JPH06187454A - Initial visual device - Google Patents

Abstract

PURPOSE:To provide an initial visual device which extracts the effective feature value to recognize an object included in a texture image. CONSTITUTION:This initial visual device consists of an image input means 1 which fetches the image information to the visual device, a Fourier transformer means 2 which devides the image information into plural areas and transforms the image information on each area into the frequency component information included in the image information on each area, a frequency component detector means 3 which detects the frequency component information on each transformed area, a frequency component deciding means 4 which decides a characteristic frequency component of each area from the frequency component information on the area, a frequency component selector means 5 which decides the characteristic frequency components of an entire image from the characteristic frequency component on each area and produces a signal that selects successively the frequency components of the entire image, and an output switch means 6 which receives the signal produced by the means 5 and successively switches the characteristic frequency components of each area sent from the means 4 to output these frequency components of the following recognition processing, etc.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an initial visual device for extracting a feature amount effective for recognizing an object in a texture image.

[0002]

2. Description of the Related Art Conventionally, in order to recognize an object in an image,
A method is often used in which a portion of the image where the contrast changes abruptly, that is, a boundary portion of light and dark is extracted as a feature amount and is used for recognition. This mimics the processing method used by human early vision.

FIG. 10 is a diagram for explaining the human initial visual mechanism, FIG. 11 is a diagram showing the excitement intensity of each cell at the light-dark boundary in the initial visual mechanism, and FIG. 12 is a conventional example of the initial visual mechanism by an electric circuit. FIG.

The human initial visual mechanism is composed of photoreceptor cells 101, horizontal cells 102 and bipolar cells 103, as shown in FIG. The mechanism is briefly described. First, the photoreceptor cells 101 receive light from the image, and only the photoreceptor cells 101 in the portion that receives the light are excited. This excitement is transmitted to the lower horizontal cells 102.
It is planarly connected to about six adjacent horizontal cells 102, and due to the transmission through this connection, the surrounding horizontal cells 10 are connected.
The excitement of 2 changes gently as it travels down the network. Further, the bipolar cell 103 is connected to the photoreceptor cell 101 and the horizontal cell 102, and excites corresponding to the difference in excitement intensity of the two cells. This excitement is further transmitted to the lower layers, and finally reaches the brain for object recognition and feature extraction.

In fact, considering the boundary of light and darkness such as the outline of an object as an image, in this initial visual mechanism, the excitement intensity of the photoreceptor cell 101 changes rapidly as shown in FIG. 11 (a). The excitation intensity of the horizontal cell 102 is shown in FIG.
It changes gently as shown in (b). Therefore, in the bipolar cell 103, corresponding to the difference in excitement intensity between these two cells, only the bright / dark boundary portion is excited exactly as shown in FIG. 11C, and this information is transmitted to the brain.

This mechanism was constructed by an electric circuit and was trial-produced as a neurochip.
A. Mr. Mead and others (CA Mead and MA Mahow
ald, “A Silicon Model of Early Visual Processin
g ”, Neural Networks, vol. 1, pp91-97 (1988)).
As shown in FIG. 12, they use a photodiode 301 for the photoreceptor cell 101 and a six resistance network 302 for connecting the horizontal cell 102 to the center and apex of the hexagon, and
3 is composed of an amplifier 303.

According to this circuit, the voltage of the photodiode 301 changes stepwise in the bright and dark portions of the image. On the other hand, the voltage of the node of the resistance network 302 changes gently. Therefore, by taking this difference, the voltage of the amplifier 303 changes only in the bright / dark boundary portion. The voltage change of these elements is similar to the change of the excitation intensity of each cell shown in FIG. 11, and the light-dark boundary, that is, the outline of the object can be detected as in the case of human initial vision.

[0008]

FIG. 13 is a diagram showing an example of an input image. The above-mentioned conventional example is a method of making good use of the human initial visual mechanism and obtaining light-dark boundary information as a feature amount useful for recognizing an object in an image. Further, it is also a method suitable for integration and miniaturization by utilizing the current LSI manufacturing technology.

However, the light-dark boundary information obtained by this method can be effectively recognized when the object information is included as an image having a clear contrast with the background as shown in FIG. There is a problem in that an image having a small contrast difference with the background as shown in b) and having only a different texture, that is, a so-called texture image cannot be recognized well.

In view of such circumstances, it is an object of the present invention to provide an initial visual device for extracting a feature amount effective for recognizing an object in a texture image.

[0011]

Therefore, as shown in FIG. 1, the initial visual device of the present invention includes a plurality of image input means 1 for taking in image information into the device and a plurality of the input image information. After dividing into areas, Fourier transform means 2 for converting the image information of each area into frequency component information of the image information of each area, and the frequency component information of each converted area are detected. Frequency component detection means 3, a frequency component determination means 4 for determining a characteristic frequency component in the area based on the detected frequency component information for each area, and a characteristic frequency in each area. Frequency component selecting means 5 for determining a characteristic frequency component forming the entire image from the components, and further generating a signal for sequentially selecting the characteristic frequency components forming the entire image; From the output switching means 6 for receiving the signal emitted from the wave number component selection means, sequentially switching the characteristic frequency components for each region sent from the frequency component determination means, and outputting to the next processing such as recognition. It is configured to achieve the object of the present invention.

[0012]

FIG. 2 is a schematic view for explaining the operation of the early vision device of the present invention. In the initial visual device of the present invention, the input image is Fourier-transformed by the configuration as described above, and the characteristic spatial frequency component included in the texture of each object included in the image is used as the feature amount of the object. Since the objects can be classified and further selected and output, it is possible to distinguish and recognize the background or other objects in the image having different textures from the object to be recognized.

Furthermore, when the input image is processed as it is, a plurality of spatial frequency information corresponding to a plurality of textures included in the input image are detected at the same time, so it is not easy to classify and select these. Absent. However, since the apparatus of the present invention is configured to divide the input image into a plurality of regions and process the images by the above-described configuration, if each region is set to a small region to a certain extent, the image shown in FIG.
As schematically shown in (b) and (c), the texture of the object is substantially the same in each region. Therefore, within each area, only the characteristic spatial frequency components corresponding to that texture become prominent, so this can be easily detected, and it becomes very easy to collect, classify, and select this information. .

[0014]

Embodiments of the present invention will be described below with reference to the drawings. [First Embodiment] FIG. 3 is a diagram for explaining the basic components, and FIG. 4 is a diagram for explaining the configurations of the Fourier transforming means and the frequency component detecting means.

First, the basic constituent elements of the Fourier transform means 2 and the frequency component detection means 3 will be described with reference to FIG. This basic component has the image input surface 2a as a substantially flat end surface, and further outputs the Fourier transform information of the image input to the image input surface 2a to the other substantially flat end surface, the Fourier surface 2b. The Fourier transform lens 2c and the detector 3a that are closely arranged on the Fourier surface 2b of the Fourier transform lens 2c and detect the information Fourier-transformed by the Fourier transform lens 2c by dividing the information for each spatial frequency to some extent.

In the present embodiment, specifically, the Fourier transform lens 2c is a gradient index rod lens as shown in FIG. 3A, and the detector 3a is circularly detected as shown in FIG. 3B. It is constructed by using a ring detector with divided regions. Further, if the basic components are arranged two-dimensionally as shown in FIG. 4, the desired Fourier transform means 2 can be configured as an array of Fourier transform lenses 2c and the frequency component detection means 3 can be configured as an array of detectors 3a. it can.

A zoom lens is used as the image input means,
When an image is input to the image input surface of the Fourier transform means 2 (the image input surfaces 2a of the basic constituent elements are two-dimensionally arranged) as described above, this image is obtained by each basic constituent element of the Fourier transform means 2. The image input surface 2a is divided into processing areas. Then, Fourier transform is performed by the Fourier transform lens 2c for each basic component in each region, and the spatial frequency component is detected by the detector 3a in each basic component for each region.

FIG. 5 is a diagram for explaining components after the frequency component judging means, FIG. 6 is a diagram showing a processing example by the initial vision device of the present invention, and FIG. 7 is a diagram showing a modified example of the frequency component detecting means. Is.

Next, the components after the frequency component judging means will be described with reference to FIG. The spatial frequency component of each area detected by the detector 3a constituting the frequency component detecting means 3 by the above-described processing is then sent to the frequency component judging means 4 as shown in FIG. The frequency component judging means 4 is composed of a threshold value means 4a and a threshold value adjusting means 4b. The spatial frequency component detected by the detector 3a in each region is threshold-value operated by the threshold value means 4a. , The strongest spatial frequency component is output as the characteristic spatial frequency component of this region.

Although the threshold value of the threshold value means 4a is set to an appropriate initial value, if the number of outputs from the threshold value means 4a is not one, the threshold value is adjusted. The means 4b judges the number and generates a signal for adjusting the threshold value of the threshold value means 4a (specifically, when the number of outputs is two or more, the threshold value is raised and the output is outputted. If not, lower the threshold). The characteristic spatial frequency component determined by the frequency component determination means 4 is then sent to the frequency component selection means 5.

The frequency component selecting means 5 is an adding means 5
a, a threshold value means 5b and a switching signal generating means 5c. First, characteristic spatial frequency components in each region are sent to the adding means 5a and added for each frequency component. This operation is an operation for classifying and quantifying the spatial frequency components that make up the entire image. This added frequency component information is further sent to the threshold value means 5b, and only the information of the frequency component exceeding the set appropriate threshold value is sent to the next switching signal generation means 5c. This switching signal generating means 5c
Then, the frequency component information that has been threshold-processed and output is ordered from the one with the largest output, and a signal is generated so that the frequency component information with the number 1 is output first. The signal generated here is sent to the next output switching means 6.

The output changeover means 6 is composed of an output changeover switch 6a and an output means 6b. First, the output changeover switch 6a receives this signal and switches the internal switch to output from the frequency component decision means 4. Of the characteristic spatial frequency component information for each area sent to the switching means 6, only the corresponding one is selected and passed.

In the next output means 6b, the output from the output changeover switch 6a is further inputted. For example, when there is an output from the output changeover switch 6a, 1 is outputted and 0 is outputted. . This value is not limited to 0 and 1 and may be given in an appropriate form according to the processing in the subsequent stage. The signal from the output means 6b is further sent to a subsequent process (for example, a neural network, a correlator, etc.) for recognition or the like.

In this state, the switching signal generating means 5c
Since only the frequency component information with the number 1 is output from this initial visual device and is not recognized, the reset signal (not shown) is switched from the subsequent processing device when the subsequent processing (for example, recognition) is completed. By sending it to the signal generating means 5c, the signal can be switched so as to output the frequency component information with the next number.

As an example, consider a case where an image divided into regions as shown in FIG. 2 is processed by the initial vision device of the present invention, and the output 1 portion is black and the output 0 portion is white. Is shown in FIG. First, the figure (a)
The background is selected and output as shown in FIG. 3B. Then, the circle portion as shown in FIG. 7B, the triangular portion as shown in FIG. It can be seen that the output is sequentially and effectively selected. In this example, since the area division is somewhat rough, the output information is also somewhat rough. However, if the recognition process in the latter stage is insufficient, the area division may be finer.

Further, in this embodiment, the gradient index rod lens is used as the Fourier transform lens 2c, but it is of course possible to use a gradient index flat plate microlens or the like, and the complexity of alignment or the like may be noticed. Alternatively, a lens array whose end face is not flat may be used.

Further, the detector 3a of the frequency component detecting means 3 is divided into four vertical and horizontal quadrants as shown in FIG. 7 so that the frequency component can be detected in each quadrant, so that the quadrant and the three quadrant are detected. The detection signals of the respective frequency components in the two quadrants and the four quadrants may be combined and the above processing may be performed by the threshold value means 4a-1 and 4a-2. By doing so, not only the frequency component of the texture but also the angle component of the texture can be identified.

[Embodiment 2] FIG. 8 is a diagram for explaining the connection of neural networks used in the frequency component judging means.
FIG. 9 is a diagram for explaining the connection of the neural net shown in FIG. 8 in more detail.

This embodiment has the same structure as that of the first embodiment except the frequency component judging means 4, and the frequency component judging means 4 is shown in FIG.
And a neural network as shown in FIG. This neural network is arranged two-dimensionally as shown in FIG. 8 and the spatial frequency information (U _{i, j} in the figure) from each detector 3a is used to detect the adjacent detectors 3a in four neighborhoods.
Spatial frequency information from (U _{i-1, j} , U _{i + 1, j} , U in the figure)
_{i, J-1} , U _{i, J + 1} ) are added and output. More specifically, as shown in FIG. 9, there is a neuron that inputs the output of the frequency component from each region of the corresponding detector 3a inside each unit that constitutes the neural network, and the neuron that is adjacent to this neuron has four neighboring units. It is connected to the neurons of and its own unit.

For simplification, in the figure, only (i, j) units and (i-1, j) units among the number of units corresponding to the detector 3a are shown, and the coupling is (i-1). , J) unit to (i, j) unit, mutual coupling within the (i, j) unit, and output from the (i, j) unit to the frequency component selecting means 5 are shown. Actually, in order to obtain the signal for determining the output from the (i, j) unit to the frequency component selecting means 5, the combination shown in this figure is further added to (i + 1, j).
It is necessary to add mutual coupling with the unit, the (i, j-1) unit, and the (i, j + 1) unit.

In this neural network, these connections are applied to all the units in the number corresponding to the detector 3a. The connections shown by the solid lines in the figure are connections for coordinating with neighboring units and for propagating the excitement of the same neuron of the self unit, and the neuron is excited by the input from this pathway. The connection shown by the dotted line is a signal for competition with other neurons in the neighboring unit and the self unit, and the neurons are suppressed by the input from this path.

Considering an actual image, when the output of the same spatial frequency component is input from the adjacent region, the excitement of the neuron corresponding to the spatial frequency component is enhanced, and the output of different spatial frequency components is output. When input, the excitation of the neuron corresponding to the spatial frequency component is suppressed. In other words, segmentation due to spatial frequency components (textures) progresses between adjacent image areas.

Further, in the self region where this segmentation has progressed, the excitement of neurons excited by self-coupling is enhanced and the excitement of neurons with less excitement is promoted, and the output of the neuron with the most excitement that survives last is characteristic. Frequency component selection means 5 as a typical spatial frequency component
Is output to.

It is apparent from the above description that the initial visual device of this embodiment can satisfactorily perform the selective output of the portions having different textures as in the first embodiment. Further, in the case of the initial visual device according to the present embodiment, even if there is a missing portion or noise in the image, there is a great merit that the signal can be interpolated or removed by exchanging the signal with the neighboring unit, that is, the neighboring pixel. . Further, it is more effective to set the coupling strength when exchanging the coupling signals.

[0035]

As described above, according to the initial vision device of the present invention, the image information is divided into a plurality of areas, and then the characteristic spatial frequency component is determined from the texture included in each area, and this information is determined. After further determining the characteristic spatial frequency components of the entire image, these are sequentially selected and output to the next processing, so that the effective feature amount for recognizing the object in the texture image can be easily It became possible to selectively extract.

[Brief description of drawings]

FIG. 1 is a diagram showing the configuration of an embodiment of an early vision device of the present invention.

FIG. 2 is a schematic diagram for explaining the operation of the early vision device of the present invention.

FIG. 3 is a diagram for explaining basic components.

FIG. 4 is a diagram for explaining the configurations of Fourier transform means and frequency component detection means.

FIG. 5 is a diagram for explaining components after the frequency component determining means.

FIG. 6 is a diagram showing a processing example by the initial vision device of the present invention.

FIG. 7 is a diagram showing a modification of frequency component detection means.

FIG. 8 is a diagram for explaining connection of neural nets used in frequency component determination means.

FIG. 9 is a diagram for explaining the connection of the neural net shown in FIG. 8 in more detail.

FIG. 10 is a diagram for explaining a human initial visual mechanism.

FIG. 11 is a diagram showing the excitation intensity of each cell at the light-dark boundary in the initial visual mechanism.

FIG. 12 is a diagram showing a conventional example of an initial visual mechanism using an electric circuit.

FIG. 13 is a diagram showing an example of an input image.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Image input means 2 ... Fourier transform means 3 ... Frequency component detection means 4 ... Frequency component determination means 5 ... Frequency component selection means 6 ... Output switching means

Claims (1)

[Claims] 1. An image input means for capturing image information in a device, the input image information is divided into a plurality of areas, and the image information of each area is held by the image information of each area. Fourier transform means for converting to frequency component information, frequency component detecting means for detecting the frequency component information for each converted region, and frequency component information for each region detected from the frequency component information for each region. A frequency component determining means for determining a characteristic frequency component, a characteristic frequency component forming the entire image from the characteristic frequency components in each area, and a characteristic component forming the entire image. Frequency component selection means for generating a signal for sequentially selecting frequency components, and each region sent from the frequency component determination means for receiving the signal emitted from the frequency component selection means An initial visual device characterized by comprising output switching means for sequentially switching characteristic frequency components for each and outputting to the next stage processing such as recognition.