JPH1091782A - Method for extracting specific site for gradation picture - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for extracting a specific site for a gradation picture in which the specific site of a lump-shaped shadow and a specific site in a structure except the lump-shaped shadow can be precisely extracted from a gradation picture such as an X-ray picture. SOLUTION: This method is constituted of a featured value extracting step 1, and a learning recognizing step 2 constituted of a learning sub-step 22 for learning the corresponding relation of a featured value vector group (b) calculated by the featured value extracting step 1 with a teaching signal (d) indicating whether or not a picture element from which the vector is extracted is a specific site to be extracted in an input picture by using a neural network, and a recognizing sub-step 23 for inputting the featured value vector (b) to the neural network, and outputting a signal indicating whether or not the picture element from which the featured value vector group (b) is the specific site to be extracted in the picture.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for extracting a specific part for a gray-scale image, and more particularly to a method for accurately extracting only a specific part to be extracted from an image composed of gray-scale information.
For example, the present invention relates to a method for automatically extracting a lesion such as a massive shadow from an X-ray image.

[0002]

2. Description of the Related Art As a conventional example of a method for accurately extracting only a specific part from a gray-scale image, there is the following method for automatically extracting a lesion indicated by a massive shadow from an X-ray image (Shimizu, Hasegawa, Toriwaki: "New abnormal shadow enhancement filter for chest X-ray image computer diagnosis and its characteristic evaluation" IEICE Technical Report PRU91-132 (1991)
Abbreviated as inDDfilter method. ). Here, the image composed of the grayscale information does not simply refer to only a black-and-white image, but to an image using a grayscale value of a specific color intensity of a color image or a grayscale image by appropriate conversion from a color image. An image composed of a wide range of grayscale information, such as an image that has been rendered. Hereinafter, it is simply referred to as a grayscale image.

An outline of this conventional method will be described with reference to FIG. First, as shown in FIG. 13B, an X-ray original image including a massive shadow as shown in FIG.
A plurality of directional differential filters (for example, 4
Is applied to the entire image. As a result, F1 (x, y), F2 (x, y), F3 (x, y) as shown in FIG.
It is assumed that four images y) and F4 (x, y) are obtained. x and y are coordinates in the image. Next, FIG.
And, as shown in the following equation (1), for the same certain pixel in the four images, the smallest filter output in four directions is set as the output of that pixel. Output this minDD (x, y)
Is defined.

[0004] Subsequently, as shown in FIG. 11E and the following equation (2),
The output obtained by the equation (1) is subjected to threshold processing, and only pixels having a value larger than a certain value are displayed as, for example, 1 (1 indicates black, 0 indicates white). Extract the shadows.

Output (x, y) = 1: if minDD (x, y)> threshold 0: else (2) In the vicinity of a massive shadow, F1 (x, y), F2 (x, y),
While F3 (x, y) and F4 (x, y) are relatively large outputs, in the vicinity of other line-shaped objects such as blood vessel shadows, one of them is large and the other outputs are large. Since there is a tendency to be small, the above operation extracts a massive shadow. Note that, in the image showing the extraction result in FIG. 13E, a line segment indicates noise.

[0006] The pixel here is not limited to one pixel in an image, but includes a pixel group in which a plurality of neighboring pixels are collected. In this method, one value, that is, a scalar amount is a characteristic amount of a certain pixel.

On the other hand, there is a method of extracting a set of a plurality of values, that is, a vector amount as a feature amount of a certain pixel, and thereafter roughly dividing the grayscale image into regions (AKJain, a
nd F. Farrokhnia, "Unsupervised texture segmentation
using Gabor filters ", Pattern Recognition, vol.24, n
o.12, pp.1167-1186,1991. K. Jain
Abbreviation of the method. ).

The outline of this method will be described with reference to FIGS. First, as shown in FIG. 8A, for an image including a plurality of textures (three in the figure, for example), identification in the image is performed as shown in FIGS. As a selective spatial frequency filter selectively reacting to the direction and a specific size, the following equation (3) (real space):
(4) A filter as represented by the expression (frequency space) may be changed in a specific direction and a specific size (hereinafter, both are also referred to as filter parameters) by changing a plurality of types (for example, four directions and five types). 20 sizes). Here, the specific direction and the specific size mean the direction and the size in the real space image. It should be noted that the concept and the size in the frequency space are greatly different. Therefore, in the future, a specific direction and a specific size in the real space will be referred to as “real space direction” and “real space size”, respectively, and the direction and size in the frequency space will be referred to as “frequency space direction” and “frequency space direction”, respectively. I will call it "size". Taking a grid composed of a plurality of line segments in a real space image as shown in FIG. 9 as an example, the direction of the line segments is “real space direction”,
The grid spacing corresponds to the “real space size”.

Next, the filter will be described. A Mexican hat-shaped filter composed of a Gaussian-type envelope function (first component on the right side) and a vibration function (second component) as in the following equation (3) is a Gabor filter.
It is called er. In addition, general Gabor fi
The expression of liter is not limited to the form of the expression (3), but generally may be any as long as it becomes a Gaussian shape after the Fourier transform as shown in the expression (4), and there are various modifications.

Gabor (x, y) = exp [−2π ｛(x ² + y ² ) (u ₀ ² + v ₀ ² )｝ / σ ² ] exp ｛j2π (u ₀ x + v ₀ y)｝ · ·· (3) F (Gabor ( x, y)) = {σ 2 / (2u 0 v 0)} exp [- πσ 2/2 {(uu 0) 2 / u 0 2 + (vv 0) 2 / v ₀ ² }] (4) In equations (3) and (4), x and y are coordinate systems in the real space (in the image), u and v are coordinate systems in the frequency space, u ₀ And v ₀ are the frequencies of the vibration function, and σ is a variable or constant that defines the spread of the Gaussian type envelope function, respectively. F () indicates the Fourier transform of the function. "Real space direction" in the filter parameters is (3),
In the case of the expression (4), the direction corresponds to the direction perpendicular to the vector (u ₀ , v ₀ ), and the “real space size” in the filter parameter is obtained by setting the size of the input image in the real space to M × M. ,
M / √ (u ₀ ² + v ₀ ² ).

FIG. 10 shows an outline of the Gabor Filter represented by the equations (3) and (4). FIG.
10A is a schematic shape (perspective view and plan view) in real space based on equation (3), and FIG. 10B is a schematic shape (perspective view and plan view) in frequency space based on equation (4). is there. FIG. 10 illustrates the concept of filter parameters (real space direction, real space size).
A qualitative explanation will be given using (a) and (b). FIG.
FIG. 10A shows a schematic shape of the filter in a real space. The filter is not centrally symmetric but has a Mexican hat shape that is line-symmetric with respect to a specific direction. Has the effect of differentiating in the direction of the thin arrow. In this case, it reacts strongly to the edge in the image in the direction perpendicular to the thin arrow (the direction of the thick arrow). This thick arrow corresponds to the “real space direction” in the filter parameters. When calculating the amount of differentiation, if one is differentiated from the other, the other is only switched between positive and negative. For convenience, in the drawing, arrows are provided in both positive and negative directions. It should be noted that the magnitude of the “real space size” is proportional to the magnitude of the extent of the foot of the mountain in FIG.

FIG. 10 (b) shows an outline in the frequency space, which is composed of two peaks at symmetrical positions with respect to the origin (center in the figure). The “real space direction” in the filter parameters corresponds to the direction perpendicular to the peak direction (v-axis direction in the example in the figure) viewed from the origin (u-axis direction in the example in the figure). I have. The magnitude of the “real space size” in the filter parameters corresponds to the magnitude of the reciprocal of the distance from the origin to the peak. In general, a filter having a large “real space size” can capture low frequency components in an image, and a filter having a small “real space size” can capture high frequency components in an image.

Subsequently, as shown in FIG.
The averaging process is performed in each of the 0 types of images. In the averaging process, as shown in the following equation (5), a window W xy of an appropriate size including a certain pixel ( x , y ) and a plurality of pixels around the pixel ( x , y ) is used to output the following equation (6) to the filter output The average value of the absolute value of the result of applying the sigmoid function φ (z) of Ga ( x , y ). I
(X, y) is the intensity of the image, N is the number of pixels in Wxy , μ is an appropriate constant, and the symbol * is the convolution product.

H Ga (x, y) = Σ | φ (Gabor (x, y) * I (x, y) | / N (x, y) ∈W xy ··· (5) φ (z) = {1-exp (−2 μz)｝ / {1 + exp (−2 μz)} (6) Subsequently, as shown in FIG.
By arranging the outputs of the pixels at the same position in
A zero-dimensional vector is defined, and the vector is defined as a feature amount of the pixel. Note that the pixel ( x , y ) in FIG.
In order to correspond to different filter parameters as the notation of the component of the feature amount vector with respect to Ga _i ( x , y ), i = 1-20.

The feature amount is extracted for all pixels in the image, and the vector group is divided into three by an unsupervised cluster classification method called a k-means method. If the feature amount is set to a three-dimensional vector instead of 20 to facilitate the illustration, the vector group is as shown in FIG. When the vector group is divided into three by the k-means method, the result is as shown in FIG. A label (in this case, A, B, and C) is attached to each of the three divided vector groups, and the label is attached to the pixel from which the vector is extracted, so that the original image becomes as shown in FIG. The area is divided. It should be noted that the region-divided boundary line usually does not completely match the boundary line of the original image, and is often slightly distorted.

[0016]

However, the above-described minDDfilter method can extract massive shadows with a certain degree of accuracy. However, among a plurality of (here, four) outputs for a certain pixel, the minimum value output is obtained. However, there is a case where the amount of information for classification is insufficient and a part other than the specific part to be extracted is also extracted. In addition, no consideration is given to a case where a specific part to be extracted has a structure other than a block shadow (for example, a linear part).

The above-described A.I. K. Since the feature of Jain uses a vector for the feature for a certain pixel, it can be said that the feature in the image is more effectively extracted than in the case of the minDDfilter method. However, the method of classifying the vector group is k-means. Since the method uses the unsupervised cluster classification method called the method, only rough region division can be performed. Therefore, further processing is required to extract a specific portion therefrom.

The rough area division is shown in FIG.
In the case of a texture whose basic structure and its arrangement are relatively clear, as in the original image of (a), it can be performed with a certain degree of accuracy. However, in the case of a texture whose basic structure and its arrangement are not clear, the rough region division is performed. Even it becomes difficult. In this case, it is very difficult to extract a specific part even if further processing is performed.

In order to understand the above, FIGS.
This will be described with reference to FIG. As shown in FIG. 8A, each texture has a relatively clear basic structure and its arrangement.
And a plurality of different types (here, 3
In the case of an original image including textures, the feature amount vector group is clearly divided into a plurality (here, three) in the feature amount vector space as shown in FIG. It is possible to perform classification by an unsupervised cluster classification method called a means method.

Incidentally, a general image does not always consist only of a texture whose basic structure and arrangement are clear. There may be a mixture of textures whose basic structure and their arrangement are not clear, or there may be no texture at all. In that case, as shown in FIG.
In many cases, clusters are not distributed in the feature vector space so that they can be clearly classified. For example, a medical image represented by an X-ray image corresponds to such a case. In other words, if the site in the image is not composed of a specific basic structure and its arrangement cannot be clearly defined, as in the case of a massive shadow, the cluster will not be clearly classified even in the feature vector space. For example, FIG.
A feature vector group showing a blocky shadow does not form a cluster that can be clearly classified in the feature vector space. In such cases, the unsupervised cluster classification method uses
It is not possible to classify clusters with high accuracy according to the distribution of clusters in the feature vector space. Therefore,
Considering the image plane, accurate region division cannot be performed.
For example, in this case, if the area is divided, a specific part to be extracted may be divided. After a specific part to be extracted has been divided, it is very difficult to extract that part. Therefore, A. K. When the method of Jain is applied to a general image, it is also difficult to roughly divide a region, and hence it is further difficult to extract a specific portion thereafter.

The above A. K. To summarize the problems of the Jain method: A-1) Since only rough segmentation can be performed, some post-processing is required to extract a specific part. A-2) When the feature amount in the image is not a cluster that can be clearly classified in the feature amount vector space, even rough segmentation becomes difficult, and it is more difficult to extract a specific part. It is.

Incidentally, the above-described A. K. In the method of Jain, when a plurality of selective spatial frequency filters are applied, the dynamic range of the output value greatly changes due to a change in the size of the real space size in the filter parameters. Specifically, the larger the real space size is, the larger the obtained output value is compared to the one having a small real space size. Therefore, when the feature value vector is created by arranging the output values as they are, components having a large real space size of the filter parameters, that is, low-frequency components in the image contribute more to the classification than other components. Will be. Therefore, when components in the case where the real space size of the filter parameter is small, that is, when high-frequency components are required for classification in an image, good classification becomes difficult. A general gray-scale image does not always include only low-frequency components or only high-frequency components. For example, in the case of an X-ray image, a single image includes a background or a heart shadow as a low-frequency component and a blood vessel shadow as a high-frequency component. Therefore, by making the low-frequency component and the high-frequency component contribute to the classification in the same manner, it becomes possible to extract a specific portion from a more versatile image using both the low-frequency component and the high-frequency component in the image.

The present invention has been made in view of such problems of the prior art, and a first object of the present invention is to provide a method for extracting a specific portion for a gray-scale image satisfying the following condition B-1): Is to provide. Then, the second purpose is B
Another object of the present invention is to provide a method for extracting a specific portion for a grayscale image that satisfies the condition of -2). Subsequently, a third object is to provide a method for extracting a specific portion for a grayscale image that satisfies the condition of B-3).

B-1) A specific portion of a massive shadow and a specific portion of a structure other than the massive shadow are accurately extracted from the grayscale image. B-2) When a vector group is created while extracting a feature vector amount for each pixel from the entire area in the grayscale image, even when the vector group does not have a cluster that can be clearly classified, X A portion, such as a block shadow of a line image, in which the portion is difficult to classify clearly in the image is accurately extracted. B-3) A low-frequency component and a high-frequency component in an image are similarly contributed to classification, thereby extracting a specific portion from a more versatile image.

[0025]

To achieve the above object, a method for extracting a specific portion of a grayscale image according to the present invention comprises: an image input sub-step of inputting an input image comprising grayscale information; A selective spatial frequency filtering sub-step of performing selective spatial frequency filtering selectively responding to a specific direction in the real space and a specific size in the real space; and in the image of the filtering result, a specific pixel and its surroundings. Averaging the result of the filtering in a window composed of a plurality of pixels and having a size proportional to the specific size of the real space, and averaging the averaged value in a specific window in the window. An averaging sub-step for outputting as a pixel feature, and at least two or more different selective spatial frequency filters A feature amount vector creating sub-step for creating a vector using the averaged output as a component of the feature amount of the specific pixel, and a feature amount obtained by the feature amount extracting step. Using a vector and a teacher signal indicating whether or not the pixel from which the vector is extracted is a specific part to be extracted in the input image, use a neural network to determine the correspondence between the feature amount vector and the teacher signal. A learning sub-step of inputting the feature vector to the neural network and outputting a signal indicating whether or not the pixel from which the feature vector is extracted is a specific part to be extracted in the image. And a learning recognition step comprising a sub-step.

The operation of the first method will be described. This method corresponds to first and fourth embodiments described later. In the feature amount extraction step, first, an input image composed of grayscale information is input (image input sub-step), and a selection is made for the input image to selectively react to a specific direction and a specific size in a real space in the image. Applying spatial frequency filtering (selective spatial frequency filtering sub-step). In this case, the processing is performed on a plurality of types by changing the real space direction and the real space size. Subsequently, an averaging process is performed on the filtering output in each of the plurality of types of images (averaging sub-step). In the averaging process, the filtering output is averaged in a window including a specific pixel and a plurality of pixels around the pixel, and having a size proportional to the specific size of the real space, and averaging the filtered output. Is the output of a particular pixel in the window. In this case, the filtering output itself may be averaged, or the sigmoid function may be averaged. In any case, the absolute value may be obtained before averaging. A vector is created by sequentially arranging the outputs subjected to the averaging process according to the plurality of types of filters (feature amount vector creation sub-step).
The vector is set as a feature amount of the pixel. The above is the operation of the feature amount extraction step.

Subsequently, a specific part is extracted from the grayscale image by using the feature amount. As a learning / recognition step for this purpose, a neural network having learning and recognition actions is used. That is, the input data of the neural network is used as a feature vector extracted in the feature extraction step, and the teacher signal is used as a signal indicating whether or not the pixel from which the vector is extracted is a specific part to be extracted in the image. At the time of learning, the correspondence is learned by simultaneously presenting a feature amount vector and a teacher signal indicating whether or not the pixel from which the vector is extracted is a specific part to be extracted in the image to the neural network ( At the time of recognition, a feature vector is input to the neural network, and a signal is output as to whether the pixel from which the vector is extracted is a specific part to be extracted in the image (recognition substep). The above is the operation of the learning recognition step.

In the above-described feature amount extraction step, a feature amount vector is extracted from each pixel of the grayscale image, and the pixel from which the vector is extracted in the learning recognition step should be extracted in the image using the feature amount vector. By learning and recognizing a signal indicating whether or not a specific portion is present, the specific portion can be extracted from the grayscale image.

The above operation will be described more specifically. First, the feature amount extraction step will be described. In the image, there are parts characterized by various real space directions and real space sizes, so to extract a specific part, simply select a specific real space direction and a specific real space size. May be applied as a feature quantity by applying a selective spatial frequency filter that responds to the above. However, in general,
It is difficult to determine in advance a specific real space direction and a specific real space size that react to a specific site. This is because the real space direction indicating the specific part to be extracted and the real space size are not always uniquely determined. Therefore, a plurality of types of selective spatial frequency filters that selectively respond to a specific real space direction and a specific real space size are prepared by changing the specific real space direction and the specific real space size, and the plural types of filters are prepared. By using the output for, it is possible to represent sites having various structures, and therefore, it is possible to extract a specific site to be output from the image. To extract a specific part, as a selective spatial frequency filter,
For example, a filter having a shape as shown in FIG. The magnitude of the extent of the foot of the mountain in FIG. 10A is proportional to the magnitude of the specific real space size that reacts, and the direction of the mountain (the direction of the thick arrow in FIG. 10A) is a response. Corresponding to a specific real space direction.

Further, the plurality of types of outputs are not directly used as components of the final feature amount vector, but are subjected to an averaging process before that. Said filter is a fairly local filter with a differentiating action. Since the specific part is a part having a certain size in the image,
It can be said that a global feature value in a window having a certain size in an image expresses a feature of a specific portion better than a local feature value. In order to make the local feature amount a global feature amount in the window, the feature amount may be averaged in the window in the image. in this case,
The output itself may be averaged, or the output obtained by applying a sigmoid function to the output may be averaged. In any case, the absolute value may be obtained before averaging. Taking the absolute value is equivalent to not considering the positive or negative of the differential result. In the case where a massive shadow is extracted, it is not necessary to distinguish between the left and right of the boundary of the massive shadow. Further, when the absolute value is obtained, the absolute amount of variation in the window can be extracted, so that a more global feature amount can be extracted. The size of the window is Gabo
The rfilter may be changed in proportion to the real space size. That is, for example, when applying a filter that responds to a large real space size in an image,
If the averaging process is also performed with a large window, the feature can be extracted with higher accuracy. When a filter that responds to an object with a small real space size is applied, the averaging process is preferably performed with a small window.

The feature extraction step has been described above. Since the output of a filter that selectively responds to a plurality of types of specific real space directions and specific real space sizes is used as the feature amount, it is possible to represent sites having various structures. In addition, since the image is averaged in the window in the image, the feature of a part having a certain size in the image can be well represented. Also, since various structural parts can be represented, it is possible to represent not only massive shadows but also specific parts of structures other than massive shadows.

The filters selectively responding to a specific real space direction and a specific real space size shown in FIGS. 10A and 10B have a simple aperture in the frequency space as shown in FIG. 10C. May be used. The convolution product is
The execution of the convolution product is often performed in the frequency space, since it is a simple product in the frequency space. Therefore, the simpler filter shape in the frequency space has an advantage that the filter design is easier when this processing is realized on hardware, for example. In the case of the shape as shown in FIG. 10C, the locality is smaller in the real space than in the case of FIG. 10A, and the filter extracts a wider range of information in the image. That is, since the feature amount becomes a global feature amount in the image, it is more effective when a specific part of the image is extracted.

Next, a neural network used in the learning recognition step will be described. A neural network having learning and recognition functions is known as an algorithm capable of learning various input / output relations with high accuracy. Particularly when the input is a vector, the neural network has higher accuracy than a function approximation such as a least squares method which is conventionally often used. It is said that.

At the time of learning, input data and a teacher signal corresponding to the input data are simultaneously presented to the neural network to learn the corresponding relationship. At the time of recognition, the input data is input to the neural network, and based on the learned correspondence, Output a signal. As the learning algorithm, for example, a back propagation method (hereinafter abbreviated as BP method) may be used (DERumelhar
t, GEHinton and RJWilliams: Learning representat
ion by back-propagating errors, Nature, 1986a).

The learning algorithm of the BP method will be briefly described. Consider a hierarchical neural network consisting of an input layer, a hidden layer, and an output layer. Elements for performing signal processing exist in the input layer, the intermediate layer, and the output layer. Generally, a layer of an element for inputting an input value is called an input layer, a layer of an element for extracting an output value is called an output layer, and layers of other elements are called an intermediate layer. The elements of the output layer and the intermediate layer are respectively coupled to the elements of the preceding layer, and a coupling weight value is assigned to each of the couplings. A value obtained by multiplying the output of each element of the preceding layer by the coupling load value is input to the elements of the output layer and the intermediate layer, and the sum is calculated. In some cases, the element may have a bias term which is a unique value, and the sum may be added thereto. The result of performing a function called an input / output function on the obtained sum value is the output value of the element. By setting the value of the connection weight and the value of the bias to various values, it is possible to make the neural network learn the input / output relationship to be obtained. As a learning method, a square error between a desired output for a certain input (called a teacher signal) and an actual output of the output layer may be minimized. The BP method is an algorithm that changes the network connection weight value and the bias value so that the square error is minimized.

The BP method as described above is used as a neural network having learning and recognition functions, and learning and recognition are performed using the feature values extracted in the above-described feature value extraction step. In this case, the feature amount vector is used as an input to the input layer. For example, the number of classes to be recognized is set to 2, the teacher signal of the BP method, that is, the desired output to be output from the output layer is set to 1 and 0, and 1 is the vector extracted. What is necessary is that the vector extracted from the inside of the specific part to be extracted, and 0 is the vector extracted from outside the specific part to be extracted.

A neural network having learning and recognition functions has a property of learning various non-linear correspondences between input and output by providing a teacher signal corresponding to an input at the time of learning. Even when there is no cluster that can be classified, it is possible to classify and extract a cluster consisting of a feature vector group representing a specific part to be extracted.

According to the first method of the present invention, the output of a filter selectively reacting to a plurality of specific real space directions and a specific real space size is used as a feature amount from a grayscale image. In addition, since the image is averaged in the window set in the image, it is possible to accurately extract a specific portion represented by a massive shadow and a specific portion represented by a structure other than the massive shadow from the grayscale image. In addition, since a neural network having learning and recognition functions is used as a learning and recognition step for learning and recognizing a feature, even if the feature vector group does not have a cluster that can be clearly classified, it represents a specific part to be extracted. Clusters composed of feature amount vector groups can be classified and extracted in the feature amount space. Therefore, even when it is difficult to clearly classify the site because it does not have a clear basic structure or arrangement in the image, such as a massive shadow of an X-ray image, it is possible to accurately extract the site. it can.

According to a second method of the present invention, in the first method, the neural network used in the learning / recognition step calculates a similarity between the feature amount vector and a weight vector group set in the neural network. A neural network for learning or recognizing, wherein the learning and recognizing step includes a labeling sub-step of pre-applying a label indicating a specific part to be extracted to at least one weight vector, wherein the feature amount vector and the A similarity calculation sub-step for calculating the similarity of the weight vector group, a weight vector selection sub-step for selecting the weight vector having the large similarity, and a label detection for detecting a label assigned to the selected weight vector Sub-step, and using the label and the teacher signal, A weight vector updating sub-step of updating the torque vector, a learning sub-step of learning a correspondence relationship between the feature amount vector and the teacher signal, and a similarity calculating a similarity between the feature amount vector and the weight vector group. A calculation sub-step, a weight vector selection sub-step for selecting a weight vector having a large similarity, and a label detection sub-step for detecting a label assigned to the weight vector, wherein the feature amount vector is extracted. A recognition sub-step of outputting a signal indicating whether or not the pixel is a specific part to be extracted in the image.

The operation of the second method will be described. This method corresponds to first and fourth embodiments described later. As a neural network of the first method, a neural network for learning or recognition is prepared by calculating the similarity between the feature vector extracted in the feature extraction step and the weight vector set in the neural network. At this time, a label indicating a specific portion is assigned to at least one weight vector in advance (label assignment sub-step).

In this neural network, at the time of learning and at the time of recognition, as a basic operation, the feature vector extracted in the feature extracting step of the first method is compared with each weight vector set in the neural network. It has the effect of calculating the similarity, selecting the weight vector having the highest similarity, and outputting the label of the weight vector.

At the time of learning, the feature vector extracted in the feature extraction step and a teacher signal indicating whether or not the pixel from which the feature vector is extracted is a specific part in the image are presented to the neural network. When the feature vector extracted from the pixel of the specific part is input, the weight vector is updated (the component of the vector is changed) so that the weight vector labeled with the specific part is selected. (Learning sub-step).

At the time of recognition, when the feature amount vector extracted in the feature amount extraction step of the first method is input, the above-described weight vector having a large similarity is selected, and the label of the weight vector is output. Outputs whether or not the pixel from which the vector is extracted is a specific part in the image (recognition sub-step). By the above operation, a specific part is extracted from the grayscale image.

Next, a neural network for learning or recognizing by calculating the similarity between the feature vector and the weight vector group will be specifically described. First, the concept is roughly explained. At the time of learning, a weight vector is appropriately updated using an input vector corresponding to a feature vector and information (label) indicating to which class the input vector belongs. Classification is performed by determining which weight vector the input vector belongs to. Which weight vector the input vector belongs to
Calculate the similarity between the input vector and each weight vector,
This is performed by using the closest one as the weight vector to which the closest one belongs. The weight vector is assigned a label for each class in advance, and the class of the input vector belongs to the class of the input vector, whereby the class classification of the input vector is performed. As the similarity,
The distance between vectors is often used.

The specific learning recognition algorithm can be summarized as follows. C-1) The weight vectors are roughly arranged (corresponding to initialization). C-2) Label the weight vectors using labeled input vectors of known class. C-3) Update the labeled weight vector sequentially using the input vector. FIG. 7 shows the relationship between the weight vector and the input data as two-dimensional vector data.

The above-mentioned C-2) corresponds to the portion of the second method "labeling sub-step of pre-assigning a label indicating a specific portion to be extracted to at least one weight vector", and C-3). Is a similarity calculation sub-step of calculating the similarity between the feature amount vector and the weight vector group, a weight vector selection sub-step of selecting the weight vector having a large similarity, and the selected weight vector A label detecting sub-step of detecting the assigned label, and a weight vector updating sub-step of updating the weight vector using the label and the teacher signal, wherein a correspondence relationship between the feature amount vector and the teacher signal is provided. Learning sub-step for learning ".

As a method of roughly arranging the weight vectors of C-1), vectors may be randomly generated,
A special vector quantization method (for example, Self-organizing ma
p, hereinafter also abbreviated as SOM. T.Kohonen, Self-Organizin
g and Associative Memory, Third Edition, Springer-Ve
rlag, Berlin, 1989).

As the labeling of C-2), for example,
The similarity between the labeled input data whose class is known and each weight vector is calculated, and the weight vector having the highest similarity may be labeled with the class of the input data. Hereinafter, this is also referred to as a standard labeling method.

The updating of the weight vector in C-3) is performed, for example, by learning vector quantization.
nation method (hereinafter abbreviated as LVQ method) (T. Kohonen,
Self-Organizing and Associative Memory, Third Editi
on, Springer-Verlag, Berlin, 1989). L
The VQ method is represented by the following equations (15) and (16).

[0050] [Update] m _C (t + 1) = m _C (t) + α (t) {x (t) −m _C (t)} at Class (x (t)) = Class (m _C (t)) m _C ( t + 1) = m _C (t) −α (t) ｛x (t) −m _C (t)｝ at Class (x (t)) ≠ Class (m _C (t)) _mi (t + 1) = _mi (T) at i ≠ C (16) where x is an input vector, _mi is a weight vector, | x
-M _i | x and the Euclidean distance m _i, subscript C is the distance is smallest was elements (winner element, i.e., m _C
Is the weight vector with the highest similarity), Class (x
(T)) and class (m _C (t)) are x and m, respectively.
The class to which _C belongs, α (t) is a positive constant, and t indicates the number of updates. While repeating updating, the magnitude of α (t) is gradually reduced.

The equations (15) and (16) will be described qualitatively. First, the weight vector m _C closest to the input vector x
Is selected (Equation (15)). Next, when the class to which the weight vector m _C belongs and the class to which the input vector x belongs are the same, the weight vector is brought closer to the input vector (the first expression in the expression (16)), and if different, the distance is moved away ( (The second expression in the expression (16)). Weight vectors other than m _C are not updated (third expression in expression (16)). By this operation, the classification boundary is adjusted so that the recognition rate is increased. Since information (label) of the class to which the input vector x belongs is given at the time of learning, this corresponds to the teacher signal.

The BP method described above is a method for realizing a complicated non-linear input / output relationship by a complicated combination of connection weight values. However, the similarity between a feature vector and a weight vector group of the present method is calculated. The neural network that performs recognition during learning is a method of creating a nonlinear boundary by arranging weight vectors, and can create a more complicated nonlinear boundary. In addition, the arrangement method is very simple. Therefore, there is an advantage that the class boundary can be easily and accurately obtained as compared with the neural network using the BP method. Further, medical images include images of various cases, and it is necessary to perform additional learning after learning. The neural network can easily perform additional learning by a method such as adding a weight vector, and is also more effective than the BP method.

When the weight vector of C-1) is roughly arranged, if the SOM known as an unsupervised neural network is used, the distribution of the weight vector approximately matches the distribution of the feature amount vector. As described above, since the weight vectors are arranged, when learning is performed later, it becomes easier to accurately determine the class classification boundary, and the convergence of the learning also becomes faster.

According to the second method of the present invention, since the neural network for learning or recognizing by calculating the similarity between the feature vector and the weight vector group is used, it is extracted in the feature vector space. A vector group indicating a specific part to be obtained can be classified and extracted more accurately and easily. Therefore, the specific part in the image can be more easily and accurately extracted.

A third method of the present invention is the method according to the first or second method, wherein the feature amount extracting step includes a dynamic range adjusting sub-step in order to make a dynamic range of each component of the feature amount vector uniform. It is a method characterized by having.

The operation of the third method will be described. This method corresponds to a second embodiment described later. In the first and second methods, after the dynamic range of each component of the feature vector obtained in the feature extraction step is provided (dynamic range adjustment sub-step), the dynamic range is input to the learning recognition step. Here, to make the dynamic range uniform means to roughly match the possible range for each component of the output data. Therefore,
Various methods can be considered as a method of equalizing the dynamic range of each component. The maximum value that can be taken for each component may be made constant (Equation 17), or the maximum value and the minimum value may be made constant. (Equation (18)). Further, the average value of each component may be fixed (Equation (19)).

[0057] However, n as the number of components, the input vector before aligning the dynamic range _{^{i X = (i X 1,}} i X 2, ···
, ^I X _n ), and the aligned input vector is given by ⁱ X ^′ = ( ⁱ
X ₁ ^′ , ⁱ X ₂ ^′ ,..., ^I _Xn ^′ ), i is the number assigned to the vector (i = 1 to k), and j is the number assigned to the vector component (j = 1 to n) And

The components of each feature vector reflect the high-frequency components and low-frequency components of the image. Therefore, according to the third method of the present invention, the dynamic range of each component of the feature amount vector obtained in the feature amount extracting step is made uniform, so that in the subsequent learning and recognition step, the low-frequency component and the high-frequency component in the image are removed. Also contributes to the classification. Therefore, it is possible to extract a specific portion from a more versatile image using both low-frequency components and high-frequency components in the image.

According to a fourth method of the present invention, in the second method, the label assigning sub-step further comprises the step of: labeling the specific portion to be extracted according to a ratio of the size of the specific portion to be extracted in the input image. Is varied by varying the ratio of the number of weight vectors to which the weight vector is assigned.

The operation of the fourth method will be described. This method corresponds to a third embodiment described later. In the neural network of the second method, when a label is assigned to a weight vector group in advance, the number of weight vectors to which a label indicating a specific portion is assigned according to the ratio of the size of the specific portion to be extracted in the image The ratio occupied by.

The effectiveness of this method will be specifically described. In general, the area ratio of the specific portion extracted from the grayscale image to the entire image is small. Therefore, the number of feature vectors extracted from a specific part to be extracted is smaller than the number of feature vectors extracted from parts other than the specific part. Therefore, in such a case, it can be said that the number of input vector data (corresponding to a feature amount vector) greatly differs for each class.

As described above, in the neural network used in the second method, labeling is performed in advance, but when the number of input vector data greatly differs from class to class, a label of a class having a small number of data is assigned. Since the number of weight vectors to be given is small, it becomes difficult to accurately create a boundary for classifying the input vector data during learning, and a sufficient recognition rate may not be obtained. As described above, when the number of input vector data greatly differs from class to class, a method for improving the recognition rate by changing the learning algorithm is disclosed in Japanese Patent Laid-Open No. 4-25 / 1990.
No. 9000. This method changes the learning method when the number of input vector data greatly differs from class to class, but does not consider the labeling process. Since the number of input vector data is small, if the number of weight vectors to which a label indicating the class is given is small, a sufficient recognition rate may not be obtained even if the learning is devised.

Therefore, when the number of input vector data greatly differs from class to class, the number of weight vectors to which labels of the class are assigned is changed according to the number of input vector data belonging to the class. I just need. For example, when the number of input vector data is small, the number of weight vectors to which the label of the class is given is increased, and when the number of input vector data is large, the label of the class is given conversely. What is necessary is just to reduce the number of weight vectors. Regardless of the magnitude of the number of input vector data belonging to a class, a sufficient weight vector can be prepared for each class, so that it is possible to accurately create a boundary for dividing a class in the input vector data space during learning.

In the labeling substep, the number of weight vectors to which a label of the class is assigned is changed according to the number of input vector data belonging to the class, for example, as follows. In the following, a case where the number of weight vectors to which labels of classes having a small number of data are assigned are increased is shown as an example.

D-1) When assigning a label, the number of input vector data for each class is determined, and then only the input vector data of the class with a small number is used. For example,
In the case of two classes (example: class A and class B)
A label is assigned to the weight vector using only the class A input vector data having a small number of input vector data, and a class B label is assigned to the weight vector to which the class label is not assigned.

D-2) When assigning a label, the number of input vector data for each class is obtained, and then, for a class having a small number of input vector data, a label is easily attached to the weight vector. For classes with a large number of, it is difficult to label the weight vector. Adjustment of ease of label attachment and difficulty of attachment may be performed, for example, as follows.

When calculating the similarity between the input vector data and the weight vector, the similarity between the input vector data and the weight vector of the class having a small number of input vector data is set to a coefficient such that the similarity becomes larger than the actual similarity. Multiplying or changing the definition of similarity makes it easier to attach a label, and the similarity between input vector data of a class with a large number of input vector data and the weight vector is smaller than the actual similarity. If the degree is multiplied by a coefficient or the definition of the similarity is changed, the label becomes difficult to be attached.

As described above, the number of weight vectors to which labels of classes having a small number of data are assigned can be increased. The description of the case of reducing the number of weight vectors to which labels of classes having a large number of data are added is omitted because it can be easily analogized from the above description of D-1) and D-2).

As described above, in the neural network used in the second method, the ratio of the number of weight vectors to which the label indicating the specific portion is occupied is determined according to the ratio of the size of the specific portion to be extracted in the image. If it is made to fluctuate, it becomes possible to accurately create a boundary for dividing a class in the feature vector space, and therefore, it is possible to more precisely extract a specific part from a grayscale image.

According to the fourth method of the present invention, in the label assigning sub-step, the number of weight vectors to which labels of the class are assigned is changed according to the number of input feature quantity vector data belonging to the class. By doing so, a sufficient weight vector can be prepared for each class regardless of the number of feature quantity vector data belonging to the class, so that boundaries for class division in the feature quantity vector space are created with high accuracy. Therefore, it is possible to more precisely extract the specific part from the grayscale image.

[0071]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Some embodiments of the present invention will be described below. [First Embodiment] The first embodiment corresponds to the first method and the second method. First, an overview of the overall processing will be described with reference to FIG.

First, in the image input sub-step 11, a gray-scale image a is input. In the selective spatial frequency filtering sub-step 12, the input gray-scale image a is subjected to selective spatial frequency filtering selectively responding to a specific real space direction and a specific real space size in the image. In this case, the processing is performed on a plurality of types by changing the real space direction and the real space size.

Subsequently, in the averaging sub-step 13, a window composed of a specific pixel and its surrounding pixels is set for each of a plurality of types of filtered outputs (filtered images) obtained as described above. Averaging is performed internally. The averaged output is
In the feature amount vector creation sub-step 14, as a feature amount for one filter of a specific pixel, a vector is created by arranging the filters according to the order of the plurality of types of filters, and the vector is stored in the particular pixel. Let it be a feature vector. By creating such a vector for each pixel in the gray image a, a plurality of feature amount vector groups b are generated for one gray image. The above is the description of the feature amount extraction step 1.

The feature vector group b created in the feature extraction step 1 is subsequently sent to the learning recognition step 2. Note that the learning recognition step 2 in the figure particularly corresponds to the second method. In the drawing, a thin broken line indicates the flow of the labeling sub-step 21 before learning recognition, a thick broken line indicates the flow of the processing of the learning sub-step 22, and a thick solid line indicates the flow of the processing of the recognition sub-step 23. First,
In the labeling sub-step 21, the weight vector group c is labeled using the feature amount vector group b in advance as described in C-2) of the operation of the second method. It is assumed that the feature amount vector group b used at this time is known as a priori knowledge (for example, in the case of an X-ray image, a doctor's diagnosis) and has been extracted from a specific part. That is, a label is given to the feature amount vector group b.

Subsequently, in the learning sub-step 22, in the similarity calculation sub-step 201, the similarity (for example, the distance between the vectors) between any one vector in the feature quantity vector group b and the weight vector group c is determined. In the weight vector selection sub-step 202, a weight vector having the maximum similarity (for example, the distance between the vectors is minimum) is selected, and the label e of the weight vector selected in the label detection sub-step 203 is detected.

Using the label e and the teacher signal d, in the weight vector updating sub-step 204, the weight vector group c is set so that the label e and the teacher signal d match as much as possible.
To update. The above-described learning is sequentially performed on all of the feature amount vector groups b (the feature amount vectors that are all labeled by a priori knowledge as described above) or the vectors appropriately extracted therefrom. repeat.

Subsequently, in the recognition sub-step 23, one of the feature amount vector groups b and the weight vector group c
Are calculated in a similarity calculation sub-step 201, and in a weight vector selection sub-step 202,
The weight vector having the maximum similarity is selected, and the label e of the weight vector selected in the label detection sub-step 203 is detected. That is, the recognition corresponds to outputting the label e for the feature amount vector. If the other vectors in the feature amount vector group b are processed in the same manner, the label e is assigned to all the feature amount vectors of all the pixels, and a desired specific portion to be extracted in the image can be extracted. .

Next, a description will be given, with reference to FIG. 2, of a gray-scale image specific site extracting apparatus 50 for implementing the above method. This device is mainly composed of a feature amount extracting device 30 for extracting a feature amount from a grayscale image, and a learning recognition device 40 for learning and recognizing using the feature amount.

The feature quantity extracting device 30 includes an image input device 31
And a plurality of selective spatial frequency filtering operators 3
2, a number of averaging calculators 33 equal to the number of selective spatial frequency filtering calculators 32, a feature vector creation calculator 34, and a feature vector group memory 35. The operation will be described.

First, a gray-scale image a is input using the image input device 31 (corresponding to the image input sub-step 11).
Then, a selective spatial frequency filtering that selectively reacts to a specific real space direction and a specific real space size in the image is performed on the input grayscale image a by using the selective spatial frequency filtering calculator 32. (Corresponding to the selective spatial frequency filtering sub-step 12). In this case, the processing is performed on a plurality of types by changing the real space direction and the real space size. In FIG. 2, a plurality of selective spatial frequency filtering calculators 32 are prepared. Subsequently, a window including a specific pixel and pixels around the specific pixel is set for each of the plurality of types of filtering outputs (filtering images) obtained as described above using the averaging calculator 33. The averaging process is internally performed (corresponding to the averaging sub-step 13). A vector is created by arranging the averaged output as a feature amount for one filter of a specific pixel in the feature amount vector creation calculator 34 in accordance with the order of the plurality of types of filters (feature amount). (Corresponds to the vector creation sub-step 14),
The vector is written to the feature amount vector group memory 35 as the feature amount of the pixel.

Next, the learning recognition device 40 will be described. The learning recognition device 40 includes a labeling operation unit 41, a similarity calculation operation unit 42, and a weight vector selection operation unit 43.
A label detection calculator 44, a label output device 45,
Teacher signal memory 46 and weight vector update calculator 47
And a weight vector group memory 48. The operation includes the operation shown in the learning sub-step 22 and the operation shown in the recognition sub-step 23. First, the operation shown in the learning sub-step 22 will be described.

First, the labeling arithmetic unit 41 compares the weight vector group recorded in the weight vector group memory 48 with the feature vector group recorded in the feature vector group memory 35, Labeling is performed as described in C-2) in the description of the operation of the second invention (corresponding to the labeling sub-step 21). Subsequently, in the similarity calculation calculator 42, the similarity (for example, the distance between the vectors) between any one of the feature amount vector groups and the weight vector group is determined, and the weight vector selection calculator 4
In step 3, the weight vector having the maximum similarity (for example, the minimum distance between the vectors) is selected, and the label detection arithmetic unit 44 is selected.
Then, the label of the selected weight vector is detected. Using the label and the teacher signal extracted from the teacher signal memory 46, the weight vector update calculator 47 updates the weight vector group so that the label and the teacher signal match as much as possible (for example, the expression (16) Method). The above learning is sequentially repeated for an arbitrary vector in the feature vector group.

Next, a description will be given in accordance with the description of the recognition sub-step 23. First, the similarity calculation calculator 42 calculates the similarity between one of the feature amount vector groups and the weight vector group, and the weight vector selection calculator 43 selects the weight vector having the maximum similarity, and The detection arithmetic unit 44 detects the label of the selected weight vector and outputs it by the label output unit 45. The specific label to be extracted is displayed by the output label.

The specific image extracting device 50 for grayscale images has been described above. As the image input device 31, for example,
A CRT, a printer, or the like may be used as the label output device 45 such as a scanner or a two-dimensional image sensor. Most of the other parts can be processed very easily by using a computer to perform the respective operations on software. In addition, if a dedicated DSP (Digital Signal Processor) or an analog circuit is used as electrical hardware, the processing speed can be increased. Further, at this time, if the hardware configuration is such that the selective spatial frequency filter and the inner product operation can be performed in parallel, higher speed can be realized. Also, if an optical parallel filtering device is used as the selective spatial frequency filtering computing device 32 or an optical inner product computing device is used as the similarity calculation computing device 42, the speed can be similarly increased. Furthermore, as an image input method, in addition to the above, digital image data formatted according to a common standard such as the IS & C standard, DICOM standard, MIPS standard, etc. can be transmitted via an Ethernet or an MO device or the like. hand,
Of course you can enter it. In this case, the image input device 31
Is an information capturing interface such as a computer or an MO device.

Next, a more specific method of processing grayscale image data will be described. The X-ray image (image size: 256 × 25) of the lung field shown in the schematic diagram of FIG.
6). An example in which a massive shadow is extracted as a specific portion will be described. According to the present invention, as described above, a thing other than the massive shadow can be extracted as the specific portion, but the massive shadow is selected as an example. That is, two kinds of labels, a label representing the inside of the massive shadow and a label representing the outside of the massive shadow, may be prepared as the label e in the learning recognition step 2.

As the selective spatial frequency filter used in the selective spatial frequency filtering sub-step 12 and selectively responding to a specific real space direction and a specific real space size in an image, Gabor as shown in equation (3) is used. fill
ter was used. The specific real space direction and the specific real space size of Gaborfilter are four sizes different in size from eight directions at intervals of 15 °, and the types of the filters are 32 in total. Note that the Gabor filter includes a case where the spread of the envelope function is constant irrespective of the frequency of the vibration function, and a case where the spread of the envelope function changes according to the frequency of the vibration function as in Equation (3). But the latter gives better results. The reason is,
A filter having a large vibration function frequency (which responds to high-frequency components in an image) should be localized in real space to detect a change occurring in a small area, and the vibration function has a small frequency. This is because the filter (which responds to low-frequency components in the image) should not be very localized in the real space in order to detect a change occurring in a wide area. Also,
In equation (3), using only the real part of the filter function can provide the same effect as using both the real part and the imaginary part. Therefore, only the real part of the filter is used to reduce the processing operation. Then, high-speed processing can be performed even if the number of pixels is large.

In the averaging sub-step 13, a simple linear function was used instead of the sigmoid function φ as a function applied to the output of the convolution of the equation (5). Through the above-described feature amount extraction step 1, 256 × 256 32-dimensional feature amount vectors are created per image. A specific part is extracted in the learning recognition step 2 using the vector group.

Update method in weight vector update sub-step 204 (C-3 in description of operation of second method)
Also used the LVQ method. In the labeling sub-step 21, a standard labeling method was used as a labeling method for labeling the weight vector (see also C-2). The weight vectors were initialized by roughly arranging the weight vectors using the SOM (C-
See also 1)).

From the X-ray images of a plurality of lung fields, a feature amount vector group b is created from each pixel of each X-ray image in the feature amount extraction step 1. Subsequently, learning is first performed in the learning sub-step 22 of the learning recognition sub-step 2 using the feature amount vector group b. In addition, as the X-ray images of a plurality of lung fields used for learning, both an image including a massive shadow and an image not including the massive shadow were used. Subsequently, in the recognition sub-step 23 of the learning recognition step 2, a specific part is extracted from the image by recognizing a feature amount vector extracted from within the specific part. Note that, at the time of recognition, the X-ray image from which the massive shadow is extracted may be the image used at the time of learning, or another new image. That is, the feature amount vector group used at the time of recognition may or may not have a label. When it is desired to extract a specific part from the X-ray image used at the time of learning, a feature amount vector group used at the time of learning is input and recognized after learning, and the specific part is extracted from the image. When it is desired to extract a specific part from an X-ray image not used at the time of learning, a characteristic amount vector group is created again in the characteristic amount extraction step 1, input after learning, recognized, and a specific part is extracted from the image. I do.

By extracting the massive shadow from the X-ray image of the lung field by the above operation, a good result with little noise as shown in the schematic diagram of FIG. 4 was obtained. What is extracted other than the massive shadow to be extracted is noise (indicated by an arrow). Noise is significantly reduced as compared to the conventional case.

In the above-described first embodiment, a neural network that performs learning or recognition by calculating the similarity between a feature vector and a weight vector group is used as the learning and recognition step 2. Of course, the BP method, which is a network, may be used.

[Second Embodiment] This embodiment corresponds to the third method. As shown in FIG. 5, the feature amount vector creation sub-step 1 of the feature amount extraction step 1 of the first embodiment is performed.
After step 4, a dynamic range adjustment sub-step 15 is provided, and the dynamic range of each component of the vector obtained in the feature quantity vector creation sub-step 14 is made uniform. As a corresponding device, as shown in FIG. 6, a dynamic range adjusting device 36 may be disposed after the feature amount vector creation calculator 34 in FIG. The method of making the dynamic range uniform has been described in the operation of the third method (17).
As shown in the equation, the maximum value was made constant.

The specific operation will be described with reference to FIG. First, the image input device 31 is used for one gray-scale image a.
, A selective spatial frequency filtering calculator 32, an averaging calculator 33, and a feature vector creation calculator 34, a feature vector group having a non-uniform dynamic range is created. At this time, the maximum value for each component, which is the output of each feature vector for the same filter, is recorded. The memory for this may be installed in the dynamic range adjusting device 36. And (17)
In accordance with the formula, each component is divided by the maximum value for each component in the dynamic range adjusting device 36. The result is newly set as a feature vector group b. The division operation may be performed by a computer, or if an optical filtering device is used as the selective spatial frequency filtering arithmetic unit 32, the preceding or following stage of the selective spatial frequency filtering arithmetic unit 32 is determined based on the maximum value. The light amount may be adjusted by inserting an ND filter or the like into the device. Further, when an analog circuit is used, a current voltage value may be adjusted by using a divider or changing a resistance value. In the above-described operation, the dynamic range of the feature amount vector is adjusted after inputting one gray-scale image a. However, the same maximum value may be used for each component after inputting a plurality of gray-scale images a. .
If the types of images at the time of learning and at the time of recognition do not change much (for example, only similar X-ray images are always targeted), the maximum value of each component is recorded at the time of learning, and at the time of recognition, Each component may be divided using the maximum value of each component at the time of learning. In addition, the equation to make constant is (18)
Equation (19) may be used. In this case, a minimum value, an average value, and the like may be recorded in addition to the maximum value.

As described above, the dynamic range of each component of the feature vector is made uniform, and as a result, the low-frequency component and the high-frequency component in the image can be similarly contributed to the classification. When the groups were classified in the learning recognition step 2 (learning recognition device 40) to extract massive shadows, a good result with less noise was obtained. Specifically, the ratio of the area of the blocky shadow to the area (that is, noise) extracted outside the blocky shadow is about 0.6 times that in the first embodiment. It should be noted that the smaller the ratio is, the better the result is that the noise is reduced.

[Third Embodiment] This embodiment corresponds to the fourth method. First, before describing the present embodiment, the labeling sub-step 21 will be described more specifically. As described above, in the labeling sub-step 21,
A label is assigned to the weight vector group c using the feature amount vector group b (the label is assigned by a priori knowledge as described in the first embodiment). Specifically, the similarity between one of the feature amount vector groups b and the weight vector group c is calculated, the most similar weight vector is selected, and the feature vector is determined for the selected weight vector. A label of the quantity vector (in this case, whether or not it is a specific part) is given. Therefore, this labeling operation is performed by the similarity calculation sub-step 201 and the weight vector selection sub-step 202.
2, the similarity calculation unit 42 and the weight vector calculation unit 43 may be shared as the label assignment calculation unit 41, or the same unit may be newly provided.

Next, the operation of the present embodiment will be described. In the labeling sub-step 21 in the first embodiment, as shown in D-1) in the description of the operation of the fourth method, when the number of feature vectors from within a specific part to be extracted in the image is small, A label of a weight vector is assigned using only the feature amount vector, and a label other than the specific portion is automatically assigned to a weight vector to which no label is assigned. Considering the device, for example, a feature vector group b
Is recorded in the feature amount vector group memory 35 separately from those inside the specific portion and those outside the specific portion, and the labeling arithmetic unit 41 stores the feature amount vector from within the specific portion. And extract the weight vector group c
You can label them. When the labeling method described in D-2) in the description of the operation of the fourth method is used, both methods are used. In this case, the ease of labeling is adjusted (for example, a coefficient suitable for the similarity). , Etc.) so that the label in the specific portion is easily applied.

Taking the case where 400 weight vector groups c are used as an example, for a gray-scale image including a specific part,
When the above-described standard labeling method is used, the ratio of the number of weight vectors to which a label is assigned to a specific part to the number of weight vectors to which labels other than the specific part are assigned is 7: 393.
On the other hand, when the labeling method described in D-1) was performed, the ratio was 20 to 380. As a result, the number of weight vectors to which the label of the specific part was given increased, and the result of extracting the massive shadow further reduced the noise.
Specifically, the ratio of the area of the blocky shadow to the area (that is, noise) extracted outside the blocky shadow is about 0.5 times as compared with the case of the first embodiment.

[Fourth Embodiment] FIG. 10C shows a filter used in the selective spatial frequency filtering sub-step 12 in the feature amount extraction step 1 of the first embodiment.
A simple aperture in the frequency space as described above was used. As a result, particularly when an optical filtering device is used as the selective spatial frequency filtering calculator 32, it is only necessary to create a simple aperture as a filter, so that the creation of the filter is facilitated. Further, as described above, in the real space, the localization is smaller than in the case of FIG. 10A, and as a result, the filter becomes a filter that extracts a wider range of information in the image. Since the feature amount becomes a more global feature amount, when the classification is performed in the learning and recognition step 2 using the feature amount vector group and a massive shadow having a certain size is extracted, a better result with less noise is obtained. . Specifically, the ratio of the area of the blocky shadow to the area (ie, noise) extracted outside the blocky shadow is about 0.7 compared to the case of the first embodiment.
Doubled.

The above embodiments may be practiced independently, but a plurality of them may be used in combination, in which case higher effects can be obtained.

As is apparent from the above, the specific portion extracting apparatus for gray-scale images for implementing the method of the present invention can be constituted, for example, as follows. [1] Image input means for inputting an input image composed of shading information, and selective spatial frequency filtering for selectively responding to the input image in a specific direction in real space and a specific size in real space in the image And a selective spatial frequency filtering operation means for performing the filtering operation, and in the image of the filtering result, a specific pixel and a plurality of pixels around the specific pixel, and a size proportional to the specific size of the real space is determined. Averaging means for averaging the result of the filtering in a window having the same, and outputting the averaged value as a characteristic amount of a specific pixel in the window; and at least two or more different selective spaces different from each other. A vector obtained by using the averaged output to the frequency filtering operation unit as a component of a feature amount of the specific pixel is obtained. A feature amount extraction device including a feature amount vector creation calculation unit to be created, a feature amount vector obtained by the feature amount extraction device, and a pixel from which the vector is extracted is a specific portion to be extracted in the input image. The correspondence between the feature amount vector and the teacher signal is learned using a teacher signal indicating whether or not the feature amount vector is input, and the pixel from which the feature amount vector is extracted is included in the image. And a learning / recognition device including a neural network that outputs a signal indicating whether or not the specific portion is to be extracted.

[2] In the above item [1], the neural network comprises a similarity calculating means for calculating a similarity between the feature quantity vector and a weight vector set in the neural network. The recognizing device includes a label assignment calculating unit that assigns a label indicating a specific part to be extracted to at least one weight vector in advance, and the learning recognizing device further selects a weight vector having a large similarity. Weight vector selection calculation means, label detection calculation means for detecting a label assigned to the selected weight vector, weight vector update calculation means for updating the weight vector using the label and the teacher signal, Label output means for outputting a label detected by the label detection calculation means. A specific image extraction device for gray-scale images.

[3] In the above [1] or [2],
The gray-scale image specific portion extraction device, further comprising a dynamic range adjusting unit for the feature amount extraction device to make a dynamic range of each component of the feature amount vector uniform.

[4] In the above item [2], the label assignment calculating means assigns a label indicating a specific site to be extracted according to a ratio of the size of the specific site to be extracted in the input image. A specific portion extraction device for a gray-scale image, wherein the ratio of the number of images is varied.

[0104]

As is apparent from the above description, according to the method for extracting a specific part for a grayscale image of the present invention, B-1) the specific part of the massive shadow and the specific part of the structure other than the massive shadow from the grayscale image are accurate. Extract well. B-2) When a vector group is created while extracting a feature vector amount for each pixel from the entire area in the grayscale image, even when the vector group does not have a cluster that can be clearly classified, X A portion, such as a block shadow of a line image, in which the portion is difficult to classify clearly in the image is accurately extracted. B-3) A low-frequency component and a high-frequency component in an image are similarly contributed to classification, thereby extracting a specific portion from a more versatile image. A method for extracting a specific portion for a grayscale image that satisfies the above condition can be obtained.

[Brief description of the drawings]

FIG. 1 is a diagram showing a processing flow of a first embodiment of a method for extracting a specific part for a grayscale image of the present invention.

FIG. 2 is a diagram showing a configuration of a gray-scale image specific site extraction device for performing the method of the first embodiment.

FIG. 3 is a schematic diagram of an X-ray image of a lung field.

FIG. 4 is a schematic diagram of massive shadows and noise extracted from an X-ray image of a lung field.

FIG. 5 is a diagram showing a main part of a processing flow of a second embodiment of the grayscale image specific site extraction method of the present invention.

FIG. 6 is a diagram showing a configuration of a main part of a gray-scale image specific site extraction device for performing the method of the second embodiment.

FIG. 7 is a diagram illustrating a relationship between a weight vector, input data, and a boundary line.

FIG. 8 is a diagram for explaining another conventional method for extracting only a specific portion from a grayscale image, which is a premise of the present invention.

FIG. 9 is a diagram for explaining a real space direction and a real space size.

FIG. 10 shows a Gabor Filter and its modified Fi.
It is a figure which shows the outline of liter.

FIG. 11 is a diagram for explaining a classification process according to the method of FIG. 8;

FIG. 12 is a diagram illustrating a case where clusters are not distributed so as to be clearly classified in a feature amount vector group space.

FIG. 13 is a diagram for explaining one method of extracting only a specific part from a gray-scale image in the related art.

[Explanation of symbols]

 a: Grayscale image b: Feature vector group c: Weight vector group d: Teacher signal e: Label 1, 1 '... Feature extraction step 11 ... Image input substep 12 ... Selective spatial frequency filtering substep 13 ... Averaging Substep 14 ... Feature vector creation substep 15 ... Dynamic range adjustment substep 2 ... Learning recognition step 21 ... Label assignment substep 22 ... Learning substep 23 ... Recognition substep 201 ... Similarity calculation substep 202 ... Weight vector selection Substep 203 ... Label detection substep 204 ... Weight vector update substep 30, 30 '... Feature amount extraction device 31 ... Image input device 32 ... Selective spatial frequency filtering calculator 33 ... Averaging calculator 34 ... Feature value vector creation Arithmetic unit 35 ... Feature vector Memory for group 36 ... dynamic range adjustment device 40 ... learning recognition device 41 ... labeling calculation device 42 ... similarity calculation calculation device 43 ... weight vector selection calculation device 44 ... label detection calculation device 45 ... label output device 46 ... teacher signal Memory 47 ... Weight vector update calculator 48 ... Memory for weight vector group 50 ... Specific part extraction device for gray image

Claims (4)

[Claims] 1. An image input sub-step of inputting an input image comprising shade information, and selectively responding to the input image to a specific direction in real space and a specific size in real space in the image. A selective spatial frequency filtering sub-step of performing spatial frequency filtering, and in the image of the filtering result, the image is composed of a specific pixel and a plurality of pixels around the specific pixel, and is proportional to a specific size of the real space. An averaging sub-step of averaging the results of the filtering in a window having a size and outputting the averaged value as a feature value of a specific pixel in the window; The averaged output to the spatial frequency filter as a component of the characteristic amount of the specific pixel. A feature vector extraction sub-step for creating a torque vector, a feature vector obtained in the feature vector extraction step, and a specific region in which a pixel from which the vector is extracted is to be extracted in the input image. A learning sub-step of learning the correspondence between the feature amount vector and the teacher signal using a neural network using a teacher signal of whether or not the feature amount vector is input to the neural network. A recognition sub-step of outputting a signal as to whether or not the pixel from which the feature amount vector is extracted is a specific portion to be extracted in the image. Specific site extraction method. 2. The learning or recognition method according to claim 1, wherein the neural network used in the learning / recognition step calculates a similarity between the feature amount vector and a weight vector group set in the neural network. A neural network, wherein the learning and recognizing step includes a label assigning sub-step of assigning a label indicating a specific part to be extracted to at least one weight vector in advance, and calculating a similarity between the feature amount vector and the weight vector group. A similarity calculation sub-step, a weight vector selection sub-step of selecting a weight vector having a large similarity, a label detection sub-step of detecting a label given to the selected weight vector, A weight vector for updating the weight vector using a teacher signal. A learning update step comprising learning a correspondence between the feature amount vector and the teacher signal; a similarity calculation substep of calculating a similarity between the feature amount vector and the weight vector group; A weight vector selection sub-step of selecting a weight vector having a large similarity, and a label detection sub-step of detecting a label assigned to the weight vector, wherein the pixel from which the feature amount vector is extracted is included in the image. And a recognition sub-step of outputting a signal indicating whether or not the specific portion is to be extracted. 3. The feature amount extracting step according to claim 1, wherein the feature amount extracting step further includes a dynamic range adjusting sub-step in order to equalize a dynamic range of each component of the feature amount vector. A method for extracting a specific part for a grayscale image. 4. The weighting vector according to claim 2, wherein the label assigning sub-step includes the step of assigning a label indicating the specific portion to be extracted according to a ratio of the size of the specific portion to be extracted in the input image. A specific portion extraction method for a gray-scale image, wherein a ratio occupied by the specific region is varied.