RU2313828C2 - Image recognition method - Google Patents

Abstract

FIELD: methods for automatic decision making in informational and controlling systems of robot-engineering complexes, in particular, technical vision systems and television systems for detecting, tracking and recognizing objects, may also be used in automated systems for controlling situation of various purposes.

SUBSTANCE: image recognition method additionally includes operation for detecting local extremes, while a priori unknown structure of image being recognized, presence and location of its component objects, are determined not by using preliminarily given object standard, but on basis of analysis of characteristics of a set of local extremes detected on image signal being analyzed.

EFFECT: expanded area of application of recognizing informational machines to include cases of a priori indeterminacy of characteristics of objects being analyzed and conditions of observation of those objects.

9 dwg

Description

FIELD OF THE INVENTION

The invention relates to the fields of information television technology, robotics, pattern recognition, image recognition, television detection and tracking systems, television systems for automated image analysis and object recognition and can be used in vision systems and automated monitoring of the situation for various purposes.

State of the art

The image recognition methodology used in the construction of technical systems is focused primarily on the classification of objects. The classification operation is interpreted in this case as identifying the object's belonging to one of the predefined classes, each of which is described using the reference values of formalized signs present and highlighted in images and signals. The implementation of appropriate recognition methods requires a preliminary statistical analysis of the properties of objects under the assumed conditions of observation and the formation of reference descriptions of recognizable classes of objects specific to each applied recognition task.

At the same time, in practical recognition problems, there is the so-called a priori uncertainty problem, consisting in the fact that objects of any classes are represented by an almost unlimited set of realizations, in which none of the microstructural properties of the analyzed signal turns out to be a sufficiently reliable sign either to isolate the object or to its identification. As applied to image analysis, such conditions arise for the following reasons. First, the analyzed objects have to be distinguished not by signal intensity, but by complex geometric and topological parameters of the form type. Secondly, there is always a large structural diversity of the analyzed objects, for example, when recognizing characters - these are different alphabets and fonts; during production robotization, this is a wide range of parts and operations with them; in surveillance and environmental control systems, these are all kinds of types of equipment, people, industrial facilities and communications. The third factor creating a priori uncertainty in the tasks of automatic image analysis is the unlimited variability of the conditions for observing objects in terms of scale, angles, underlying surfaces, and lighting conditions.

Thus, the conditions of a priori uncertainty limit the use of recognition systems with specific a priori standards of objects. To overcome these limitations, the so-called structural-linguistic recognition methods are developed, in which the simplest typical image elements — primitives — that make up the images of objects are detected.

Analogs of the claimed technical solution are, for example, a method and device for recognizing objects according to the USSR copyright certificate No. 1697533, 1990, IPC G06K 9/00; Method and device for object recognition according to the patent of Germany No. 03327445, IPC G06F 15, G06K 9/00.

Closest to the claimed technical solution, selected as a prototype, is a hardware-software complex for the analysis of biological products according to Russian certificate No. 16628 for utility model dated January 20, 2001, IPC 7 G01N 33/48, G06K 9/00.

The recognition method implemented in the prototype consists of a sequence of operations for normalizing the input image signal, detecting on the normalized input signal the regions as structural elements of the image, measuring the characteristics of the detected regions and classifying the detected regions as the desired objects. In this case, as the simplest image areas displaying the desired and analyzed objects, such areas-areas of the normalized image that are formed by adjacent ones, i.e. spatially connected, pixels having brightness values (image signal intensities) in a given range.

The main disadvantage of the prototype method is the presence of an error in detecting image elements, or image areas that represent the analyzed objects. This disadvantage is manifested in the mismatch of the detected areas with the actual images (silhouettes) of objects and / or their informative fragments. This error is the greater, the more complex the plot and the greater the mismatch between the optimal observation conditions for the given data and the actual values of the parameters of the operations of processing the input signal.

The specified disadvantage of the prototype is due to the fact that the detection of areas of objects in it is carried out by spatially connected values of the normalized input signal. In this case, the correctness of the silhouettes of objects falling into the luminance ranges processed by the connection depends on the parameters of the quantization operations (discretization of the brightness level) and spatial filtering during normalization of the input image.

Disclosure of invention

The objective of the invention is to expand the scope of the recognition method for cases of fundamental a priori uncertainty by reducing the dependence of the reliability of detection of image elements constituting images of objects and / or their informative fragments on the values of the parameters of the normalization operation of the input signal and other processing.

The problem is achieved by the fact that in the image recognition method containing the operation of normalizing the input signal, the operation of quantizing the signal, the operation of extracting from the connected areas and the classification operation, an additional operation for detecting local extremes, an operation for generating an image description, an operation for generating an auxiliary signal and an operation for grouping regions are introduced . In this case, the detection of local extrema is performed on the input image signal and on the normalized input signal by independently comparing the values of neighboring samples in each signal, and the signal of the detected local extrema is converted and stored by the image description generation operation. The generated image description signal is then used in generating an auxiliary signal for non-linear conversion of a linear combination of the input image signal and the normalized input signal. Next, the generated auxiliary signal is quantized and the quantized signal is converted by analyzing the spatial connectivity of its neighboring samples into a signal of the selected areas, which are converted and stored by the image description generating operation. Also, the signal of the selected areas is converted by statistical and logical analysis into a signal of groups of areas, which are also converted and stored by the operation of generating an image description. The signals of the connected regions and groups of regions are formed using the signals of the input image and the generated image description. Also, the signals of connected regions and groups of regions and the input image signal are converted by a classification operation by comparison with the reference signals into a classification signal, which is also converted and stored by the operation of generating an image description.

The technical result achieved in this case in comparison with the prototype is as follows:

- the set of detectable local extrema actually reveals and describes the information structure of the image, recognizing it by determining the centers and boundaries of all texture heterogeneities present in the image, i.e. desired primitives;

- the structure of the set of detected local extrema, at least in terms of the boundaries of the structural elements of the image, does not depend on the absolute brightness values of the primitives and the shape of the brightness distribution over their space. Moreover, data on local extremes can themselves be directly used to improve the quality of the original image. This eliminates the dependence of recognition reliability on normalization parameters, i.e. ways of preprocessing the analyzed image and increasing the signal-to-noise ratio.

The introduced operation of detecting local extremes provides additional information independent of the absolute brightness values for marking the boundaries of the desired structural image elements or regions of objects, necessary for subsequent processing of the image by the operation of identifying regions of objects by the condition of spatial connectivity.

Brief Description of the Drawings

Figure 1 shows a block diagram of a device that implements the proposed method.

Figure 2 shows an embodiment of a block for detecting local extremes.

Figure 3 shows an embodiment of a normalization block of the input signal.

Figure 4 shows an embodiment of an image description forming unit.

Figure 5 shows an embodiment of the auxiliary signal generating unit.

Figure 6 shows an embodiment of a block of quantization of a signal.

Figure 7 shows an embodiment of a block allocation by connecting areas.

On Fig shows an embodiment of a block grouping areas.

Figure 9 shows an embodiment of a classification block.

In figure 1 is indicated:

1 Local extremum detection unit,

2 Block normalization of the input signal,

3 Block forming the image description,

4 Block forming an auxiliary signal,

5 Block quantization signal

6 Block allocation by connected areas,

7 Area grouping unit,

8 Classification block.

In figure 2 is indicated:

Image input signal,

Normalized input

9 analog to digital converter,

10 analog to digital converter,

11 storage device

12 Comparator,

13 Comparator,

14 Comparator,

15 logical unit,

16 storage device.

In figure 3 is indicated:

17 filter unit,

18 differentiation unit,

19 storage device

20 Amplifier unit.

21 Adder.

In figure 4 is indicated:

22 Converter,

23 storage device

24 storage device,

25 storage device,

26 storage device,

27 Converter.

Figure 5 is indicated:

28 Converter,

29 Converter,

30 Amplifier,

31 Amplifier

32 adder.

In Fig.6 indicated:

33 Comparator,

34 storage device.

In Fig.7 indicated:

35 storage device,

36 Numbering logic block,

37 storage device

38 Checksum verification unit,

39 The calculator characteristics of the areas

40 storage device.

On Fig indicated:

41 storage device

42 block statistical analysis,

43 Block logical combination,

44 The calculator characteristics of the groups of regions,

45 storage device.

In Fig.9 indicated:

46 storage device

47 storage device

48 Comparator,

49 Block sampling standards.

The implementation of the invention

The proposed method can be implemented in the form of a device, a block diagram of which is presented in figure 1. The device consists of a series connection of a local extremum detection unit 1, an image description generating unit 3, an auxiliary signal generating unit 4, a signal quantization unit 5, an area connectivity unit 6, an area grouping unit 7, and a classification unit 8, to the second input of which an output is connected block 6 allocation of connected areas, and the output of block 8 classification is connected to the second input of block 3 of the formation of the image description. The device also includes an input signal normalization unit 2, the output of which is connected to the second input of the local extreme detection unit 1 and to the second input of the auxiliary signal generating unit 4, and the input signal normalization unit 2 is connected to the input of the local extreme detection unit 1, to which an input image signal is supplied and which is the input of the device, and is also connected to the third input of the auxiliary signal generating unit 4, to the second input of the isolation allocation unit 6 areas and to the second input of block 7 grouping areas. Moreover, the output of the image description generating unit 3 is also connected to the third input of the area grouping unit 6 and the third input of the area grouping unit 7, the output of which is also connected to the third input of the image description forming unit 3, and the output of the area connecting unit 6 is connected also to the fourth input of the image description generating unit 3.

A device that implements the proposed method works as follows.

The input image signal supplied to the input of the device and the normalized input signal generated in the input signal normalization unit 2 are converted in the local extremum detection unit 1 into a local extremum signal, which is stored in the image description generating unit 3. Then, in the auxiliary signal generating unit 4, a primary auxiliary signal is formed from the input image signal, the normalized input signal, and the local extremum signal. The primary auxiliary signal is then quantized in level in the signal quantization unit 6, and thus, binary secondary auxiliary signals are obtained. Secondary auxiliary signals in the block 6 allocation by connectivity areas are converted according to the rule of spatial connectivity into a signal of a numbered silhouette image, the readings of which are represented by the number of the luminance region (i.e. the structural element of the image), which includes this sample. Here, in block 6, the selection by connectivity of the regions, for each region thus selected, the brightness, geometric and topological characteristics, including shape indicators, are calculated by the input image signal. The characteristics signal is stored in the image description generating unit 3, and also converted in the area grouping unit 7 into a signal of the area groups based on a logical and statistical analysis of the signal parameters of the regions identified by the connectedness. Here, in block 7 grouping areas, similarly as in block 6, the brightness, geometric and topological characteristics are calculated by the input image signal, including shape indicators, revealed groups of areas. The signal of the characteristics of the groups of regions is also stored in the block 3 forming the image description. At the final stage of recognition of the input image signal, in block 8 of the classification, the comparison of the signals of the characteristics of the regions and groups of regions identified by connectivity with the reference signals of object classes is generated and a signal for evaluating the similarity of the regions and groups of regions identified by connectivity to the object classes specified by the reference signals is generated. The similarity rating signal is also stored in the image description generating unit 3.

The blocks of the device that implements the proposed method, operate as follows.

In the block 1 for detecting local extremes, the implementation of which is shown in Fig. 2, the input image signal and the normalized input signal are each supplied, respectively, to analog-to-digital converters 9 and 10, the output signals of which are stored in the storage device 11. Then, for each counting of stored digital images by comparators 12, 13 and 14, the number of neighboring samples is detected, respectively, with larger, smaller and equal signal values. According to these parameters, in the logical block 15, an evaluation signal of the type of local extremum is generated in the analyzed sample, which is stored in the storage device 16. Thereby, data are generated on the local structural elements — local extrema — that make up the information structure of the input image signal and the objects represented on it.

In block 2 of the normalization of the input signal, the implementation of which is shown in figure 3, the input image signal is converted in the filter unit 17 and the differentiation unit 18 into a series of intermediate signals, which, together with the input image signal and additionally stored signals in the storage device 19, amplify in block 20 amplifiers, each with its own coefficient, and then summed in the adder 21.

In block 3 of the formation of the image description, the implementation of which is shown in figure 4, sequential storage of the detected characteristics of the analyzed input image signal is performed. Moreover, each of the generated local extremum signals, distinguished by the connectivity of regions, groups of regions and a classification signal, is converted in the input converter 22 to a uniform tabular form of a list of image regions and their characteristics, and then stored in memory devices 23, 24, 25 and, respectively 26. The selection of the stored characteristics of the analyzed image is performed using the output transducer 27. In this case, the previously obtained data on the parameters of the input image signal are used To obtain its subsequent parameters, particularly the data on the detected local extrema also fed to block 4 of the auxiliary signal forming unit 6 for allocating areas and connected to the block 7 of grouping regions.

In block 4 for generating an auxiliary signal, an embodiment of which is shown in FIG. 5, a primary auxiliary signal is formed to select image areas by spatial connectivity. The formation of the primary auxiliary signal is performed by amplification by the converter 28 in the input image signal and by the converter 29 in the normalized input signal of those sections that correspond to the specified types of local extrema, by amplification in the amplifiers 30 and 31 of these converted signals and their summing by the adder 32.

In block 5 of the quantization of the signal, the implementation of which is shown in Fig.6, produce the formation of binary secondary auxiliary signals for the direct selection of areas of spatial connectivity. Secondary auxiliary images are formed from the primary auxiliary signal by quantization, i.e. discretization by level in the comparator 33 and their storage in the storage device 34.

In block 6, the selection of connected regions, the implementation of which is shown in Fig.7, converts the quantized signal into a signal of a numbered silhouette image, the readings of which are represented by the number of the region (or element), which includes this sample. For this, secondary auxiliary binary images, or fragments thereof, selected according to the specified characteristics of the signal of local extrema, are alternately recorded in the storage device 35. The samples of the stored binary images are numbered in the logical numbering unit 36 according to the criterion of spatial connectivity and the resulting numbered silhouette image is stored in the memory 37. For the stored numbered image in the checksum verification unit 38, its checksum is calculated and the numbering of this image is repeated by renumbering to repeating the checksum value. The numbered silhouette image obtained in this way is considered to be the result of the current secondary auxiliary binary signal, and for the areas highlighted on it, the brightness characteristics, geometric and topological characteristics of the regions are calculated in the calculator 39 of the regions, including shape indicators and characteristics. The resulting signal of the connected areas and their characteristics is accumulated in the storage device 40. Thereby, data on non-local structural elements — brightness areas — that make up the input image signal and the objects represented on it are generated.

In block 7 grouping regions, an implementation option of which is shown in Fig. 8, the plurality of areas identified by connectivity is divided into subsets (clusters), the elements - regions of which have similar characteristics. For this, samples 41 of previously calculated characteristics of the regions identified by connectivity and detected local extrema are alternately recorded in the storage device 41, their statistical analysis is performed in the statistical analysis block 42, and then, in the logical union block 43, an area membership group (cluster) signal is generated according to a specified criterion. Then, in the calculator 44 of the characteristics of the groups of regions, the same set of characteristics is calculated for the groups of regions — clusters — as for the regions identified by connectedness, and these data are accumulated in the storage device 45. Thereby, data are generated about the cluster structural elements — groups of brightness regions, - constituting the input image signal and the objects represented on it.

In the classification block 8, the implementation variant of which is shown in Fig. 9, the comparator 48 compares the characteristics of the connected regions and groups of regions, which are alternately recorded in the memory 46, with reference descriptions of the specified classes of objects stored in the memory 47 and alternately selected block 49 sample standards. According to the results of the comparison, a similarity assessment signal is generated in the comparator 48, which is also stored in the image description generating unit 3.

Thus, in the process of image recognition by the proposed method, a step-by-step description of its information structure is formed in the form of a list of elements (or areas-objects) making up the image with their characteristics. A primary approximate description of the structure of the recognized image is obtained immediately after the detection of local extremes. At the next stages of image recognition, this information structure of the analyzed image is clarified by more accurately identifying connected areas of objects and measuring the full set of their characteristics present in the input image, as well as identifying clusters of areas. The revealed informational structure of the input image is supplemented by data on the similarity of the identified regions and clusters of regions to given classes of objects.

Claims (1)

A method for recognizing image signals, comprising an operation for normalizing an image signal, an operation for quantizing a signal, an operation for detecting connected areas, a classification operation, characterized in that an operation for detecting local extremes, an operation for generating an image description signal, an operation for generating an auxiliary signal and an operation for detecting groups of regions are introduced, this successively generates a normalized image signal, a signal of detected local extremes, a description signal images, an auxiliary signal, a quantized signal, a signal of detected connected regions, a signal of detected groups of regions, a classification signal and an image description signal, which is formed by storing the signal of the detected local extremes first, then storing the signal of the detected connected regions, then storing the signal of the detected region groups and then storing classification signal, while the detection of local extremes is performed on the image signal and on the normalized signal from brazings by means of independent comparison of values of neighboring samples in each signal, and the signal of detected local extrema is formed as a list of detected local extrema and their characteristics, which are calculated on the corresponding signal for each detected extremum; the auxiliary signal is formed as a linear combination of an image signal and a normalized image signal, which is subjected to non-linear transformation under the control of the image description signal; the auxiliary signal is quantized and the detection of the connected regions is performed on the quantized signal under the control of the image description signal, and the characteristics of the connected regions are calculated on the image signal; the detection of groups of regions is carried out under the control of the image description signal and by identifying connected areas with similar characteristics in limited areas of the image signal, the signal of the detected groups of regions is formed as a list of the detected groups of regions and their characteristics, which are calculated for each detected group of regions on the image signal and on the signal Connected areas detected Further, the signal of the detected connected regions, the signal of the detected group of regions, and the image signal are compared with the reference signals, a similarity score of the compared signals is calculated, and a classification signal is generated, which is also stored by the operation of generating the image description signal.

Images