Computer vision

A convolution whose kernel covers a single pixel but all input channels. It mixes information across channels at each location, and is used to cut or raise the channel count cheaply, as in bottleneck blocks and Inception modules.

Example. A 1 x 1 convolution can compress a 56 x 56 x 256 feature map to 56 x 56 x 64 before an expensive 3 x 3 convolution.

Related: Bottleneck block, Inception module, Channel

A geometric change of an image that keeps straight lines straight and parallel lines parallel, covering rotation, scaling, shearing and translation. It is expressed as a 2 x 3 matrix applied to pixel coordinates.

Example. Rotating a scanned form by 5 degrees to fix a slight tilt is an affine transformation.

Related: Interpolation, Homography, Data augmentation

The deep CNN that won the 2012 ImageNet competition by a large margin, using ReLU activations, dropout and GPU training. Its success started the modern deep learning wave in computer vision.

Example. AlexNet has eight learned layers and was trained on two GPUs, a large jump in depth and scale at the time.

Related: ImageNet, VGG, Convolutional neural network

A predefined reference box of a chosen size and aspect ratio placed at many positions on the image. Detectors predict, for each anchor, whether it holds an object and how to adjust it, which simplifies handling many object shapes.

Example. At each feature-map cell, a detector might use 9 anchors: three scales times three aspect ratios.

Related: Bounding box regression, Region proposal network, Aspect ratio

The ratio of an image or box's width to its height. Preserving it during resizing avoids distorting shapes, and detectors use several aspect ratios for their reference boxes to match tall people or wide cars.

Example. A 1920 x 1080 frame has an aspect ratio of 16:9, about 1.78, while a standing person's box may be 1:3.

Related: Image resizing, Anchor box, Bounding box

A pooling method that outputs the mean of the values in each window. It smooths features rather than picking the strongest, and is often applied to the whole map at the end of a network to summarise it.

Example. For the 2 x 2 block [[1, 3], [2, 9]], average pooling outputs (1 + 3 + 2 + 9) / 4 = 3.75.

Related: Pooling layer, Max pooling, Global average pooling

An image-sized grid of ones and zeros marking which pixels belong to a region, with 1 for inside and 0 for outside. Segmentation models output masks, and masks can be used to cut out objects.

Example. A mask of a tumour is 1 on the tumour's pixels and 0 elsewhere, letting software measure its area by counting the ones.

Related: Image segmentation, Thresholding, Dice coefficient

A residual block that shrinks channels with a 1 x 1 convolution, does a 3 x 3 convolution at the smaller width, then expands with another 1 x 1 convolution. It gives deep networks like ResNet-50 lower cost per layer.

Example. A ResNet-50 block might go 256 to 64 channels, apply a 3 x 3 convolution, then go back to 256.

Related: 1x1 convolution, ResNet, Residual connection

A rectangle that tightly encloses an object, described either by two corners (x1, y1, x2, y2) or by a centre plus width and height. It is the standard output format of detectors.

Example. A box (50, 40, 150, 240) encloses an object 100 pixels wide and 200 pixels tall.

Related: Object detection, Intersection over union, Anchor box

A classic multi-step edge detection algorithm: blur the image, compute gradients, thin edges to one pixel by non-maximum suppression, then keep strong edges and weak edges connected to them using two thresholds (hysteresis). It gives clean, thin edge maps.

Example. With thresholds 100 and 200, pixels with gradient above 200 are kept as edges, and those between 100 and 200 are kept only if touching a strong edge.

Related: Edge detection, Sobel operator, Gaussian blur

One of the separate layers of values that together describe each pixel or feature map location. A colour photo normally has three channels (red, green, blue), while inner layers of a neural network can have hundreds of channels, each responding to a different learned pattern.

Example. An RGB image of shape 224 x 224 x 3 has 3 channels, and a convolutional layer with 64 filters outputs a feature map with 64 channels.

Related: RGB, Feature map, Grayscale image

A benchmark dataset of 60,000 small 32 x 32 colour images across 10 classes such as aeroplane, cat and truck, with 50,000 for training and 10,000 for testing. It is a quick test bed for new architectures and training tricks.

Example. Students compare a plain CNN against ResNet on CIFAR-10 because each run finishes in minutes on one GPU.

Related: MNIST, ImageNet, Image classification

An extra learnable token prepended to the patch sequence in a vision transformer, whose final output is used as a summary of the whole image for classification. It has no image content of its own; attention lets it gather information from every patch.

Example. After the last layer, the CLS vector goes through a linear layer to give 1,000 ImageNet class scores.

Related: Vision transformer, Patch embedding, Image classification

An augmentation that randomly changes an image's brightness, contrast, saturation and hue. It makes a model robust to different lighting and camera settings. Values are drawn at random each time, so the model rarely sees the same colours twice.

Example. Brightness jitter of plus or minus 20% mimics photos taken in shade versus sunlight.

Related: Data augmentation for images, HSV colour space, Contrast

A number between 0 and 1 attached to a prediction, showing how strongly the model believes it. Detections below a chosen threshold are discarded, so the threshold trades missed objects against false alarms.

Example. With a threshold of 0.5, a detection scored 0.42 is dropped and one scored 0.87 is kept.

Related: Object detection, Non-maximum suppression, Precision and recall for detection

CTC, a loss and decoding method for sequence labelling when input and output lengths differ and alignment is unknown. It lets a model output characters per image column and then merges repeats and blanks into the final text.

Example. The frame outputs 'h h - e - l l - l o' collapse to 'hello' after merging repeats and removing blanks.

Related: Text recognition, Optical character recognition, Sequence labelling

Finding the outlines of shapes in an image, usually from a binary or edge image, and returning each outline as a list of boundary points. It is used to measure shapes, count objects and find document borders.

Example. OpenCV's findContours can trace the four corners of a receipt photo so it can be straightened.

Related: Edge detection, Connected components, OpenCV

CLIP, a model trained on hundreds of millions of image-caption pairs to place matching images and texts near each other in a shared embedding space. It enables zero-shot classification by comparing an image with text prompts.

Example. To classify a photo, compare it with the texts 'a photo of a dosa' and 'a photo of an idli' and pick the closer one.

Related: Zero-shot learning, Image embedding, Contrastive learning

A training approach that pulls together embeddings of related pairs, such as two augmented views of one image, and pushes apart unrelated ones. It learns useful visual features without labels.

Example. Two random crops of the same photo are treated as a positive pair and crops of other photos as negatives.

Related: Self-supervised learning, Image embedding, Data augmentation for images

An operation that slides a small grid of numbers, called a kernel, across an image and at each position sums the products of the kernel values with the pixels beneath it. It detects local patterns such as edges and reuses the same weights everywhere, making it efficient for images.

Example. Sliding a 3 x 3 kernel over a 5 x 5 image with no padding and stride 1 gives a 3 x 3 output, each value a weighted sum of 9 pixels.

Related: Convolutional neural network, Stride, Padding

A neural network built mainly from convolution layers, pooling and a final classifier, designed for grid-like data such as images. Layers learn a hierarchy from edges to textures to object parts, and weight sharing keeps parameter counts manageable.

Example. A CNN that reads handwritten digits might use two convolution-and-pool stages followed by a small layer that outputs ten class scores.

Related: Convolution, Pooling layer, ResNet

The same sliding-window sum as convolution but without flipping the kernel first. Deep learning libraries actually implement this operation and call it convolution; since kernels are learned, the flip makes no practical difference.

Example. PyTorch's Conv2d computes cross-correlation, so a hand-designed edge kernel copied from a signal-processing textbook may produce a mirrored result.

Related: Convolution, Convolutional neural network, Kernel

An augmentation that pastes a rectangular patch from one image onto another and mixes the labels in proportion to the patch area. It keeps more useful pixels than Cutout while still regularising.

Example. If a pasted dog patch covers 25% of a cat photo, the label becomes 75% cat and 25% dog.

Related: Mixup, Cutout, Data augmentation for images

An augmentation that blanks out a random square patch of the input image. It forces the model not to depend on one region and improves robustness to occlusion. It acts as a cheap regulariser during training.

Example. A 32 x 32 patch of a 224 x 224 photo is set to zero so the model must recognise the object from the remaining parts.

Related: Mixup, CutMix, Data augmentation for images

Creating extra training examples by applying random changes such as flips, crops, rotations and colour shifts to existing images while keeping their labels. It reduces overfitting and teaches the model to ignore irrelevant variation.

Example. Randomly flipping and cropping each cat photo lets 5,000 images act like many times more distinct training examples.

Related: Random crop, Colour jitter, Overfitting

A family of semantic segmentation models that use dilated convolutions and atrous spatial pyramid pooling to capture context at several scales while keeping resolution high. It is a common baseline for scene parsing and satellite mapping.

Example. DeepLab can label pixels of buildings, roads and vegetation in satellite tiles.

Related: Dilated convolution, Semantic segmentation, Feature pyramid network

A factorised convolution that first filters each channel on its own with a spatial kernel and then mixes channels with a 1 x 1 convolution. It gives similar accuracy to a standard convolution with far fewer operations, and is central to mobile models.

Example. For 3 x 3 kernels with 128 input and 128 output channels, a standard convolution uses 147,456 weights while the separable version uses 1,152 + 16,384 = 17,536.

Related: MobileNet, 1x1 convolution, EfficientNet

An overlap score for masks: Dice = 2 x |A and B| / (|A| + |B|), ranging from 0 to 1. It is common in medical segmentation, and 1 minus Dice is often used as a loss.

Example. If a predicted mask has 80 pixels, the true mask 100 and they share 70, Dice = 140 / 180 = 0.78.

Related: Intersection over union, Binary mask, Semantic segmentation

A convolution whose kernel taps are spaced apart by gaps, enlarging the receptive field without extra parameters or lost resolution. It is popular in segmentation, where wide context is needed but the output must stay detailed.

Example. A 3 x 3 kernel with dilation 2 covers a 5 x 5 area while still using only 9 weights.

Related: Receptive field, Semantic segmentation, DeepLab

Identifying the structure of a page image, such as titles, paragraphs, tables, figures and columns, and the reading order. It lets OCR output be organised rather than a flat block of text.

Example. Analysis of a bank statement separates the header, the transaction table and the footer notes.

Related: Optical character recognition, Table recognition, Text detection

Finding locations where image brightness changes sharply, which usually mark object boundaries. Classical methods convolve the image with gradient kernels and threshold the result, and early CNN layers learn similar detectors on their own.

Example. Running an edge detector on a photo of a coin outputs a thin bright circle where the coin meets the table.

Related: Sobel operator, Canny edge detector, Image gradient

A family of CNNs that scales network depth, width and input resolution together using a fixed ratio, called compound scaling. It reached strong accuracy for its compute budget. It showed that balanced scaling beats growing only one dimension.

Example. EfficientNet-B0 is the small baseline, and larger variants B1 to B7 scale all three dimensions up.

Related: MobileNet, ResNet, Depthwise separable convolution

Finding the location of faces in an image, giving each a bounding box, without deciding whose face it is. It is used for camera autofocus and as a first step in face recognition.

Example. A phone camera draws a square around each face in a group photo.

Related: Face recognition, Object detection, Keypoint detection

An improved R-CNN that runs the CNN once on the whole image and then pools features for each proposed region using RoI pooling. Sharing the computation made training and inference much faster.

Example. One backbone pass over a photo yields a feature map from which 2,000 proposals are read out.

Related: R-CNN, Faster R-CNN, RoI pooling

A two-stage detector that adds a region proposal network sharing the backbone's features to generate proposals, then classifies and refines them. It is a strong, widely used baseline for accurate detection.

Example. Faster R-CNN is often chosen when accuracy matters more than speed, for example in analysing satellite images.

Related: Region proposal network, Fast R-CNN, Two-stage detector

A numeric vector that summarises the appearance around a keypoint so it can be compared with keypoints in other images. Good descriptors stay similar despite changes in scale, rotation or lighting.

Example. SIFT describes each keypoint with a 128-number vector built from local gradient directions.

Related: Feature detection, SIFT, ORB

Finding distinctive points in an image, such as corners and blobs, that can be located again reliably in other images. These keypoints support matching, tracking, mapping and stitching. Reliable keypoints are the starting point for matching and geometry.

Example. A Harris corner detector picks out the corners of window frames in a building photo as keypoints.

Related: Feature descriptor, Feature matching, SIFT

The output grid produced when a filter is applied across an image or an earlier layer's output. Each value shows how strongly the filter's pattern appears at that location, and a layer with many filters produces a stack of feature maps.

Example. A vertical-edge filter applied to a photo of a building gives a feature map with high values along window frames and low values on plain walls.

Related: Convolution, Channel, Receptive field

Pairing keypoints from two images by finding descriptors that are closest to each other, often with a ratio test to reject ambiguous pairs. The matches feed tasks like stitching, 3D reconstruction and object recognition.

Example. With Lowe's ratio test at 0.75, a match is kept only if the best distance is under 75% of the second-best distance.

Related: Feature descriptor, Homography, RANSAC

An architecture that builds feature maps at several scales, combining deep, semantically strong low-resolution features with shallow high-resolution ones through a top-down path. It improves detection of objects of different sizes.

Example. A distant, tiny pedestrian is detected on a fine-scale map while a nearby bus is found on a coarse one.

Related: Upsampling, Object detection, Multi-scale detection

Continuing to train a pretrained model on a smaller task-specific dataset, usually with a low learning rate, so its existing features adapt to the new problem. It needs far less data than training from scratch.

Example. A team fine-tunes an ImageNet ResNet on 2,000 labelled X-rays for 10 epochs at a learning rate of 0.0001.

Related: Transfer learning, Pretrained backbone, Feature extraction

A network with no fully connected layers, made only of convolutions, so it accepts any image size and outputs a spatial map. It made dense per-pixel prediction practical. It typically ends with upsampling so the output matches the input size.

Example. Feeding a 512 x 512 image gives a 512 x 512 class map; feeding a 256 x 256 one gives a 256 x 256 map.

Related: Semantic segmentation, Transposed convolution, Convolutional neural network

Smoothing an image by averaging each pixel with its neighbours using weights that follow a bell curve. It suppresses noise and fine detail, and is a standard preprocessing step before edge detection or when building image pyramids.

Example. A 5 x 5 Gaussian kernel with sigma 1 weights the centre pixel most and distant neighbours least, softening a grainy night-time photo.

Related: Convolution, Canny edge detector, Image noise

Averaging each entire feature map down to a single number, so a stack of C channels becomes a vector of length C. It replaces large fully connected layers at the end of a CNN, cutting parameters and letting the network accept varying image sizes.

Example. A final 7 x 7 x 512 feature map becomes a 512-value vector that feeds one small classification layer.

Related: Pooling layer, Convolutional neural network, Feature map

Gradient-weighted Class Activation Mapping, an explanation method that highlights image regions that most increased a CNN's score for a class, by weighting a late feature map with gradients. The result is a heatmap over the image.

Example. For a 'tumour' prediction on a scan, Grad-CAM shows a bright area over the suspicious region.

Related: Saliency map, Feature map, Explainability

An image with a single channel per pixel that records only brightness, from black through grey to white. Converting to grayscale cuts data to a third of RGB and is used when colour carries no useful information, such as in document scanning.

Example. A common conversion is gray = 0.299 R + 0.587 G + 0.114 B, so an orange pixel (255, 165, 0) becomes about 173.

Related: Channel, RGB, Thresholding

A contrast-boosting technique that redistributes pixel brightness values so they spread more evenly across the full range. It makes details in under- or over-exposed images easier to see. It is a common preprocessing step for dim medical or surveillance images.

Example. Applying it to an X-ray whose values sit between 80 and 140 stretches them across roughly 0 to 255, making faint structures visible.

Related: Image histogram, Contrast, Image preprocessing

A hand-designed feature that divides an image into small cells and records a histogram of gradient directions in each, normalised over larger blocks. It captures shape and was the basis of classic pedestrian detectors paired with a linear SVM.

Example. A 64 x 128 pedestrian window produces a feature vector of 3,780 numbers with the standard HOG settings.

Related: Image gradient, Support vector machine, Sliding window

A 3 x 3 matrix that maps points between two views of a flat surface, allowing perspective changes that affine maps cannot. It is used to straighten photographed documents and stitch panoramas.

Example. Given the four corners of a tilted cheque in a phone photo, a homography warps it into a flat, front-on rectangle.

Related: Affine transformation of images, Image stitching, Feature matching

An augmentation that mirrors an image left to right. It is safe for most natural scenes but wrong where orientation carries meaning, such as text or digits. It doubles the variety of many datasets at almost no cost.

Example. Flipping a photo of a dog keeps the label; flipping the digit 2 makes it look like a different symbol.

Related: Data augmentation for images, Random crop, Colour jitter

A way of describing colour by hue (the type of colour), saturation (how vivid it is) and value (how bright it is) instead of red-green-blue mixtures. Separating colour from brightness makes it easier to pick objects by colour under changing light.

Example. To find ripe red tomatoes on a conveyor belt, a script keeps pixels whose hue falls in the red range and saturation is high, ignoring shadow-driven brightness changes.

Related: RGB, Thresholding, Colour jitter

Assigning a label, or a set of labels, to a whole image, such as 'dog' or 'no defect'. It is the simplest vision task and the base for many pretrained backbones.

Example. A model looks at a photo of a rickshaw and outputs 'auto-rickshaw' with probability 0.93.

Related: Convolutional neural network, Top-k accuracy, ImageNet

A fixed-length vector representing an image such that visually or semantically similar images have nearby vectors. It powers image search, duplicate detection and recommendation. It is usually taken from a late layer of a neural network and compared by cosine similarity or distance.

Example. An e-commerce app embeds every product photo and, for a query photo of a kurta, returns the products with the closest vectors by cosine similarity.

Related: Feature extraction, Cosine similarity, Contrastive learning

The rate and direction of brightness change at each pixel, computed from differences between neighbouring pixels. Its magnitude highlights edges and its direction says which way they run; it is the base of edge detectors and descriptors like HOG.

Example. If the pixel to the right is 200 and the one to the left is 100, the horizontal gradient there is about (200 - 100) / 2 = 50.

Related: Edge detection, Sobel operator, Histogram of oriented gradients

A count of how many pixels take each brightness value, plotted as a bar chart. It reveals whether an image is too dark, too bright or low in contrast and guides adjustments like equalisation.

Example. A dark photo has most of its histogram bars bunched near 0 to 60 out of a 0 to 255 range.

Related: Histogram equalisation, Otsu's method, Grayscale image

Rescaling pixel values before feeding an image to a model, typically dividing by 255 and then subtracting a per-channel mean and dividing by a standard deviation. It puts inputs on a consistent scale so training is stable and pretrained weights receive the statistics they expect.

Example. With ImageNet statistics, a red-channel value of 0.5 becomes (0.5 - 0.485) / 0.229, which is about 0.066.

Related: Pixel, Transfer learning, ImageNet

A set of copies of one image at progressively smaller resolutions, each usually half the size of the last. Searching every level lets an algorithm find objects or features of different sizes with one fixed-size detector, and it underlies many classic matching methods.

Example. A 640 x 480 image yields levels of 320 x 240, 160 x 120 and 80 x 60; a face detector with a fixed 24 x 24 window finds large faces at the coarse levels.

Related: Sliding window, Feature pyramid network, Image resizing, Gaussian blur

Changing an image to a different width and height by interpolating pixel values. Vision pipelines resize because networks expect fixed-size inputs, and the choice of interpolation (nearest, bilinear, bicubic) and whether the aspect ratio is kept can change accuracy.

Example. Resizing a 640 x 480 photo to 224 x 224 without cropping squashes it horizontally, while resizing the short side to 224 and centre-cropping preserves shapes.

Related: Resolution, Aspect ratio, Interpolation

Dividing an image into regions by labelling pixels, so the outline of each object or area is known rather than just a box. It is used in medical imaging, autonomous driving and satellite analysis.

Example. A model colours every pixel of a road scene as road, car, pedestrian or sky.

Related: Semantic segmentation, Instance segmentation, Panoptic segmentation

A large labelled image dataset with over a million training photos in 1,000 categories, used in the ILSVRC competition. Networks pretrained on it became the standard starting point for many other vision tasks.

Example. A ResNet pretrained on ImageNet is loaded and its final layer swapped to classify 5 types of crop disease.

Related: Transfer learning, Image classification, Pretrained backbone

A building block that applies several convolutions of different sizes (1 x 1, 3 x 3, 5 x 5) and pooling in parallel and joins their outputs, letting the network capture patterns at multiple scales. 1 x 1 convolutions keep the cost down.

Example. GoogLeNet stacked Inception modules to reach strong accuracy with far fewer parameters than VGG.

Related: 1x1 convolution, VGG, Convolutional neural network

Labelling each pixel with its class and also with which individual object it belongs to, so two cars get separate masks. It combines detection with pixel-level outlines. It is needed when objects must be counted or handled one by one.

Example. A model outputs three separate masks for three people standing side by side.

Related: Semantic segmentation, Mask R-CNN, Panoptic segmentation

A measure of how much two boxes or masks overlap: IoU = area of overlap / area of union. It ranges from 0 (no overlap) to 1 (perfect match) and decides whether a detection counts as correct.

Example. Two 10 x 10 boxes overlapping in a 5 x 10 region have IoU = 50 / (100 + 100 - 50) = 0.33.

Related: Bounding box, Non-maximum suppression, Mean average precision

An extension of Faster R-CNN that adds a branch predicting a pixel mask for each detected object, using RoI Align for precise alignment. It is a standard instance segmentation model.

Example. Given a photo of a crowd, Mask R-CNN outputs a box, class and mask for each person.

Related: Faster R-CNN, Instance segmentation, RoI pooling

A pooling method that outputs the largest value inside each window of a feature map. It keeps the strongest detection of a pattern while discarding its exact position, giving mild translation invariance.

Example. For the 2 x 2 block [[1, 3], [2, 9]], max pooling outputs 9.

Related: Pooling layer, Average pooling, Translation invariance

A noise-removal filter that replaces each pixel with the median of the values in its neighbourhood. Unlike averaging, it removes isolated bright or dark specks while keeping edges sharp. It is especially good against isolated speckle noise.

Example. For the neighbourhood 10, 12, 11, 255, 9, the median is 11, so the 255 speck disappears.

Related: Image noise, Gaussian blur, Salt-and-pepper noise

An augmentation that blends two training images and their labels with a random weight, creating soft examples. It smooths decision boundaries and can improve generalisation and calibration. The blended images look unnatural to people but still help training.

Example. Mixing a cat image with weight 0.7 and a dog image with weight 0.3 gives the label [0.7 cat, 0.3 dog].

Related: CutMix, Cutout, Overfitting

A family of compact CNNs that use depthwise separable convolutions to run fast on phones and embedded devices with modest accuracy loss. Width and resolution multipliers let engineers trade accuracy for speed.

Example. A Bengaluru start-up runs a MobileNet on-device to check that an uploaded ID photo is not blurry before sending it to the server.

Related: Depthwise separable convolution, EfficientNet, Edge deployment

Image operations that reshape white regions in a binary image using a small pattern called a structuring element. Erosion shrinks regions, dilation grows them, opening removes small specks and closing fills small holes.

Example. Applying erosion then dilation (opening) to a thresholded image removes isolated white dots while keeping large blobs nearly unchanged.

Related: Thresholding, Binary mask, Connected components

A post-processing step that removes duplicate detections of the same object: keep the highest-scoring box, discard other boxes that overlap it by more than an IoU threshold, and repeat. It leaves one box per object.

Example. Five boxes cluster around one face; with threshold 0.5, NMS keeps the highest-confidence one and drops the rest.

Related: Intersection over union, Object detection, Confidence score

Finding every object of interest in an image and giving each a class label and a bounding box. Unlike classification, it says both what is present and where, and it handles several objects at once.

Example. A traffic camera model outputs 'car' at box (120, 80, 300, 220) with confidence 0.91 and 'motorcycle' at another box.

Related: Bounding box, Intersection over union, Mean average precision

A widely used open-source library for computer vision and image processing, with functions for reading images, filtering, geometric transforms, feature detection, video capture and classic algorithms. It is available for Python and C++.

Example. cv2.imread('bill.jpg') loads an image as a NumPy array, and cv2.cvtColor converts it to grayscale.

Related: Image resizing, Edge detection, Contour detection

Software that converts images of printed or handwritten text into machine-readable characters. It underlies digitising documents, reading number plates and extracting data from invoices. Accuracy depends on image quality, fonts, language and layout.

Example. An OCR tool reads a photo of a GST invoice and returns text such as 'Total: 4,720.00'.

Related: Text detection, Text recognition, Document layout analysis

Oriented FAST and Rotated BRIEF, a fast keypoint detector and binary descriptor that is free of patent constraints and cheap enough for real-time use on phones and robots. It compares descriptors using Hamming distance.

Example. A robot's camera uses ORB to track a few hundred keypoints per frame while it moves around a warehouse.

Related: SIFT, Feature matching, SLAM

An automatic way to choose a threshold by trying every cutoff on the brightness histogram and picking the one that best separates the two groups of pixels, by minimising the variance within groups. It works well when the histogram has two clear peaks.

Example. On a scanned page with dark ink on pale paper, Otsu's method finds a cutoff near the valley between the ink peak and the paper peak.

Related: Thresholding, Image histogram, Binary mask

Extra values, usually zeros, added around the border of an image or feature map before convolution. It lets the kernel be centred on edge pixels and controls output size; 'same' padding keeps the size unchanged while 'valid' uses none.

Example. A 3 x 3 kernel with padding 1 and stride 1 maps a 32 x 32 input to a 32 x 32 output; with no padding it gives 30 x 30.

Related: Convolution, Stride, Feature map

A task that unifies semantic and instance segmentation: every pixel gets a class, and pixels of countable things such as people get an instance identity, while stuff such as sky and road gets only a class.

Example. A driving scene is labelled with separate IDs for each car and pedestrian, plus single regions for road and sky.

Related: Semantic segmentation, Instance segmentation, Image segmentation

The first step of a vision transformer: cut the image into equal square patches, flatten each and project it with a learned linear layer into a vector, giving the tokens the transformer reads.

Example. A 16 x 16 x 3 patch has 768 raw values, which is projected to a 768-dimensional embedding in ViT-Base.

Related: Vision transformer, Position embedding, Image tokens

The smallest addressable element of a digital image, holding one or more numbers that describe brightness or colour at a single location. An image is a grid of pixels, so its resolution and file size grow with the number of pixels along its width and height.

Example. A 1920 x 1080 photo contains 1920 x 1080 = 2,073,600 pixels, and in an 8-bit RGB image each pixel stores three values between 0 and 255.

Related: Resolution, Channel, Grayscale image

A layer that shrinks a feature map by summarising each small window with one number, commonly the maximum or the average. It reduces computation, adds tolerance to small shifts of the object, and lets later layers see larger regions.

Example. A 2 x 2 max pool with stride 2 turns a 56 x 56 feature map into 28 x 28, keeping the strongest value in each block.

Related: Max pooling, Average pooling, Convolutional neural network

Detection versions of the standard metrics: a predicted box is a true positive if its IoU with a real object passes a threshold. Precision = TP / (TP + FP) and recall = TP / (TP + FN), so missed objects and false boxes are both counted.

Example. Of 10 predicted boxes, 8 match real objects and there are 12 real objects: precision = 0.8, recall = 8 / 12 = 0.67.

Related: Mean average precision, Intersection over union, Confidence score

A network already trained on a large dataset, whose early and middle layers serve as a general feature extractor for a new task. Detection and segmentation models attach task-specific heads to it.

Example. A Faster R-CNN detector uses an ImageNet-pretrained ResNet-50 as its backbone.

Related: Transfer learning, ImageNet, ResNet

Regions with CNN features, an early deep detector that generated region proposals, ran a CNN on each cropped region and classified it with SVMs. It was accurate but very slow because each region was processed separately.

Example. R-CNN would push about 2,000 cropped regions through a CNN for a single photo.

Related: Fast R-CNN, Faster R-CNN, Region proposal

An augmentation that cuts a randomly positioned sub-region from an image and resizes it to the input size. It teaches models that objects can appear at different positions and scales.

Example. A 256 x 256 image is randomly cropped to 224 x 224 each epoch, so the model sees slightly different views.

Related: Data augmentation for images, Horizontal flip, Image resizing

Random Sample Consensus, an algorithm that fits a model to data containing many wrong points by repeatedly fitting to random small samples and keeping the model that most points agree with. In vision it removes bad feature matches when estimating a homography.

Example. Of 500 matches, 150 are wrong; RANSAC finds the homography supported by the other 350 and ignores the outliers.

Related: Feature matching, Homography, Stereo vision

The region of the original input image that can influence one unit in a given layer. It grows as layers are stacked, so early units see small edges and deeper units see whole objects, which is why depth matters for recognising large structures.

Example. Two stacked 3 x 3 convolutions with stride 1 give each output unit a 5 x 5 receptive field; three give 7 x 7.

Related: Convolutional neural network, Feature map, Dilated convolution

A candidate image region that might contain an object, generated before classification. Early detectors used hand-made methods such as selective search; later ones learned proposals with a network. Proposals narrow the search from every possible window to a manageable few thousand.

Example. Selective search may propose about 2,000 regions per image for R-CNN to classify.

Related: R-CNN, Region proposal network, Sliding window

A small network attached to a backbone that scores anchors for objectness and refines their boxes, outputting candidate regions. It replaced slow hand-crafted proposal methods in Faster R-CNN. Because it shares computation with the classifier, proposals cost little extra time.

Example. For each location on a 38 x 50 feature map with 9 anchors, the RPN outputs 9 objectness scores and 9 box adjustments.

Related: Faster R-CNN, Anchor box, Region proposal

Residual Network, a CNN family built from blocks with residual connections, available in depths such as 18, 34, 50, 101 and 152 layers. It won ImageNet 2015 and remains a common backbone and baseline.

Example. ResNet-50 is often used as the default pretrained backbone when fine-tuning on a new image dataset.

Related: Residual connection, Pretrained backbone, Bottleneck block

The number of pixels in an image along its width and height, which sets how much fine detail it can hold. Higher resolution keeps small objects visible but costs more memory and compute, so vision models usually resize inputs to a fixed working resolution.

Example. A model trained on 224 x 224 inputs receives a 4000 x 3000 phone photo only after it is resized down, so a tiny distant object may shrink to a few pixels.

Related: Pixel, Image resizing, Aspect ratio

A colour representation in which each pixel is described by three intensities, red, green and blue, that mix additively to produce other colours. It matches how screens and cameras work and is the default input format for most vision models.

Example. Pure yellow is stored as (255, 255, 0) in 8-bit RGB, meaning full red, full green and no blue.

Related: Channel, HSV colour space, Grayscale image

A layer that takes a region of any size from a feature map and converts it to a fixed-size grid, such as 7 x 7, by max pooling in sub-cells. It lets a fixed classifier handle variable-size proposals; RoI Align is a more precise variant.

Example. A 23 x 41 region is divided into a 7 x 7 grid of cells, and each cell's maximum forms the output.

Related: Fast R-CNN, Faster R-CNN, Mask R-CNN

Assigning a class label to every pixel, without distinguishing separate objects of the same class. All pixels of any car are simply 'car'. The result is a map the same size as the image, where every pixel carries a class label.

Example. In a street photo, two adjacent cars form one merged 'car' region in the output map.

Related: Instance segmentation, U-Net, Image segmentation

Scale-Invariant Feature Transform, a classic algorithm that detects keypoints at several scales and describes them with 128-number vectors of local gradient orientations. It matches the same object across different sizes, angles and lighting.

Example. SIFT can match a photo of a monument taken from far away with a close-up of the same monument.

Related: Feature detection, Feature descriptor, ORB

A detector that predicts class scores and boxes directly from the feature map in one pass, without a separate proposal step. It is generally faster and suits real-time use. It is faster because there is no separate second pass per candidate.

Example. SSD and YOLO are single-stage detectors that can run on live video.

Related: Two-stage detector, YOLO, SSD

A pair of 3 x 3 kernels that estimate the horizontal and vertical brightness gradients of an image, with some smoothing built in. Combining their outputs gives edge strength and direction cheaply.

Example. The horizontal Sobel kernel is [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], which responds strongly to vertical edges.

Related: Edge detection, Image gradient, Canny edge detector

The number of pixels a kernel or pooling window moves each step as it slides across an image. A stride of 1 visits every position, while a stride of 2 skips every other one and roughly halves the output height and width.

Example. A 224 x 224 input with a 7 x 7 kernel, padding 3 and stride 2 gives a 112 x 112 output, as in the first layer of ResNet.

Related: Convolution, Padding, Pooling layer

The first stage of many OCR systems: locating where text appears in an image and returning boxes or polygons around words or lines, before any characters are read. Separating finding from reading makes each stage easier to train and debug.

Example. On a shop-front photo, the detector draws boxes around the sign's Hindi and English lines.

Related: Optical character recognition, Text recognition, Bounding box

The stage of OCR that reads the characters in a cropped text region and outputs a string. Modern systems use a CNN or transformer encoder with a sequence decoder. It must cope with fonts, blur, skew and different lengths of text.

Example. A cropped image of a licence plate is turned into the string 'KA05MN1234'.

Related: Text detection, Connectionist temporal classification, Optical character recognition

Converting a grayscale image to a black-and-white one by setting pixels above a cutoff to white and the rest to black. It is a simple way to separate foreground from background, as in document scanning.

Example. With threshold 128, a pixel of 200 becomes 255 and a pixel of 90 becomes 0.

Related: Otsu's method, Binary mask, Grayscale image

A metric that counts a prediction as correct if the true class is among the model's k highest-scoring classes. Top-5 is common on datasets with many similar classes, where the exact label can be ambiguous.

Example. If the true label is 'Labrador' and the model ranks it second, top-1 counts it wrong while top-5 counts it correct.

Related: Image classification, ImageNet, Softmax

The property that a system gives the same answer when an object moves to a different position in the image. Convolution shares weights and pooling discards exact position, so CNNs are approximately, not perfectly, invariant to shifts.

Example. A classifier should still say 'cat' whether the cat sits in the left or right half of the photo.

Related: Max pooling, Convolution, Data augmentation

A learned layer that increases spatial size, working roughly in reverse of a strided convolution. It is used in decoders and segmentation networks to turn small feature maps back into full-resolution outputs, though it can leave checkerboard artefacts.

Example. A stride-2 transposed convolution turns a 14 x 14 feature map into a 28 x 28 one.

Related: Upsampling, U-Net, Semantic segmentation

A detector that first proposes candidate regions and then classifies and refines each one. This usually gives higher accuracy but slower speed than single-stage designs. The second stage sees a cropped, focused view of each candidate, which helps precision.

Example. Faster R-CNN is a two-stage detector; YOLO is single-stage.

Related: Single-stage detector, Faster R-CNN, Region proposal network

A segmentation network shaped like a U: an encoder shrinks the image to capture context, a decoder expands it back, and skip connections copy detailed features from encoder to decoder. It works well even with small datasets, notably in medical imaging.

Example. U-Net is widely used to outline organs and lesions in CT and MRI slices.

Related: Semantic segmentation, Transposed convolution, Residual connection

Increasing the height and width of a feature map or image, by fixed interpolation or by a learned layer. Segmentation, super-resolution and feature pyramids all need it to bring coarse features back to fine resolution.

Example. Bilinear upsampling by 2 turns a 16 x 16 mask prediction into a 32 x 32 one.

Related: Transposed convolution, Interpolation, Feature pyramid network

A family of CNNs from the Visual Geometry Group at Oxford built from stacks of small 3 x 3 convolutions followed by pooling. Its simple, uniform design is easy to understand, though it has very many parameters.

Example. VGG-16 has 16 weight layers and is still used as a feature extractor in tasks like style transfer.

Related: AlexNet, ResNet, Transfer learning

A model that splits an image into fixed-size patches, treats each patch as a token and processes them with a transformer encoder using self-attention. With enough data it matches or beats CNNs on classification.

Example. A 224 x 224 image cut into 16 x 16 patches becomes a sequence of 196 tokens.

Related: Patch embedding, Self-attention, CLS token

You Only Look Once, a family of single-stage detectors that treat detection as one prediction step over a grid on the image, giving fast, real-time detection. Many versions from different authors exist, so behaviour varies by version.

Example. A YOLO model on a CCTV stream can label vehicles frame by frame fast enough for live monitoring.

Related: Single-stage detector, Non-maximum suppression, Object detection

Recognising classes the model was never explicitly trained on, usually by relating images to text descriptions. It avoids collecting labelled data for every new category. It relies on shared meaning, such as text embeddings, to bridge seen and unseen classes.

Example. CLIP can label a photo as 'a kolam pattern' though no training step used that class name.

Related: Contrastive language-image pretraining, Image classification, Transfer learning