flange distance of 17.526mm. Common on industrial cameras. Not compatible with CS-mount without adapter.
CS-Mount
A lens mounting standard similar to C-mount but with shorter flange distance (12.526mm). Common on smaller industrial cameras and CCTV. A C-mount lens can be used on a CS-mount camera with a 5mm spacer ring. (A CS-mount lens cannot be used on a C-mount camera.)
Camera Link
A high-speed camera interface standard. Supports up to 6.8 Gbps bandwidth. Used in high-end industrial vision systems requiring maximum frame rates.
Classification
Assigning an image or region to a category (e.g., "good", "defective", "scratch", "dent"). Can be binary (pass/fail) or multi-class. Deep learning excels at classification tasks.
CNN (Convolutional Neural Network)
A type of deep learning architecture designed for image processing. Uses convolution layers to learn visual features automatically. The foundation of modern AI-based defect detection.
Coaxial Illumination
Lighting where light travels along the camera's optical axis via a beam splitter. Creates even illumination on flat reflective surfaces. Defects appear dark against bright background.
Contrast
The difference in brightness between features in an image. High contrast makes defects easier to detect. Lighting technique significantly affects contrast.
CUDA
NVIDIA's parallel computing platform for GPUs. Enables accelerated AI inference and image processing. Required for running many deep learning frameworks efficiently.
Darkfield Illumination
Very low-angle lighting (typically <15° from surface). Smooth surfaces appear dark; surface defects like scratches scatter light toward camera and appear bright.
Data Augmentation
The technique of artificially expanding a training dataset by rotating, flipping, cropping, or slightly blurring images. Essential in defect detection where you often lack enough examples of defective parts. Helps models generalise better and reduces overfitting.
Deep Learning
A subset of machine learning using neural networks with multiple layers. Learns features automatically from data rather than requiring manual programming. State-of-the-art for complex defect detection.
Depth of Field (DOF)
The range of distances that appear acceptably sharp in an image. Affected by aperture (smaller = deeper DOF), focal length, and working distance. Critical for inspecting 3D objects.
Diffuse Illumination
Soft, scattered light that minimises shadows and reflections. Dome lights and cloudy-day panels provide diffuse illumination. Good for curved or reflective surfaces.
Dilation
A morphological operation that expands bright regions in a binary image. Used to fill small holes and connect nearby objects. Opposite of erosion.
Dome Light
A hemispherical enclosure with internal LEDs providing diffuse illumination from all angles. Eliminates shadows and reflections on curved, shiny objects.
Edge Detection
Finding boundaries between regions in an image. Common algorithms include Canny, Sobel, and Laplacian. Fundamental operation for shape analysis and measurement.
Encoder
A sensor that measures position or motion. In line scan applications, encoders synchronise camera triggering with conveyor movement to prevent image stretching.
Erosion
A morphological operation that shrinks bright regions in a binary image. Used to remove small noise spots and separate touching objects. Opposite of dilation.
Exposure Time
Duration the camera sensor collects light for each frame. Longer exposure = more light but more motion blur. Measured in milliseconds or microseconds.
F-Number (f-stop)
Ratio of lens focal length to aperture diameter. Lower f-numbers (f/1.4) mean more light but shallower depth of field. Higher f-numbers (f/16) mean less light but deeper DOF.
F1-Score
The harmonic mean of precision and recall. Formula: 2 × (Precision × Recall) / (Precision + Recall). The best single metric for defect detection models because datasets are usually highly imbalanced (mostly good parts, few bad parts). Range 0-1, with 1 being perfect.
False Negative
When a defect exists but the system fails to detect it (missed defect). Critical to minimise in quality control applications.
False Positive
When the system incorrectly flags a good part as defective. Wastes product and reduces throughput. Must balance against false negative rate.
Faster R-CNN
A deep learning architecture for object detection. Two-stage detector: first proposes regions, then classifies them. More accurate but slower than single-stage detectors like YOLO.
Field of View (FOV)
The area visible to the camera. Determined by sensor size, lens focal length, and working distance. Measured in mm or degrees.
Focal Length
The distance from lens to sensor when focused at infinity. Shorter focal length = wider FOV. Longer = narrower FOV (telephoto). Measured in mm.
FPS (Frames Per Second)
The number of images a camera can capture per second. Higher FPS needed for fast-moving objects. Limited by sensor readout speed and interface bandwidth.
Gain
Electronic amplification of the camera signal. Increasing gain brightens the image but adds noise. Better to use more light than high gain when possible.
GigE Vision
A camera interface standard using Gigabit Ethernet. Long cable runs (100m), standard cabling, multi-camera capability. Industry standard for industrial vision.
Global Shutter
Sensor type where all pixels capture light simultaneously. Essential for imaging moving objects without distortion. Compare with rolling shutter.
GPU (Graphics Processing Unit)
Processor designed for parallel computation. Essential for accelerating deep learning inference and training. NVIDIA GPUs with CUDA are most common.
Grayscale
Image with only intensity information, no colour. Each pixel has a single value (typically 0-255). Simpler to process than colour. Most defect detection uses grayscale.
Ground Truth
The absolute, human-verified reality of what is in an image. Used as the reference for training and evaluating AI models. If a model's prediction doesn't match the ground truth, it's counted as an error. Quality of ground truth directly affects model accuracy.
HALCON
Commercial machine vision software from MVTec. Industry-leading library with comprehensive tools for inspection, measurement, and AI. Used in enterprise applications.
Histogram
A graph showing the distribution of pixel intensities in an image. Useful for setting thresholds, checking exposure, and image analysis.
HSV (Hue, Saturation, Value)
A colour space that separates colour (hue) from intensity (value). Often easier to work with than RGB for colour-based detection.
Image Acquisition
The process of capturing images from a camera. Includes triggering, exposure control, and transferring image data to the processing system.
Inference
Running a trained AI model on new images to make predictions. Inference speed (in FPS or ms per image) determines if real-time operation is possible.
IoU (Intersection over Union)
A metric for evaluating object detection accuracy. Measures overlap between predicted and ground truth bounding boxes. Higher IoU = better localisation.
Jetson
NVIDIA's embedded AI computing platform. Models include Nano, Xavier, and Orin. Popular for edge AI and real-time vision applications.
JPEG
A lossy image compression format. Small file sizes but loses quality. Avoid for machine vision; use lossless formats (PNG, TIFF, BMP) for inspection images.
Kernel
A small matrix used in image filtering operations (convolution). Common kernels: blur, sharpen, edge detect, emboss. Size affects processing area.
Keyence
Japanese manufacturer of industrial automation equipment including vision systems, sensors, and inspection equipment. Known for turnkey solutions.
Latency
The delay between capturing an image and receiving inspection results. Critical for real-time systems. Measured in milliseconds.
LED (Light Emitting Diode)
Solid-state light source used in most machine vision lighting. Advantages: instant on/off, long life, low heat, stable output, various wavelengths available.
Lens
Optical element that focuses light onto the camera sensor. Key parameters: focal length, aperture, mount type, and optical quality.
Line Scan Camera
Camera with a 1D sensor that captures one row of pixels at a time. Used for continuous web inspection (paper, fabric, film) where objects move past the camera.
mAP (mean Average Precision)
A metric for evaluating object detection models. Considers both precision and recall across different confidence thresholds. Higher mAP = better performance.
Machine Learning
Algorithms that improve through experience (training data). Includes traditional methods (SVM, random forests) and deep learning. Enables systems to learn from examples.
Megapixel (MP)
One million pixels. A 5MP sensor has approximately 5 million pixels (e.g., 2592×1944). Higher megapixels = more detail but more data to process.
Monochrome Camera
Camera that captures only intensity, no colour information. More sensitive than colour cameras (no Bayer filter light loss). Often preferred for machine vision.
Morphology
Image processing operations based on shapes. Includes erosion, dilation, opening, closing. Used to clean up binary images and extract shape features.
Neural Network
A computing system inspired by biological brains. Composed of layers of interconnected nodes that learn to recognise patterns. Foundation of deep learning.
Noise
Random variations in pixel values not related to the actual scene. Sources: sensor noise, electronic interference, compression. Reduces image quality and detection accuracy.
Nyquist Theorem
Sampling theorem stating you need at least 2 samples per feature to accurately capture it. In vision: need at least 2 pixels across the smallest defect to detect it. In practice, vision engineers aim for 3-4 pixels across a defect to guarantee reliable detection, accounting for sub-pixel blurring and edge effects.
Object Detection
Finding and locating objects within an image. Outputs bounding boxes and class labels. Different from classification (whole image) or segmentation (pixel-level).
OCR (Optical Character Recognition)
Reading text from images. Used for serial numbers, date codes, labels. Requires good image quality and appropriate lighting.
ONNX
Open Neural Network Exchange format. Allows trained models to be used across different frameworks (PyTorch → TensorRT). Industry standard for deployment.
OpenCV
Open-source computer vision library. Supports Python, C++, and other languages. Comprehensive tools for image processing, feature detection, and machine learning.
Pattern Matching
Finding instances of a template image within a larger image. Used for alignment, verification, and locating features. Can be correlation-based or edge-based.
Photometric Stereo
A technique using multiple images of the same object taken with different lighting directions. Software combines them to separate 3D surface textures (scratches, dents) from 2D colour variations (stains, prints). Powerful for surface defect detection where you need to distinguish shape from colour.
Pixel
The smallest unit of a digital image. Each pixel has position (x, y) and value(s). Monochrome: one intensity value. Colour: typically three values (R, G, B).
PLC (Programmable Logic Controller)
Industrial computer for automation control. Vision systems often interface with PLCs via digital I/O or industrial protocols to trigger rejects and control lines.
Polarizer / Polarised Light
Filters placed over the lens and/or light source to block light waves vibrating in certain directions. Essential for removing glare from reflective parts (metal, glass, plastic) to see the defects underneath. Cross-polarisation uses filters at 90° to each other for maximum glare reduction.
Precision
Of all items flagged as defective, what percentage actually are? High precision = few false positives. Formula: True Positives / (True Positives + False Positives).
Quantisation
Reducing neural network precision (e.g., 32-bit float to 8-bit integer) to speed up inference. Some accuracy loss but significant performance gain. INT8 is common.
QR Code
A 2D barcode that can store more data than 1D barcodes. Common in tracking and traceability. Machine vision systems often include QR reading capability.
Recall
Of all actual defects, what percentage did the system detect? High recall = few missed defects. Formula: True Positives / (True Positives + False Negatives).
Resolution
The level of detail in an image. Can refer to pixel count (e.g., 1920×1080) or spatial resolution (pixels per mm). Higher resolution captures smaller features.
Ring Light
Circular LED array mounted around the camera lens. Provides even, shadow-free illumination. Good general-purpose lighting for flat, matte surfaces.
ROI (Region of Interest)
A defined area within an image for focused processing. Reduces computation by ignoring irrelevant areas. Can be rectangular, circular, or arbitrary shape.
Rolling Shutter
Sensor type where rows are captured sequentially (top to bottom). Moving objects appear skewed or distorted. Cheaper than global shutter but problematic for moving targets.
Segmentation
Classifying each pixel in an image. Semantic segmentation: pixel classes. Instance segmentation: separate objects. More detailed than bounding boxes.
Sensor
The light-sensitive chip in a camera. Common types: CCD (older, high quality) and CMOS (modern, faster, cheaper). Converts light to electrical signals.
Shutter
Controls when the sensor is exposed to light. Electronic shutter (no moving parts) is standard. Global shutter exposes all pixels simultaneously; rolling shutter is sequential.
SNR (Signal-to-Noise Ratio)
Ratio of useful signal to noise. Higher SNR = cleaner images. Affected by lighting, sensor quality, and gain settings.
SSD (Single Shot Detector)
A deep learning architecture for real-time object detection. Faster than Faster R-CNN but generally less accurate. Good for embedded applications.
Strobe
Pulsed lighting synchronised with camera capture. Provides much higher peak intensity than continuous light. Essential for high-speed inspection to freeze motion.
Structured Light
Projected patterns (lines, grids) used for 3D measurement. Pattern deformation reveals surface height. Used for weld inspection, volume measurement, and profile analysis.
Telecentric Lens
A lens designed to eliminate perspective error (parallax). Objects appear the same size regardless of their distance from the lens. Mandatory for accurate dimensional measurement of defects where parts may vary in height. More expensive than standard lenses but essential for metrology applications.
Template Matching
Finding a known pattern within an image. The template is slid across the image, computing similarity at each position. Used for alignment and verification.
TensorRT
NVIDIA's library for optimising and running deep learning inference on GPUs. Dramatically speeds up models through optimisation and precision reduction.
Threshold
A value used to separate pixels into categories. In binary thresholding, pixels above threshold become white, below become black. Simple but effective technique.
Throughput
The number of items inspected per unit time. Affected by cycle time, camera frame rate, and processing speed. Key metric for production systems.
Transmitted Illumination
Light passing through the object (backlight). Creates silhouettes. Reveals internal defects in transparent materials. Ideal for dimensional measurement.
Trigger
Signal that tells the camera when to capture an image. Can be software trigger (command) or hardware trigger (electrical signal from sensor or PLC).
Transfer Learning
Taking a model pre-trained on a massive dataset (like ImageNet) and fine-tuning it on your specific defect images. Saves hundreds of hours of GPU training time and works well even with limited defect samples. The standard approach for deploying deep learning in industrial inspection.
USB3 Vision
Camera interface standard using USB 3.0. Up to 5 Gbps bandwidth. Plug-and-play, common cabling. Popular for desktop development and lighter industrial use.
UV (Ultraviolet)
Light wavelengths shorter than visible (10-400nm). UV illumination causes some materials to fluoresce, revealing otherwise invisible features like cracks or contamination.
Validation
Testing a vision system on data it wasn't trained on to ensure it generalises. Critical for ensuring real-world performance matches development results.
Vignetting
Darkening at the corners of an image due to lens limitations. More pronounced at wide apertures. Can be corrected in software or with higher-quality lenses.
Working Distance
Distance from the front of the lens to the object being imaged. Affects field of view and depth of field. Must allow room for lighting and part handling.
X-Ray Inspection
Using X-rays to see inside objects. Reveals internal defects, assembly errors, and foreign objects. Common in electronics, food, and pharmaceutical inspection.
YOLO (You Only Look Once)
A family of real-time object detection models. Single-stage detector that's very fast. Recent versions like YOLOv8 through YOLOv11 dominate industrial applications, with each iteration bringing improvements to edge deployment and parameter efficiency. Popular for defect detection due to speed.
Zoom Lens
A lens with adjustable focal length. Offers flexibility but typically lower optical quality than fixed (prime) lenses. Variable FOV without changing lenses.
Related Resources
Missing a term? Contact us and we’ll add it.