Glossaryπ
Deep learning-relatedπ
Adamπ
Adam (adaptive moment estimation) is a first-order gradient-based optimization algorithm for stochastic objective functions, which computes individual adaptive learning rates. In the deep learning methods, this algorithm can be used to minimize the loss function.
Anchorπ
Anchors are fixed bounding boxes. They serve as reference boxes, with the aid of which the network proposes bounding boxes for the objects to be localized.
Annotationπ
An annotation is the ground truth information, what a given instance in the data represents, in a way recognizable for the network. This is, for example, the bounding box and the corresponding label for an instance in object detection.
Anomalyπ
An anomaly means something deviating from the norm, something unknown.
Backboneπ
Part of a pretrained classification network. Its task is to generate various feature maps, which is why the classifying layer has been removed.
Batch sizeπ
The dataset is divided into smaller subsets of data, which are called batches.
The batch size determines the number of images taken into a batch and thus processed simultaneously.
Hyperparameter: 'batch_size'
Bounding boxπ
Bounding boxes are rectangular boxes used to define a part within an image and to specify the localization of an object within an image.
Change strategyπ
The strategy, when and how hyperparameters are changed during the training of a DL model
Class agnosticπ
Class agnostic means without the knowledge of the different classes.
In HALCON, we use it for reduction of overlapping predicted bounding boxes. This means, for a class-agnostic bounding box suppression, the suppression of overlapping instances is done ignoring the knowledge of classes, thus strongly overlapping instances get suppressed independently oftheir class.
Classπ
Discrete categories (for example, βappleβ, βpeachβ, βpearβ) that the network distinguishes. In HALCON, the class of an instance is given by its corresponding annotation.
Classifierπ
In the context of deep learning: The classifier takes an image as input and returns the inferred confidence values, expressing how likely the image belongs to every distinguished class. For example, the three classes βappleβ, βpeachβ, and βpearβ are distinguished. When giving an image of an apple to the classifier, the resulting confidences βappleβ: 0.92, βpeachβ: 0.07, and βpearβ: 0.01 could be returned.
COCOπ
Abbreviation for βcommon objects in contextβ, a large-scale object detection, segmentation, and captioning dataset. There is a common file format for each of the different annotation types.
Confidenceπ
Number expressing the affinity of an instance to a class.
In HALCON, the confidence is the probability, given in the range of [0,1].
Alternative name: score
Confusion matrixπ
Table comparing the classes predicted by the network (top-1) with the ground truth class affiliations. It is often used to visualize the performance of the network on a validation or test set.
Convolutional Neural Networks (CNNs)π
Neural networks used in deep learning, characterized by the presence of at least one convolutional layer in the network. They are particularly successful for image classification.
Dataπ
In the context of deep learning, the instances to be recognized (for example, images) and their appropriate information concerning the predictable characteristics (for example, the labels in case of classification)
Data augmentationπ
Generation of altered copies of samples within a dataset. This is done to augment the richness of the dataset, for example, through flipping or rotating.
Datasetπ
Dataset refers to the complete set of data used for a training. The dataset is split into three, if possible disjoint, subsets:
The training set contains the data on which the algorithm optimizes the network directly.
The validation set contains the data to evaluate the network performance during training.
The test set is used to test possible inferences (predictions), thus to test the performance on data without any influence on the network optimization.
Deep learningπ
Originally used to describe the training of neural networks with multiple hidden layers, today it is rather used as a generic term for several different concepts in machine learning. In HALCON, βdeep learningβ is used for methods using a neural network with multiple hidden layers.
Epochπ
In the context of deep learning, an epoch is a single training iteration over the entire training data, i.e., over all batches. Iterations over epochs should not be confused with the iterations over single batches (for example, within an epoch).
Errorsπ
In the context of deep learning, we refer to error when the inferred class of an instance does not match the real class (for example, the ground truth label in case of classification). Within HALCON, error in particular refers to the top-1 error.
Feature mapπ
Output of a given layer
Feature pyramidπ
Group of feature maps, whereby every feature map origins from another level, i.e., it is smaller than its preceding levels.
Headπ
Heads are subnetworks. For certain architectures, they attach on selected pyramid levels. These subnetworks proceed information from previous parts of the total network o generate spatially resolved output, for example, for the class predictions. They generate the output of the total network and hence constitute the input of the losses.
Hyperparameterπ
Like every machine learning model, CNNs contain many formulas with many parameters. During training, the model learns from the data in the sense of optimizing the parameters. However, such models can have other, additional parameters, which are not directly learned during the regular training. These parameters have values set before starting the training. We refer to this last type of parameters as hyperparameters to distinguish them from the network parameters that are optimized during training. Hyperparameters are solver-specific parameters. Examples: the initial learning rate or the batch size
Inference phaseπ
Stage when a trained network is applied to predict (infer) instances (which can be the total input image or just a part of it) and eventually their localization. Unlike during the training phase, the network is not changed anymore in the inference phase.
In-distributionπ
In-distribution refers to data that comes from the same underlying distribution as the data on which a model was trained. When a model encounters in-distribution data during inference, the data is similar in terms of its statistical properties, features, and patterns to what the model has seen before during training.
Intersection over Unionπ
The intersection over Union (IoU) is a measure to quantify the overlap of two areas.
We can determine the parts common in both areas, the intersection, as well as the united areas, the union.
The IoU is the ratio between the two areas intersection and union.
The application of this concept may differ between the methods.
Labelπ
Arbitrary strings used to define the class of an image. In HALCON, these labels are given by the image name or by the directory name, for example, βapple_01.pngβ, βpear.pngβ, βpeach/01.pngβ.
Layer, hidden layerπ
Building block in a neural network, thus performing specific tasks (for example, convolution, pooling). It can be seen as a container, which receives weighted input, transforms it, and returns the output to the next layer. Input and output layers are connected to the dataset, i.e., the images or the labels, respectively. All layers in between are called hidden layers.
Learning rateπ
Weighting with which the gradient is considered when updating the arguments of the loss function.
In simple words, when optimizing a function, the gradient defines the direction into which is optimized, and the learning rate determines how far along this direction it is stepped.
Hyperparameter: 'learning_rate'
Alternative names: Ξ», step size
Levelπ
Denotes the whole group of layers within a feature pyramid network, whose feature maps have the same width and height. The input image represents level 0.
Lossπ
Compares the prediction from the network with the given information, what to find in the image (and, if applicable, also where), and penalizes deviations.
The loss function is optimized during the training process to adapt the network to a specific task.
Alternative names: objective function, cost function, utility function
Momentumπ
Used for the optimization of the loss function arguments; ΞΌ β [0,1).
When the loss function arguments are updated (after having calculated the gradient), a fraction ΞΌ of the previous update vector of the past iteration step is added.
This has the effect of damping oscillations.
When ΞΌ is set to 0, the momentum method has no influence.
In simple words, when updating the loss function arguments, we still remember the step we did for the last update.
Now we go a step in direction of the gradient with a length according to the learning rate and additionally we repeat the step we did last time, but this time only ΞΌ times as long.
Hyperparameter: 'momentum'
Non-maximum suppressionπ
In object detection, non-maximum suppression is used to suppress overlapping predicted bounding boxes. When different instances overlap more than a given threshold value, only the one with the highest confidence value is kept while the other instances, not having the maximum confidence value, aresuppressed.
Out-of-Distributionπ
Refers to data that significantly differs from the data on which a model was trained. When a model encounters out-of-distribution data during inference, the dataβs statistical properties, features, or patterns are unfamiliar to the model, leading to potential challenges in making accurate predictions.
Overfittingπ
Happens when the network starts to βmemorizeβ training data instead of learning how to find general rules for the classification. This becomes visible when the model continues to minimize error on the training set but the error on the validation set increases. Since most neural networks have a huge amount of weights, these networks are particularly prone to overfitting.
Regularizationπ
Technique to prevent neural networks from overfitting by constraining their weights.
Simply put, regularization favors simpler models that are less likely to fit to noise in the training data and generalize better.
L2 regularization works by adding an extra term to the loss function, which penalizes the squared magnitude of the weights.
Weight decay, on the other hand, directly shrinks the magnitude of the weights by a small factor in each iteration of the training.
Hyperparameters: 'weight_prior', 'weight_decay'
Alternative names:
regularization parameter, weight decay parameter, Ξ»
In HALCON, Ξ» is used for the learning rate. Within formulas, Ξ± is used for the regularization parameter.
Retrainingπ
Defined as updating the weights of an already pretrained network, i.e., during retraining the network learns the specific task.
Alternative name: fine-tuning.
SGDπ
See Stochastic gradient descent.
Solverπ
Optimizes the network by updating the weights in a way to optimize (i.e., minimize) the loss.
Stochastic gradient descentπ
Iterative optimization algorithm for differentiable functions. A key feature of the SGD is to calculate the gradient only based on a single batch containing stochastically sampled data and not all data. With deep learning methods, this algorithm can be used to calculate the gradient to optimize (i.e., minimize) the loss function.
Top-k errorπ
For a given image class, the classifier infers confidences of how likely the image belongs to every distinguished class.
Thus, for an image we can sort the predicted classes according to the confidence value the classifier assigned.
The top-k error tells the ratio of predictions where the ground truth class is not within the k predicted classes with highest probability.
In case of a top-1 error, we check if the target label matches the prediction with the highest probability.
In case of a top-3 error, we check if the target label matches one of the top 3 predictions (the three labels getting the highest probability for this image).
Alternative name: top-k score
Transfer learningπ
Refers to the technique where a network is built upon the knowledge of an already existing network. In concrete terms this means taking an already (pre-) trained network with its weights and adapt the output layer to the respective application to get your network. In HALCON, we also see the following retraining step as a part of transfer learning.
Underfittingπ
Occurs when the model over-generalizes. In other words, it is not able to describe the complexity of the task. This is directly reflected in the error on the training set, which does not decrease significantly.
Weightsπ
In general, weights are the free parameters of the network, which are altered during the training due to the optimization of the loss. A layer with weights multiplies or adds them with its input values. In contrast to hyperparameters, weights are optimized and thus changed during the training.
Filtering-relatedπ
Smoothingπ
Applying a filter function on the given data to capture the main data patterns while removing noise
Random noiseπ
Stationary variation of brightness or color information by a small random amount for every pixel with an assumed mean of 0 over the total image
Systematic noiseπ
Predictable noise, for example caused by the specific setup used to acquire the images
Identification-relatedπ
Data code symbolπ
Two-dimensional graphical symbol that encodes characters and numbers. It is constructed by dark and light dots, bars, or small squares called modules. There are different types of data codes. Two common types are stacked codes and matrix codes.
Stacked codeπ
Type of data code symbol that contains a stack of 1D bar codes arranged in rows and columns. To ensure that the complete stack of 1D bar codes is processed, the symbol contains a start and a stop pattern. Additionally, the symbol is framed by a quiet zone.
Matrix codeπ
Type of data code symbol that uses graphical patterns consisting of dark and light modules arranged in two dimensions. The symbol consists of three components: a finder pattern, a data pattern, and a quiet zone.
Moduleπ
Dark and light dots, bars, or small squares that are used to build a bar or data code symbol
Quiet zoneπ
Homogeneous frame around the symbolβs border, which makes the symbol better distinguishable from the background or from other objects in the image
Finder patternπ
Pattern that is used to find the symbol and its orientation in the image. The pattern differs depending on the used bar or data code type.
Image Source-relatedπ
Deviceπ
The actual imaging device. This includes different kinds of sensors, ranging from conventional machine vision cameras to 3D sensors and, in special cases, controllers or virtual devices.
GenICamπ
The GenICam standard allows machine vision devices to expose their configuration properties to the connected application via a feature description XML file.
HALCON image sources rely on this standard to provide access to the device configuration (see set_image_source_param and get_image_source_param).
Usually, HALCON users do not need to deal with these XML files directly.
The standard defines a wide range of features, many of which go beyond the scope of this documentation.
For more detailed information, consult the official standard documents, particularly the GenICam Standard Features Naming Convention (SFNC) and the GenICam GenTL SFNC.
These resources are hosted by the European Machine Vision Association (EMVA).
GenTLπ
The GenICam GenTL standard allows machine vision applications and frameworks from different vendors to access machine vision cameras and devices from different other vendors. It specifies a clearly defined software interface that allows to hide different implementation details from the using application. An application like HALCON using different GenTL producers like the plugins can also be referred to as GenTL consumer.
Image source handleπ
Provides a generic interface to a wide variety of underlying image acquisition standards and technologies. From the viewpoint of the HALCON library, any image source handle provides the same basic operations. The image source handle ties together software and hardware components that are needed to acquire images.
Interfaceπ
The hardware interfacing with the imaging device, for example, where the cable coming from the imaging device is plugged into the host machine. Thus, in the image source context, the term βinterfaceβ does not refer to what was formerly called the image acquisition interface. If framegrabbers are used, each framegrabber card is typically modeled as one interface. Per default, HALCON image sources hide the complexity of multiple interface by automatically using the correct one.
Moduleπ
The individual units of an image source (plugin, interface, device, stream). Each module allows access to its parameters. The module names therefore also apply to parameter groups of the same name.
Pluginπ
Abstract the various technologies used to transfer images from the imaging device to the receiving application, also referred to as GenTL Producer.
Plugins exist in form of .cti files on each target machine.
Plugins can be selected and installed with the MVTec Software Manager (SOM).
Plugins (i.e., GenTL Producers) provided by other manufacturers are automatically available after installing the manufacturerβs proprietary SDK.
Consult the manufacturers documentation for more in depth instructions.
At runtime, plugins not provided by MVTec are found by HALCON through the environment variable GENICAM_GENTL64_PATH (or GENICAM_GENTL32_PATH), which lists all the paths that contain .cti files.
Streamπ
A device can allow access to none, one, or multiple streams of images. The typical machine vision camera just gives access to one stream. Per default, HALCON image sources hide the complexity of multiple streams.
Measurement-relatedπ
Annular arcπ
A circular arc with an associated width
Measure objectπ
Data structure containing a specific region of interest that is prepared for the extraction of straight edges, which lie perpendicular to the major axis of a rectangle or an annular arc
Metrology modelπ
Data structure containing all metrology objects, all information needed for the measurement, and the measurement results
Metrology objectπ
Data structure for the object to be measured with 2D metrology. The metrology object is represented by a specific geometric shape for which the shape parameters are approximately known. Additionally, it contains parameters that control the measurement, for example, parameters that specify the dimension and distribution of the measure regions.
Measure regionsπ
Rectangular regions that are arranged perpendicular to the boundaries of the approximate objects. Within these regions, the edges that are used to get the exact shape parameters of the metrology objects are extracted.
RANSACπ
RANSAC (RANdom SAmple Consensus) is an iterative method to estimate parameters of a mathematical model from a set of observed data that contains outliers. It is a non-deterministic algorithm producing a reasonable result only with a certain probability, which increases as more iterations are allowed. The algorithm is used in the measurement process to fit the edge positions inside all measure regions to geometric shapes.
Morphology-relatedπ
Input regionπ
Region that is modified by morphological operators
Non-linear operatorπ
Operator that does not necessarily preserve structures of the input image
Structuring elementπ
Region that is used to scan the input image
Texture inspection-relatedπ
Training imagesπ
Images that are used for training
Test imagesπ
Images that are compared to the trained texture inspection model
Patchπ
Collection of neighboring pixels
Texture featureπ
The pixel values within a patch
Sampleπ
A texture feature that is used for training
Novelty scoreπ
In the test process, the texture features of the test images are compared to the texture inspection model and their novelty score is calculated. The larger this value, the more probable it is that the individual texture feature does not fit to the texture inspection model.
Novelty thresholdπ
Threshold determined during the training process. Texture features with a novelty score below the novelty threshold fit to the texture inspection model. Texture features with a higher novelty score do not.