JPH02280078A - Object image pickup system - Google Patents

Abstract

PURPOSE:To bring a fine three-dimentional form of an object body to image pickup by irradiating the object body with an ultrasonic wave and measuring and analyzing its scattered wave by using an array type receiver. CONSTITUTION:A signal is sent from a pulse generator 4 and a burst wave is sent to an object body 1 from an ultrasonic irradiater 2 and a scattered wave is measured by M pieces by changing the time by a receiver array 3 and processed by a computer 5. In this case, by utilizing a fact that the scattered wave and a surface existence function of an object are in a linear function, the existence function is calculated and a rough image is obtained on an indicator 6. Subsequently, when information passes through inter-unit coupling (full line) through a neural network by an input layer unit of NXNXM pieces, an output layer unit of LXLXK pieces, and an intermediate layer unit whose number of pieces is variable, it is set to weight multiple of its coupling, the information is outputted based on a bias value of each unit and the network is allowed to learn the inter-coupling weight for realizing a desirable input/ output corresponding function and the bias value. Since data for learning can be constituted artificially, a desired network is constituted at a high speed and a fine object image is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an object imaging method, such as a three-dimensional camera using ultrasound, a three-dimensional copier, a three-dimensional facsimile,
It can also be applied to robot eyes, etc.

For example, a method known as acoustic holography is a method in which a target object is irradiated with ultrasonic waves, the scattered waves are received by multiple receivers, and the shape of the target object is imaged from the measured values. Although this has been considered in the past, practical imaging has not yet been achieved due to the length of the ultrasonic wavelength and the limited size of the receiving array.

In addition, a method of constructing ultrasound images using neural networks has been proposed in "Robot Eye System Using Ultrasound Using Neural Networks" (IEICE, February 22, 1989). High-definition reconstructed images could not be obtained for objects with unlearned shapes.

The present invention was made in order to solve the above-mentioned drawbacks, and it can be solved by irradiating an object with ultrasonic waves and measuring and analyzing the scattered waves using an array receiver. The purpose of this invention is to provide an object imaging method that images a high-definition three-dimensional shape of a mold.

In order to achieve the above object, the present invention (1) emits ultrasonic waves to a target object, and reduces the sound pressure of the scattered waves to N (N>0)
When reconstructing the shape of a target object by measuring it with an array type receiver, the operation is divided into two stages.In the first stage,
The approximate shape of the target object is reconstructed from the sound pressure of the scattered waves, and in the second stage, the precise shape of the target object is reconstructed from the approximate shape of the target object. Emit ultrasonic waves and reduce the sound pressure of the scattered waves to N (N>0)
In order to reconstruct the shape of the target object using the array type receiver, the operation is divided into two stages.In the first stage, the approximate shape of the target object is reconstructed from the sound pressure of the scattered waves. In the second stage, when reconstructing the precise shape of the target object from the approximate shape of the target object, both the first stage and the second stage calculate the correspondence from the given data and the desired output data. (3) Emit ultrasonic waves to a target object, measure the sound pressure of the scattered waves with N array receivers (N > 0), and determine the shape of the target object. In order to reconstruct, prepare N objects with different positions and shapes as target objects, define the scattered waves from them, and utilize the fact that the target objects and the sound pressure of the scattered waves have a linear relationship. By? ll! l
Construct a relational expression that reconstructs the shape of the target object from the determined sound pressure, and (4) radiate ultrasonic waves to the target object and calculate the sound pressure of the scattered waves by N pieces (N>0) In order to reconstruct the shape of the target object by measuring it with an array type receiver,
The operation is divided into two stages: in the first stage, the approximate shape of the target object is reconstructed from the sound pressure of the scattered waves, and in the second stage, the precise shape of the target object is reconstructed from the approximate shape of the target object. In the case where the method according to claim 3 is used in the first stage and the neural network is used in the second stage, or (5) the ultrasonic wave is emitted to the target object and the sound pressure of the scattered wave is In order to reconstruct the shape of the target object by measuring with N (N>0) array type receivers, the operation is divided into two stages.In the first stage, the approximate shape of the target object is calculated from the sound pressure of the scattered waves. In the second stage, the first stage reconstructs the precise shape of the target object from the approximate shape of the target object.
In the second stage, when a neural network is used, input data and teacher output data used for learning the neural network are combined with artificially created output data. The feature is that input data obtained by conversion is used instead. Hereinafter, the present invention will be explained based on examples.

The object imaging method of the present invention is roughly divided into the following two stages.

First stage: A method in which the target object is irradiated with ultrasonic waves and the approximate shape of the target object is constructed from the scattered waves.

Step 2: A method for constructing the precise shape of a target object from the approximate shape of the two target objects.

In the first stage, the target object is irradiated with ultrasonic waves, and the sound pressure of the scattered waves is measured by N receivers (N>0).

The shape of the target object is constructed from the measured values. Since there is a linear relationship between the measured sound pressure and the shape of the target object, a matrix representing a linear transformation connecting the two can be created by specifically expressing the shapes of N different target objects and the sound pressure of the scattered waves at that time. It can be determined by measurement.

In the second stage, the approximate shape of the target object obtained in the first stage is refined using a neural network.

First, the first embodiment will be described. FIG. 1 shows a scattered wave measuring device for explaining the present invention in detail, and in the figure,
1 is a target object, 2 is an ultrasonic irradiator, 3 is a receiver array consisting of NXN receivers, 4 is a pulse generator, 5 is a computer, and 6 is a display. The signal sent from the pulse generator 4 is irradiated from the ultrasonic irradiator 2 toward the target object as an ultrasonic burst wave (ro represents the position of the irradiator). At this time, the scattered waves from the target object are measured by the receivers (p, q) of the receiver array 3 at different times, and the value is (g (p, q + r)) r = 11213"""M and sends it to the computer 5.The calculation in the computer 5 is performed as follows.

FIG. 2 is a diagram for explaining the present invention in detail, and shows a digitalized reproduction area of a target object. At this time, the function ξ(itj+k) is defined as follows.

At this time, there is a linear relationship between the scattered wave g (p+q+r) and the surface existence function ξ(i, j, k) of the target object, so
Using a certain constant a(P+Q+r+l+j,k), 1=O
It can be expressed as j=Ok=O. a(P+Q+rtltJ+k)
To find this, it is best to place a small object whose surface exists only in each voxel (i, , +, k) and measure the scattered waves at that time.

P=(ppq+r) I=(1+j+k) If we write it as , both expressions are.

g(P)=Σa(P,I)ξ(I) Since it is an IE voxel, the inverse matrix of matrix (a([',I)) P, I is (a'(1,P)) r,p (Given a matrix.

Methods for calculating the inverse of the matrix are well known. ).

ξ(■)=Σa-” (I,P)g(P)PEill
ξ(i) can be found from J fixed time and place number. ξ(1)
Since ξ is a function representing whether or not the surface of the target object exists, if this ξ is calculated by the computer 5 and displayed on the display 6, the approximate shape of the target object can be obtained.

Next, a second stage embodiment will be described. When M and N are small, the reconstructed image obtained in the first stage deteriorates significantly and is not suitable for practical use. Furthermore, it is often impossible to increase M and N due to device limitations. Therefore, in the second stage, a method for constructing an image with higher resolution from the image obtained in the first stage will be described.

FIG. 3 is a diagram for explaining a neural network that constructs an image on LXLXK from the reconstructed image on the NXNXM reproduction area obtained in the first stage, where circles indicate each unit of the network and solid lines indicate Each represents the connection between each unit. Input) t! Units belonging to J are NXN
The number of units belonging to the output layer is LXLXK, and the number of units belonging to the intermediate layer is variable (the number of intermediate layers can be selected depending on the desired performance of the network). When each piece of information passes through the inter-unit connections shown by the solid line, the weight of that connection is multiplied, and the information output from each unit is calculated as follows: 1/(1+exp (-x+0)). Here, 0 is the bias value of that unit. For a network with such a structure, when an input cover pattern and a desired output pattern are given, the well-known pack propagation method is used to calculate
A neural network can be made to learn weights and bias values between connections to realize these correspondences.

Therefore, in the second stage, a method of configuring input data and output data for making this network learn is shown.

One possible method for composing input data and output data is to construct them through actual measurements using a large number of actual objects with surfaces of various shapes, but in the case of three-dimensional objects, it is not practical to construct objects with sufficiently diverse shapes. It's not a good idea. Therefore, a method for artificially configuring input data and output data will be described below.

Figures 4(a) and 4(b) are diagrams for explaining the method of configuring input data and output data for learning by a neural network, where (,) indicates desired output data, and (b) indicates the desired output data. Each represents data when output data is expected. Although the actual data is expressed three-dimensionally, it is displayed two-dimensionally for ease of viewing.

Hereinafter, it is assumed that L=sN and tM. Here, S.

t is an integer. In FIG. 4, 5=T=3 is shown. First, a sufficient number of figures representing the surface of an object as shown in FIG. 4(a) are created on the LXLXK area. Fourth
In figure (a), the voxels with diagonal lines indicate the presence of the surface of the object.

The value here is set to 1. Further, voxels without diagonal lines indicate that there is no surface of the object, and the value here is O. Let these be the desired output data of the neural network. Next, the LXLXK region is divided into NXNXM regions (that is, 5XsXt voxels are combined into one voxel). NXNX obtained in this way
The value of M on the area is determined as follows. That is, among the old voxels included in the newly created voxel, the number of old voxels with a value of 1 is divided by 5XsXt, and the value on the new voxel is set as the value on the new voxel. In FIG. 4, the value of 5XsXt is shown as 9.

Since the input data and output data have been configured as described above, the NXNXM data obtained using this data will be as shown in Figure 4 (a), so if the network created in this way is used, the target It is possible to image the precise shape of an object.

Example - As is clear from the above description, according to the present invention, high-definition object imaging is possible by measuring reflected ultrasonic waves using a receiver array consisting of a small number of receivers.

Furthermore, high-definition images can be obtained even for shapes that were not used for learning when constructing the neural network.

Furthermore, since the data used for training the neural network can be constructed artificially rather than actually d1, a desired network can be constructed faster than ever before.

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the object imaging method according to the present invention, and is a block diagram of a scattered wave measuring device. FIG. 2 is a diagram digitizing the reproduction area of the target object. Diagrams showing the neural network, Figures 4(a) and (b),
FIG. 3 is a diagram showing a method of configuring input data and output data for learning by a neural network. 1...Target object, 2...Ultrasonic irradiation, 3...Receiver array, 4...Pulse generator, 5...°computer, 6...Display.

Claims (1)

[Claims] 1. Emit ultrasonic waves to a target object, and set the sound pressure of the scattered waves to N
When reconstructing the shape of a target object by measuring with multiple (N>0) array type receivers, the operation is divided into two stages.
In the first stage, the approximate shape of the target object is reconstructed from the sound pressure of the scattered waves, and in the second stage, the precise shape of the target object is reconstructed from the approximate shape of the target object. Imaging method. 2. Emit ultrasonic waves to the target object, and reduce the sound pressure of the scattered waves to N
In order to reconstruct the shape of the target object by measuring with N > 0 array type receivers, the operation is divided into two stages.In the first stage, the approximate shape of the target object is determined from the sound pressure of the scattered waves. In the second stage, when reconstructing the precise shape of the objective object from the approximate shape of the target object, the first stage,
2. The object imaging method according to claim 1, wherein both of the second stages are constructed by learning a correspondence relationship between given data and desired output data for the given data. 3. Emit ultrasonic waves to the target object, and reduce the sound pressure of the scattered waves to N
In order to reconstruct the shape of the target object by measuring with N > 0 array type receivers, prepare N objects with different positions and shapes as the target objects, measure the scattered waves from them, and reconstruct the shape of the target object. An object imaging method characterized by constructing a relational expression for reconstructing the shape of a target object from the measured sound pressure by utilizing the fact that there is a linear relationship between the sound pressure and the sound pressure of a scattered wave. 4. Emit ultrasonic waves to the target object, and reduce the sound pressure of the scattered waves to N
In order to reconstruct the shape of the target object by measuring with N > 0 array type receivers, the operation is divided into two stages.In the first stage, the approximate shape of the target object is determined from the sound pressure of the scattered waves. In the second stage, when reconstructing the precise shape of the target object from the approximate shape of the target object, the first stage uses the method according to claim 3, and the second stage uses the neural Claim 1 characterized in that a network is used.
Object imaging method described. 5. Emit ultrasonic waves to the target object, and reduce the sound pressure of the scattered waves to N
In order to reconstruct the shape of the target object by measuring with N > 0 array type receivers, the operation is divided into two stages.In the first stage, the approximate shape of the target object is determined from the sound pressure of the scattered waves. In the second stage, the method according to claim 3 is used in order to reconstruct the precise shape of the object from the approximate shape of the target object, and in the second stage, a neural network is used. Claim 4 characterized in that when using the neural network, the input data and teacher output data used when making the neural network learn are substituted with artificially created output data and input data obtained by converting the same. Object imaging method described.