Machine learning techniques aim to automatically differentiate patterns using mathematical algorithms. These techniques are commonly used to classify images, to make decisions within the business world (for example, to decide which bank clients can receive a loan or how much each client should pay for insurance depending on their background), as well as within many other areas of science and technology. Mainly two types of techniques can be distinguished: supervised and unsupervised.

In supervised learning, the computer is trained by providing previously labeled patterns, so that the algorithm used must find the boundaries that separate the possible different types of patterns. Adaboost and some neural networks are part of this group.

In unsupervised learning, the computer is trained with patterns that have not been previously classified and it is the computer itself that must group the different patterns into different classes. K-means and some neural networks are part of this group.

Both techniques are widely used in computer vision, especially in image classification and segmentation.[2]​.

Object detection is the part of computer vision that studies how to detect the presence of objects in an image based on their visual appearance, either based on the type of object (a person, a car) or the instance of the object (my car, the neighbor's car). Generally two parts can be distinguished in the detection process: the extraction of features from the content of an image and the search for objects based on these features.

Feature extraction consists of obtaining compact mathematical models that "summarize" the content of the image in order to simplify the process of learning the objects to recognize. These characteristics are commonly called descriptors.

There are various types of descriptors, which will have better or worse performance depending on the type of object to be recognized and the conditions of the recognition process (light controlled or not, distance to the object to be recognized known or not). Basic histograms of color or light intensity, LBP (Local Binary Pattern, used especially for textures) or more advanced descriptors such as HOG (Histogram of Oriented Gradients) or SIFT can be used.

Different machine learning techniques can be used for the classification process. There are different methods, such as logistic regression, or more advanced methods based on machine learning techniques such as SVM or AdaBoost (Adaptive Boost).

The biggest challenges of both feature extraction and classification is finding descriptors and classifiers "Classifier (mathematics)") that are invariant to the changes that an object may have, such as its position or illumination.[3]​[4]​.

The term video analytics describes a wide number of new technologies and developments in the field of video surveillance and security. These changes are producing more effective and efficient security systems.

Video analysis is essential in the video surveillance and security sector. In a traditional CCTV (closed circuit television) system it is common to display the content of up to 16 cameras simultaneously. This task is complicated for a security guard since there are studies that assure that after 22 minutes of supervision he loses up to 95 percent of the activity in the scene. With video analysis, the security guard is alerted when there is movement or indicates which camera is most likely to have suspicious or dangerous activity.

One of the core capabilities of video analytics is motion detection: technology that identifies and alerts when motion occurs. Sophisticated motion detection adaptations include sensors that detect motion in unauthorized directions. See Motion Detector.

Some applications of video analysis are detecting abandoned objects in crowded places, monitoring works of art in museums, and detecting unauthorized vehicles entering certain areas. Vehicle license plate detection and traffic congestion are features that are being developed.

Despite the constant evolution of video analysis, some of its applications are not very rigorous. Technologies such as facial recognition and suspicious movement detection are still unreliable and often produce false alarms. See Facial recognition system.

La visión 3D informática se encarga de proporcionar la capacidad de emular la visión humana a un ordenador. Con dicha capacidad el ordenador podrá generar un modelo tridimensional de un objeto o escena, generalmente a partir de una imagen en 2D. Existen técnicas o sistemas que permiten captar la profundidad de los objetos o en una escena, como por ejemplo: sistemas estereoscópicos, mediante múltiples cámaras, sistemas basados en el tiempo de vuelo o los sistemas basados en el escáner de luz estructurada.

Las aplicaciones son numerosas, en el campo de la automoción por ejemplo, Google self-Driving car utiliza la detección de objetos, junto con radares y sensores para conducir por la vía pública de una forma autónoma.

3D reconstruction from 2D image

Obtaining a 3D reconstruction from a 2D image poses a series of problems since there are many objects with similar shapes or these same objects may appear in the image in a different way. The reconstruction process must be limited by assuming the similarity of the shapes between the objects. To compare the query image, a series of 3D models must be created for each class. These models are used to generate a database of objects of the class with their respective depth maps. We will call this supervised learning because all the objects to be treated will be known. This will give us examples of feasible 3D maps associated with shapes that will be used to estimate the image.

An example of the blocks that a 3D recognition system must contain from a 2D image:

• - Image acquisition: In this stage, a two-dimensional projection of the light reflected by the objects in the scene is captured.

• - Segmentation "Segmentation (image processing)"): detection of edges and regions: Allows you to separate the different objects in the scene.

• - Feature Extraction: A numerical representation in vector form is obtained for each image.

• - Recognition and localization: Using techniques, such as triangulation, the object is located in 3D space.

• - Interpretation or estimation: The scene is estimated from the information obtained.

3D reconstruction from multiple 2D images

The goal of 3D reconstruction with multiple images is to find out the geometry of a scene captured by a collection of images. Typically, the camera position and internal parameters are assumed to be known or can be estimated from the image set. The automatic correspondence of the objects in the image is usually ambiguous and incomplete, it is recommended to have prior knowledge about the objects.

• - Multi-camera: The objective of this technique is to capture the object or scene with several cameras calibrated to obtain different points of view and thus generate depth data.

• - Steroscopic cameras: They allow us to emulate human vision by creating two images as if they were eyes. See Stereoscopic cameras.

• - Structured light scanner: By projecting a light pattern, the device is capable of capturing the shape and characteristics of the object. See Structured light scanner.

• - Time of Flight (TOF): Using a sensor that emits an infrared signal, this signal hits the scene or object and bounces back to the camera. The camera generates a grayscale image that gives us depth information. See 3D Television: TOF.

Currently, a large part of the best-known computer vision algorithms have already been implemented in specific libraries for this purpose, that is, it is generally not necessary to develop them. In this way, a computer vision systems developer can focus on exploiting the functionalities offered by these libraries to solve their own problems in the most suitable way that best suits each case.

One of the best-known libraries is OpenCV, free and open, available for popular languages ​​such as Python or C++. This library provides the fundamental tools for reading and saving images by programming code, basic functions to improve the quality of the images, some segmentation methods such as image thresholding and even a small range of machine learning functions, although this is a little further outside the main idea of ​​the library.

It should be noted that there is also an implementation intended for Android devices, whose functionalities can be exploited from any code implemented in Android Studio, making the appropriate configurations.

There are also other Python libraries such as Scikit-Image or Scikit-Learn, which are of great importance in the field of image processing and machine learning. In general, it is observed that Python is a widely used language in this area of ​​computer vision.

Finally, it is worth highlighting the existence of other libraries specialized in Deep Learning (known in English by the term Deep Learning[5]​), such as TensorFlow or Torch (PyTorch being its specific implementation in Python). Likewise, these latter libraries support hardware acceleration provided by NVIDIA graphics cards through the CUDA library.

Given that the graphics processors of graphics cards adapt very well to the calculation of matrix operations (a characteristic aspect of machine learning, being even more critical in deep learning), currently this has become the most common way of working with deep models, hence the great success of the CUDA library.

Underwater computer vision is a specialized branch of computer vision focused on image perception and analysis in aquatic environments. Its development faces adverse conditions such as light attenuation, water turbidity and color distortion, which reduce visibility and make object detection difficult. [6][7]​.

Modern systems combine deep learning models with segmentation and depth estimation algorithms to interpret underwater scenes without requiring additional supervised training [6]. These methods allow us to distinguish masses of water, structures and mobile elements such as fish or floating objects. In parallel, underwater robotics has incorporated cameras and optical sensors in autonomous vehicles (AUVs and UUVs) to perform visual inspection, mapping, and environmental monitoring tasks. [8]​.

Likewise, research on the optical properties of water has made it possible to calibrate visual systems by studying the relationship between transparency, chlorophyll and available light for the capture of precise images [7]​. Recent advances integrate hydrodynamic simulations and computer vision, consolidating underwater vision as a key tool in the exploration and conservation of the marine ecosystem.[9]​[10]​.

• - Vision transformer.

• - Transformer (machine learning model) "Transformer (machine learning model)").

• - Wikimedia Commons hosts a multimedia category on Computer Vision.

• - Videos about computer vision.

Machine learning techniques aim to automatically differentiate patterns using mathematical algorithms. These techniques are commonly used to classify images, to make decisions within the business world (for example, to decide which bank clients can receive a loan or how much each client should pay for insurance depending on their background), as well as within many other areas of science and technology. Mainly two types of techniques can be distinguished: supervised and unsupervised.

In supervised learning, the computer is trained by providing previously labeled patterns, so that the algorithm used must find the boundaries that separate the possible different types of patterns. Adaboost and some neural networks are part of this group.

In unsupervised learning, the computer is trained with patterns that have not been previously classified and it is the computer itself that must group the different patterns into different classes. K-means and some neural networks are part of this group.

Both techniques are widely used in computer vision, especially in image classification and segmentation.[2]​.

Object detection is the part of computer vision that studies how to detect the presence of objects in an image based on their visual appearance, either based on the type of object (a person, a car) or the instance of the object (my car, the neighbor's car). Generally two parts can be distinguished in the detection process: the extraction of features from the content of an image and the search for objects based on these features.

Feature extraction consists of obtaining compact mathematical models that "summarize" the content of the image in order to simplify the process of learning the objects to recognize. These characteristics are commonly called descriptors.

There are various types of descriptors, which will have better or worse performance depending on the type of object to be recognized and the conditions of the recognition process (light controlled or not, distance to the object to be recognized known or not). Basic histograms of color or light intensity, LBP (Local Binary Pattern, used especially for textures) or more advanced descriptors such as HOG (Histogram of Oriented Gradients) or SIFT can be used.

Different machine learning techniques can be used for the classification process. There are different methods, such as logistic regression, or more advanced methods based on machine learning techniques such as SVM or AdaBoost (Adaptive Boost).

The biggest challenges of both feature extraction and classification is finding descriptors and classifiers "Classifier (mathematics)") that are invariant to the changes that an object may have, such as its position or illumination.[3]​[4]​.

The term video analytics describes a wide number of new technologies and developments in the field of video surveillance and security. These changes are producing more effective and efficient security systems.

Video analysis is essential in the video surveillance and security sector. In a traditional CCTV (closed circuit television) system it is common to display the content of up to 16 cameras simultaneously. This task is complicated for a security guard since there are studies that assure that after 22 minutes of supervision he loses up to 95 percent of the activity in the scene. With video analysis, the security guard is alerted when there is movement or indicates which camera is most likely to have suspicious or dangerous activity.

One of the core capabilities of video analytics is motion detection: technology that identifies and alerts when motion occurs. Sophisticated motion detection adaptations include sensors that detect motion in unauthorized directions. See Motion Detector.

Some applications of video analysis are detecting abandoned objects in crowded places, monitoring works of art in museums, and detecting unauthorized vehicles entering certain areas. Vehicle license plate detection and traffic congestion are features that are being developed.

Despite the constant evolution of video analysis, some of its applications are not very rigorous. Technologies such as facial recognition and suspicious movement detection are still unreliable and often produce false alarms. See Facial recognition system.

La visión 3D informática se encarga de proporcionar la capacidad de emular la visión humana a un ordenador. Con dicha capacidad el ordenador podrá generar un modelo tridimensional de un objeto o escena, generalmente a partir de una imagen en 2D. Existen técnicas o sistemas que permiten captar la profundidad de los objetos o en una escena, como por ejemplo: sistemas estereoscópicos, mediante múltiples cámaras, sistemas basados en el tiempo de vuelo o los sistemas basados en el escáner de luz estructurada.

Las aplicaciones son numerosas, en el campo de la automoción por ejemplo, Google self-Driving car utiliza la detección de objetos, junto con radares y sensores para conducir por la vía pública de una forma autónoma.

3D reconstruction from 2D image

Obtaining a 3D reconstruction from a 2D image poses a series of problems since there are many objects with similar shapes or these same objects may appear in the image in a different way. The reconstruction process must be limited by assuming the similarity of the shapes between the objects. To compare the query image, a series of 3D models must be created for each class. These models are used to generate a database of objects of the class with their respective depth maps. We will call this supervised learning because all the objects to be treated will be known. This will give us examples of feasible 3D maps associated with shapes that will be used to estimate the image.

An example of the blocks that a 3D recognition system must contain from a 2D image:

• - Image acquisition: In this stage, a two-dimensional projection of the light reflected by the objects in the scene is captured.

• - Segmentation "Segmentation (image processing)"): detection of edges and regions: Allows you to separate the different objects in the scene.

• - Feature Extraction: A numerical representation in vector form is obtained for each image.

• - Recognition and localization: Using techniques, such as triangulation, the object is located in 3D space.

• - Interpretation or estimation: The scene is estimated from the information obtained.

3D reconstruction from multiple 2D images

The goal of 3D reconstruction with multiple images is to find out the geometry of a scene captured by a collection of images. Typically, the camera position and internal parameters are assumed to be known or can be estimated from the image set. The automatic correspondence of the objects in the image is usually ambiguous and incomplete, it is recommended to have prior knowledge about the objects.

• - Multi-camera: The objective of this technique is to capture the object or scene with several cameras calibrated to obtain different points of view and thus generate depth data.

• - Steroscopic cameras: They allow us to emulate human vision by creating two images as if they were eyes. See Stereoscopic cameras.

• - Structured light scanner: By projecting a light pattern, the device is capable of capturing the shape and characteristics of the object. See Structured light scanner.

• - Time of Flight (TOF): Using a sensor that emits an infrared signal, this signal hits the scene or object and bounces back to the camera. The camera generates a grayscale image that gives us depth information. See 3D Television: TOF.

Currently, a large part of the best-known computer vision algorithms have already been implemented in specific libraries for this purpose, that is, it is generally not necessary to develop them. In this way, a computer vision systems developer can focus on exploiting the functionalities offered by these libraries to solve their own problems in the most suitable way that best suits each case.

One of the best-known libraries is OpenCV, free and open, available for popular languages ​​such as Python or C++. This library provides the fundamental tools for reading and saving images by programming code, basic functions to improve the quality of the images, some segmentation methods such as image thresholding and even a small range of machine learning functions, although this is a little further outside the main idea of ​​the library.

It should be noted that there is also an implementation intended for Android devices, whose functionalities can be exploited from any code implemented in Android Studio, making the appropriate configurations.

There are also other Python libraries such as Scikit-Image or Scikit-Learn, which are of great importance in the field of image processing and machine learning. In general, it is observed that Python is a widely used language in this area of ​​computer vision.

Finally, it is worth highlighting the existence of other libraries specialized in Deep Learning (known in English by the term Deep Learning[5]​), such as TensorFlow or Torch (PyTorch being its specific implementation in Python). Likewise, these latter libraries support hardware acceleration provided by NVIDIA graphics cards through the CUDA library.

Given that the graphics processors of graphics cards adapt very well to the calculation of matrix operations (a characteristic aspect of machine learning, being even more critical in deep learning), currently this has become the most common way of working with deep models, hence the great success of the CUDA library.

Underwater computer vision is a specialized branch of computer vision focused on image perception and analysis in aquatic environments. Its development faces adverse conditions such as light attenuation, water turbidity and color distortion, which reduce visibility and make object detection difficult. [6][7]​.

Modern systems combine deep learning models with segmentation and depth estimation algorithms to interpret underwater scenes without requiring additional supervised training [6]. These methods allow us to distinguish masses of water, structures and mobile elements such as fish or floating objects. In parallel, underwater robotics has incorporated cameras and optical sensors in autonomous vehicles (AUVs and UUVs) to perform visual inspection, mapping, and environmental monitoring tasks. [8]​.

Likewise, research on the optical properties of water has made it possible to calibrate visual systems by studying the relationship between transparency, chlorophyll and available light for the capture of precise images [7]​. Recent advances integrate hydrodynamic simulations and computer vision, consolidating underwater vision as a key tool in the exploration and conservation of the marine ecosystem.[9]​[10]​.

• - Vision transformer.

• - Transformer (machine learning model) "Transformer (machine learning model)").

• - Wikimedia Commons hosts a multimedia category on Computer Vision.

• - Videos about computer vision.