Using the computer to carry on the automatic extraction of remote sensing information must use digital images, due to the features in the same band, the same feature in the spectrum of different wavelengths have different characteristics, based on a certain feature in each band spectrum curve is analyzed, according to the characteristics of the corresponding enhancement processing, can be similar to identify and extract the target on the remote sensing image, the early automatic classification and image segmentation is mainly based on the spectral characteristics of later development for combining spectral characteristics, texture feature and shape feature, spatial relation characteristics such as combination of computer information extraction,
1.2.1 automatic classification
Commonly used method of information extraction is a computer automatic classification of remote sensing image, first of all, the remote sensing image interpretation, indoor and field investigation, and aims to build various types of ground objects and the corresponding relationship between image feature and validate the result of the indoor anticipation, after work into indoor, choose the training sample and carries on the statistical analysis, with the appropriate classifier for remote sensing data classification, the classification results are the post-processing, the final precision evaluation, remote sensing image classification is generally based on the spectral features of ground objects shape characteristics, spatial relations features and so on, at present most studies are based on spectral characteristics,
Before computer classification, often have to do some preprocessing, such as correction, enhancement, filtering and so on, in order to highlight the target characteristics or eliminate the different parts of the same type target because of different exposure conditions, terrain change, scanning observation Angle caused the different brightness difference, such as
By using remote sensing image classification, that is, to a single like yuan or more homogeneous like a tuple corresponds to its characteristics of the name is given, its principle is to use the image recognition techniques to realize automatic classification of remote sensing image, the computer to identify and classify the main sign is the spectral characteristics of the object, the image of the other information such as size, shape, texture, etc has yet to make full use of,
Computer image classification method, there are two kinds of common, namely, supervised classification and unsupervised classification, supervised classification, first of all, from the desire classification in the area of the image area selected some training samples, in such training area feature category is known, and use it to establish classification standards, and then the computer will be effected according to the same standard to the entire image identification and classification, it is a kind of by the known samples, extrapolating unknown categories of methods; Unsupervised classification is a kind of no a priori (known) category standard classification method, and the research object and area, no known classes or training samples for standard, but the use of image data itself to the characteristics of the gathered crowds in characteristics measurement space, to form the first data set, and then check these data sets, represented by the object category,
Compared with supervised classification and unsupervised classification has the following advantages: no need to have prior knowledge of the area studied, and the results of classification accuracy under the same conditions, more save on time and cost, but in fact, unsupervised classification and supervised classification is better than high precision, so the supervised classification used more widely,
1.2.2 texture feature analysis
Tiny terrain is repeated regularly in the image, it reflects the tonal change frequency, texture form a lot, including spot, spot, and curves, gate, on the basis of these forms according to the thickness, density, width, length, straight Angle and sympathetic conditions also can be subdivided into more types, each type of object has itself on the image texture pattern, therefore, can from the image of the character recognition feature, texture reflects the brightness changes (gray) space, there are three main signs: a partial sequence of sex within the area is larger than the sequence repeats itself; Sequence is composed of basic part of the random arrangement; The unity of the parts are roughly evenly, anywhere in the area of the texture have roughly the same structure size, the sequence is often referred to as the basic part of the texture primitives, so you can think texture is by some kind of uncertainty by primitive rule or statistical rule of arrangement, the former is called uncertainty texture (such as artificial texture), the latter is the random texture (or natural texture), description of the texture can be through the thick degree of texture, smoothness, granularity, randomness, directivity, linearity, periodic, repeatability of these qualitative or quantitative characterization of the concept of features,
Corresponding number of texture feature extraction algorithm can be divided into two categories, namely structure method and statistical method, the texture structure method as the basic texture yuan according to the arrangement of the specific rules of cyclical repetitive patterns, so often based on the traditional Fourier spectrum analysis method to determine the texture and arrangement rules, moreover structure yuan statistics and texture analysis of grammar is also a common method of extraction, the structure method of irregular texture in the extraction of natural landscape when they meet with difficulties, the texture is very difficult to by repeated texture yuan, said yuan and texture extraction and arrangement rules express itself is a very difficult question, for the most part in remote sensing image texture is randomness, obey statistical distribution, generally USES the statistical texture analysis, currently use more methods include: co-occurrence matrix, the fractal dimension method, markov random field method and so on, the co-occurrence matrix is a more traditional texture description methods, it can describe the image texture feature from multiple sides,
1.2.3 image segmentation
Image segmentation is the image into each area and extract the characteristic of interested target technology and process, the feature can be pixel gray scale, color, texture and other predefined goals can correspond to a single area, also can correspond to multiple regions,
Image segmentation is a key step from image processing to image analysis, occupy the important position in image engineering, on the one hand, it is the foundation of target expression, has important influence on feature measurement; Because, on the other hand, image segmentation and target expression based on segmentation, feature extraction and the parameter measurement of the original image is transformed into more abstract and more compact form, making it possible to higher level of image analysis and understanding,
Image segmentation is the basis of image understanding, image segmentation and rely on image understanding and in theory, is closely related to each other, and under the general image segmentation is a very difficult problem, in the early period of the current general image segmentation as the image processing stage, is the technology for object segmentation, is related to the problem, such as the use of the most commonly used to image segmentation threshold processing,
Image segmentation are of three different kinds of ways, one is to each pixel into the object or area of the corresponding pixel clustering methods namely area method, the other is done by directly determine the regional boundary integral boundary method, the third is the first detection of edge pixels and edge pixels to connect a boundary formation division,
1) threshold and image segmentation
Threshold is in at the time of the partitioning as the difference between object and background pixel threshold, is equal to or greater than the threshold value of pixels belongs to object, and the other belongs to the background, the method for the obvious difference between the object and the background scenery very effective segmentation (contrast), in fact, in the image processing system of any practical application, are used threshold technology, in order to effectively object segmentation and background, people developed a variety of threshold processing technology, including the global threshold, adaptive threshold, the optimal threshold, etc.,
2) gradient and image segmentationWhen the object has obvious contrast with the background, the boundary of the highest point in the image gradient, by tracing the image with the highest point of the gradient of the way to obtain the boundary of the object, can realize the image segmentation, this method is easily affected by noise and deviated from the object boundary, usually need to be done before the tracking of the gradient image smoothing processing, such as boundary search tracking algorithm is used to realize again,
3) boundary extraction and contour tracking
In order to obtain the image edge people many kinds of edge detection method is proposed, such as Sobel and Canny edge, the LoG, the edge image, on the basis of the need to smooth, morphology processing such as noise points, burr, empty and don't need the parts, such as through refinement, edge connectivity and tracking method so as to obtain the outline of the object boundary,
4) the Hough transform to
For image in some conform to the dominant characteristic of parameter model, such as straight line, circle, ellipse, etc., can be based on the parameters of clustering, the method of extraction, the characteristics of the corresponding
5) regional growth
A region growing method is according to the same object of pixels in the area of a similar nature to gather the pixel point method, from the initial area (such as a small neighborhood or even each pixel), will have the same properties of adjacent pixels or other region merge into the present area so as to gradually increase in the area, until can't merge the points or other small region, the regional similarity metric can include the average grey value of pixels, texture, color and other information,
Region growing method is a common method, without the prior knowledge to use, can achieve the best performance, can be used to more complex image segmentation, such as natural scenery, but, region growing method is an iterative method, space and time cost are large,
1. The object-oriented remote sensing information extraction
Based on pixel level of information extraction based on a single pixel unit, too much focus on local rather than the entire map spot near the geometric structure, thus seriously restricted the accuracy of information extraction, and object-oriented remote sensing information extraction, considering the characteristics of spectrum, shape, size, texture, adjacent relation and a series of factors, thus has higher precision of classification result, object-oriented remote sensing image analysis technique for image classification and information extraction method is as follows:
First to image segmentation of image data, from a 2 d array of image information in recovering the image of the landscape in the scene reflected by the target object space shape and combination, the smallest unit of image is no longer a single pixel, but one object, the subsequent image analysis and processing based on object,
Then USES the fuzzy classification algorithm of decision support, each object simply does not simply be assigned to a category, but given the probability of each object belongs to one kind of, making it easy for users to adjust according to actual condition, at the same time, can also be produced according to the maximum probability to determine the classification results, when establishing expert decision support system, set up the classification of different scale levels, at every level definition of object spectrum characteristics respectively, shape features, texture features and adjacent relations, among them, the spectral characteristics including the mean, variance, the gray level ratio; Shape features include area, length, width, length of the boundary, aspect ratio and shape factor, the density, the main direction, symmetry, position, for linear features including line length, line width, line length-width ratio, the curvature, curvature and the ratio of the length, etc., for planar features including area, perimeter, compactness, polygon number of edges, and the length of each side of the variance, and the average length of each side, the length of the longest edge; Variance texture characteristics including object, area, density, symmetry, and the main direction of the mean and variance, etc., and assign different weights by defining a variety of features, classification standard, then the image classification, classification, first on a large scale "parent", according to the actual need for defining characteristic features of interest in a small scale, cent gives "subclass",
CodePudding user response:
A good theory textbooksCodePudding user response:
How do you don't have a formula?