Image Feature Overview

From http://blog.csdn.net/passball/article/details/5204132
I found a good article on the Internet about image feature extraction. The project I made for myself was a bit similar and was released for your reference.

Feature extraction is a concept in computer vision and image processing. It refers to the use of computers to extract image information and determine whether each image point belongs to an image feature. Feature Extraction divides points in an image into different subsets, which are usually isolated points, continuous curves, or continuous areas.

Feature Definition

So far, no universal and precise definition of features has been provided. The precise definition of features is often determined by the problem or application type. A feature is an "interesting" part of a digital image. It is the starting point of many computer image analysis algorithms. Therefore, whether an algorithm is successful is usually determined by the features it uses and defines. Therefore, the most important feature of feature extraction is "repeatability": the features extracted from different images in the same scenario should be the same.

Feature extraction is a preliminary operation in image processing, that is, it is the first operation for an image. It checks each pixel to determine whether the pixel represents a feature. If it is a part of a larger algorithm, this algorithm generally checks only the feature areas of the image. As a prerequisite for feature extraction, the input image is usually smoothed in the scale space through the Gaussian fuzzy kernel. Then, one or more features of an image are calculated using the local derivative operation.

Sometimes, if feature extraction requires a lot of computing time, but the available time is limited, a high-level algorithm can be used to control the feature extraction class, so that only the part of the image is used to find the feature.

Because many computer image algorithms use feature extraction as their basic computing steps, a large number of feature extraction algorithms have been developed, and various features are extracted. Their computational complexity and repeatability are also very different.

Edge
An edge is a pixel that forms the border (or edge) between two image regions. Generally, the shape of an edge can be arbitrary or include an intersection. In practice, edges are generally defined as a subset of points with a large gradient in an image. Some common algorithms also associate the points with high gradients to form a more complete description of the edge. These algorithms may also impose limitations on edges.

The edge is a one-dimensional structure.

Corner
The angle is a feature similar to the midpoint of an image. It has a two-dimensional structure in the local area. Early algorithms first perform edge detection, and then analyze the edge trend to find the sudden turning (angle) of the edge ). Later, the developed algorithm does not need to perform edge detection. Instead, it can directly search for the height curvature in the image gradient. Later, we found that sometimes we can find areas with the same feature as the angle in the image without any angle.

Region
Different from the angle, the area describes a regional structure in an image, but the area may only consist of one pixel. Therefore, many area detection can also be used to monitor the angle. A regional monitor detects an area in the image that is flat for the Angle Monitor. Area detection can be imagined to narrow down an image and then perform corner detection on the scaled image.

Ridge
A long-stripe object is called a ridge. In practice, a ridge can be seen as a one-dimensional curve representing the axis of symmetry. In addition, each ridge pixel has a ridge width. Extracting a ridge from a gray gradient image is more difficult than extracting the edge, angle, and area. In aerial photography, ridge detection is often used to identify the road, which is used to distinguish blood vessels in medical images.

Feature Extraction
After a feature is detected, it can be extracted from the image. This process may require many computers for image processing. The results are called feature descriptions or feature vectors.

Common image features include color features, texture features, shape features, and spatial relationship features.

Color Features

(1) features: a color feature is a global feature that describes the surface properties of a scene corresponding to an image or an image area. Generally, color features are based on Pixel features. At this time, all pixels belonging to the image or image area have their respective contributions. Because the color is not sensitive to changes in the direction and size of the image or image area, the color features cannot well capture the local features of objects in the image. In addition, if the database is large when only color feature queries are used, many unnecessary images are often retrieved. Color histogram is the most commonly used method to express color features. Its advantage is that it is not affected by image rotation and moving changes. Further, with the help of normalization, it is not affected by image scale changes, the disadvantage is that the color space distribution is not displayed.

(2) Common Feature Extraction and matching methods

(1) color histogram

Its advantage is that it can briefly describe the global distribution of colors in an image, that is, the proportion of different colors in the entire image, it is particularly suitable for describing images that are difficult to automatically split and images that do not need to consider the spatial location of objects. Its disadvantage is that it cannot describe the local distribution of colors in the image and the spatial location of each color, that is, it cannot describe a specific object or object in the image.

The most common color space: RGB color space and HSV color space.

Color histogram feature matching methods: histogram intersection, distance, center distance, reference color table, and accumulative color histogram.

(2) Color Set

Color histogram is a global Color Feature Extraction and matching method, which cannot distinguish local color information. The color set is an approximation of the color histogram. First, the image is converted from the RGB color space to the color space (such as the HSV space) of the visual balance, and the color space is quantified into several handles. Then, the image is divided into several areas by Automatic Color Segmentation technology. Each area is indexed by a color component of the quantified color space, so that the image is expressed as a binary color index set. In image matching, compare the distance between different image color sets and the spatial relationship between color areas

(3) color moment

The mathematical basis of this method is that any color distribution in an image can be represented by its moments. In addition, because the color distribution information is mainly concentrated in the lower-order moment, only the first-order moment (mean), second-order moment (variance) and third-order moment (skewness) of the color are used) it is enough to express the color distribution of the image.

(4) color aggregation Vector

The core idea is to divide the pixels of each stock in the histogram into two parts. If the area occupied by certain pixels in the stock is larger than the given threshold, then, the pixels in the region are used as aggregate pixels, otherwise they are used as non-aggregated pixels.

(5) color correlation Diagram

Binary texture features

(1) features: texture features are also global features, which also describe the surface properties of the scene corresponding to the image or image area. However, because texture is only a feature of the surface of an object and cannot fully reflect the essential properties of an object, it is impossible to obtain high-level image content only by using texture features. Unlike color features, texture features are not pixel-based features. They need to be calculated statistically in areas containing multiple pixels. In pattern matching, this regional feature has great superiority and won't be able to match successfully due to local deviation. As a statistical feature, texture features often have rotation immutability and have strong resistance to noise. However, texture features also have their disadvantages. One obvious drawback is that when the image resolution changes, the calculated texture may have a large deviation. In addition, due to the potential impact of illumination and reflection, the texture reflected in the 2-D image is not necessarily the real texture of the surface of the 3-D object.

For example, the reflection in the water and the impact of smooth metal surface mutual reflection will lead to texture changes. Because these are not the characteristics of objects, texture information is sometimes "misleading" when applied to search ".

Texture features are an effective method for retrieving texture images with large differences such as width and density. However, when there is little difference between easy-to-distinguish information, such as the width and density of textures, texture features often cannot accurately reflect the differences between human visual senses and textures.

(2) Common Feature Extraction and matching methods

Classification of texture feature description methods

(1) A typical statistical method is a texture feature analysis method called the gray-scale symbiotic matrix. gotlieb and kreyszig, based on the study of various statistical features in the symbiotic matrix, through experiments, the four key features of the gray-scale symbiosis matrix are obtained: energy, inertia, entropy, and correlation. Another typical method in the statistical method is to extract texture features from the image's self-correlation function (that is, the image's energy spectrum function), that is, through the calculation of the image's energy spectrum function, extracting features such as texture fineness and directionality

(2) Geometric Method

The so-called geometric method is a texture feature analysis method based on the texture elements (basic texture elements) theory. According to the texture primitive theory, a complex texture can be composed of several simple texture elements which are arranged in regular order. There are two kinds of algorithms that affect the comparison in the geometric method: voronio board feature method and structure method.

(3) Model Method

The model method is based on the Image Construction Model and uses the model parameters as texture features. A typical method is the random field model method, such as the Markov (Markov) Random Field (MRF) model method and the Gaussian random field model method.

(4) Signal Processing

Texture Feature Extraction and matching mainly include: gray-level co-occurrence matrix, Tamura texture features, self-regression texture model, and wavelet transformation.

The Feature Extraction and matching of the gray level co-occurrence matrix mainly depends on four parameters: energy, inertia, entropy, and correlation. Based on Human Visual Perception psychology of texture, Tamura texture features propose six attributes: roughness, contrast, direction, line image, normalization, and rough. Simultaneous Auto-regressive (SAR) is an application example of the MRF model.

Tri-shape features

(1) features: Various search methods based on shape features can effectively use the objects of interest in the image for retrieval. However, they also have some common problems, including: ① At present, there is still a lack of comprehensive mathematical models for shape-based search methods; ② the search results are often unreliable if the target has deformation; ③ many shape features only describe the local properties of the target, to fully describe the target, it usually requires a high computing time and storage capacity. ④ the target shape information reflected by many shape features is not exactly the same as the human's intuitive feeling, or, the similarity of the feature space is different from that of the human visual system. In addition, the 3-D objects displayed in the 2-D image are only the projection of objects on a certain plane of the space. The shape reflected in the 2-D image is often not
The real shape of a 3-D object may produce various distortion due to changes in the viewpoint.

(2) Common Feature Extraction and matching methods

I typical shape feature description methods

Generally, shape features have two types of Representation Methods: contour features and regional features. The contour feature of an image mainly targets the outer boundary of an object, while the regional feature of an image is related to the entire shape area.

Several typical methods for describing shape features:

(1) Boundary feature method this method obtains image shape parameters by describing boundary features. Among them, the methods for detecting parallel lines and the histogram of the boundary direction are classic methods. Using the global features of the image, we can connect edge pixels to form a region Closed Boundary. The basic idea is the parity of point-line; the boundary direction histogram method first obtains the edge of the image from a differential image, and then generates a histogram of the edge size and direction. The general method is to construct the Gray Gradient Direction matrix of the image.

(2) Fourier shape descriptor Method

The basic idea of Fourier shape deors is to use the Fourier transform of the Object Boundary as the shape description, and convert the two-dimensional problem into one-dimensional problem by using the closeness and periodicity of the region boundary.

Three shape expressions are derived from the boundary points, namely the curvature function, the center distance, and the complex coordinate function.

(3) geometric parameter method

Shape expression and matching use a simpler method of regional feature description, for example, a shape factor (such as a moment, area, perimeter, etc.) related to shape quantitative measurement ). In the QBIC system, geometric parameters such as roundness, eccentric heart rate, spindle direction, and algebraic moment are used to search images based on shape features.

It should be noted that the extraction of shape parameters must be based on image processing and image segmentation. The accuracy of the parameters must be affected by the segmentation effect, shape parameters cannot be extracted.

(4) Shape Invariant Moment Method

Use the moment of the target region as the shape description parameter.

(5) Other methods

In recent years, work on shape representation and matching also includes the finite element method (finite element method or FEM), rotation function (Turning) and wavelet descriptor (wavelet deor.

Ⅱ Shape Feature Extraction and Matching Based on Wavelet and relative moment

This method first obtains the multi-scale edge image by using the wavelet transform modulus maximum. Then seven immutations of each scale are calculated and converted to 10 relative moments, the relative moment on all scales is used as the image feature vector to unify the region and closed and non-closed structures.

Four spatial relationship features

(1) features: the so-called spatial relationship refers to the relationship between the space locations or relative directions of multiple targets in the image, these relationships can also be divided into connection/adjacent relationship, overlapping/overlapping relationship, and inclusion/inclusive relationship. Space Location Information can be divided into two types: relative space location information and absolute space location information. The previous relationship emphasizes the relative situation between targets, such as the upper-lower-left relationship. The latter relationship emphasizes the distance and azimuth between targets. Obviously, relative spatial locations can be introduced from absolute spatial locations, but it is often relatively simple to express relative spatial location information.

The use of spatial relationship features can enhance the ability to differentiate the image content, but spatial relationship features are often sensitive to image or target rotation, reversal, scale changes, and so on. In addition, in practical applications, it is often not enough to use only spatial information, and scene information cannot be effectively and accurately expressed. In order to search, in addition to using spatial link features, other features are also required.

(2) Common Feature Extraction and matching methods

There are two ways to extract spatial relationship features: one is to automatically split the image and divide the objects or color areas contained in the image, then, the image features are extracted based on these regions and indexed. The other method is to divide the image evenly into several rule sub-blocks, and then extract features from each image sub-block, and create an index.
The pose estimation problem is: determining the orientation of a three-dimensional target object. Pose estimation is applied in Robot Vision, motion tracking, single camera calibration, and many other fields.

Sensors Used for pose estimation in different fields are different. Here we mainly talk about visual pose estimation.

Visual pose estimation based on the number of cameras in use can be divided into single-camera visual pose estimation and multi-object visual pose estimation. Different algorithms can be divided into Model-Based Attitude Estimation and learning-based Attitude Estimation.

A Model-Based Attitude Estimation Method

The model-based method usually uses the Geometric relationship of an object or the feature points of an object to estimate. The basic idea is to use a geometric model or structure to represent the structure and shape of an object. By extracting features of some objects, a corresponding relationship is established between the model and the image, then, we use ry or other methods to estimate the spatial posture of an object. The model used here may be either a simple ry, such as a plane, a cylinder, or a geometric structure, or a 3D model obtained by laser scanning or other methods.

The model-based pose estimation method compares the real image with the merged image for similarity calculation and updates the object pose. Currently, the model-based method first degrades the optimization problem into the matching problem of multiple local features to avoid optimizing search in the global state space, it relies heavily on accurate detection of local features. When the noise is too large to extract accurate local features, the robustness of this method is greatly affected.

2. Learning-Based Attitude Estimation

With the help of the machine learning method, the learning-based method is used to learn the ing between two-dimensional observation and three-dimensional posture in training samples with different attitudes obtained first, apply the learned decision rules or regression functions to the samples and use the obtained results as the pose estimation of the samples. The learning-based method generally adopts global observation features, and does not need to detect or recognize the local features of objects, which is highly robust. The disadvantage is that the dense sampling required for continuous Estimation in a high-dimensional space cannot be obtained, so the accuracy and continuity of Attitude Estimation cannot be ensured.

The learning-based attitude estimation method is derived from the idea of attitude recognition. Pose recognition requires that multiple pose categories be defined in advance, each category contains a certain pose range, and then several training samples are marked for each pose category, the model classification method is used to train the pose classifier for Gesture Recognition.

This method does not need to model objects. Generally, it performs matching analysis based on the global features of the image, this can effectively avoid the ambiguity of the local feature method in feature matching when the posture and occlusion are complex. However, the pose recognition method can only divide the pose into several pose categories defined in advance, and it cannot continuously and accurately estimate the posture.

The learning-based method generally adopts global observation features, which can ensure the robustness of the algorithm. However, the attitude estimation accuracy of this method depends largely on the Training adequacy. To accurately obtain the ing between two-dimensional observation and three-dimensional attitude, sufficient intensive samples must be obtained to learn decision-making rules and regression functions. In general, the number of samples required increases exponentially with the dimension of the state space. For a high-dimensional state space, in fact, it is impossible to obtain the intensive sampling required for accurate estimation. Therefore, it is impossible to obtain intensive sampling and ensure accuracy and continuity of estimation. It is a fundamental difficulty that the learning-based Attitude Estimation Method cannot overcome.

Unlike typical pattern classification problems such as pose recognition, pose estimation outputs a high-dimensional pose vector, rather than a class label of a certain category. Therefore, this method requires learning a ing from a high-dimensional observation vector to a high-dimensional attitude vector. At present, this is still a very difficult problem in the machine learning field.

Features are the best way to describe the pattern, and we usually think that each dimension of a feature can describe the pattern from different angles. Ideally, dimensions are complementary and complete.

The main purpose of feature extraction is to reduce dimensionality. The main idea of feature extraction is to project the original sample to a low-dimensional feature space to obtain the low-dimensional sample features that best reflect the nature of the sample or distinguish the samples.

General image features can be divided into four categories: intuitive features, gray statistical features, conversion coefficient features and algebraic features.

Intuitive features mainly refer to geometric features. geometric features are relatively stable and are not easy to extract because of Face attitude changes and illumination conditions, and the measurement accuracy is not high, it is closely related to image processing technologies.

Algebraic features are extracted based on statistical learning methods. The algebraic features have high recognition accuracy. The Algebraic Feature extraction methods can be classified into two types: linear projection feature extraction method and non-linear feature extraction method.

Traditionally, the feature extraction method obtained based on Principal Component Analysis and Fisher linear discriminant analysis is collectively referred to as linear projection analysis.

The basic idea of the feature extraction method based on linear projection analysis is to find a first-line transformation based on certain performance goals and compress the original signal data into a low-dimensional space, this makes the data distribution in the subspaces more compact, provides a means for better data description, and greatly reduces the computing complexity. In linear projection analysis, PCA and LDA are the most representative, the Feature Extraction Algorithm Based on these two methods has become the most classic and widely used method in the pattern recognition field.

The main disadvantage of linear projection analysis is that it needs to learn a large number of existing samples and is sensitive to positioning, illumination and non-linear deformation of objects. Therefore, the collection conditions have a great impact on the recognition performance.

Nonlinear Feature Extraction method is also one of the hot topics of research. "Core Technique" was first used in SVM, and kpca and kfa are the promotion and application of "Core Technique.

The basic idea of the kernel projection method is to transform the samples in the original sample space to a high-dimensional or even infinite-dimensional space through some form of Nonlinear ing, the kernel technique is used to solve the problem by linear analysis in the new space. Because the linear direction in the new space also corresponds to the non-linear direction of the original sample space, the Projection Direction obtained from the kernel-based projection analysis also corresponds to the non-linear direction of the original sample space.

The kernel projection method also has some weaknesses: the geometric meaning is unclear, and it cannot know the distribution mode of the sample after the non-explicit ing. The selection of parameters in the kernel function has no corresponding selection criteria, most training samples can only be selected using empirical parameters. This is not suitable for many training samples because the dimension of the sample is equal to the number of training samples after the kernel ing. If the number of training samples is large, the vector dimension after the core ing will be very high and will encounter computing difficulties.

In the application field, kpca is far from widely used in PCA. If the common dimension reduction kpca is better than PCA, it is more obvious that the feature space is not a general European space. PCA can learn a sub-space through a large number of natural images, but kpca cannot.

Transform coefficient feature refers to performing Fourier transform and wavelet transform on the image first, and then recognizing the obtained coefficient as the feature.