JP7154537B2 - Image feature extraction method and image feature extraction device - Google Patents

Description

The present disclosure relates to technology for extracting image features from an image of a flow field such as a combustion space.

For example, in the operation of a thermal power plant such as a conventional boiler, operators may judge the combustion state in the furnace based on the results of observation by the camera inside the furnace or by visual observation, but the judgment criteria are determined by the operator. different. For this reason, when the combustion condition deteriorates due to an increase in NOx or unburned ash, it may be necessary to dispatch a skilled instructor to the site to diagnose the combustion condition. Problems such as lower efficiency will arise. In addition, the combustion state can be determined from state quantities such as unburned content in ash, NOx concentration, and CO concentration. It takes time to obtain measurement results, such as the need to heat with a meter and measure the weight change at that time (JIS M 8815). For this reason, it is difficult to use the determination result of the combustion state based on the measurement result of the state quantity for the operation control in real time. It is also possible to diagnose the combustion state from the distribution of combustion gas temperature and concentration obtained by using laser CT measurement, but the drawback is that the laser equipment is expensive.

For example, Patent Literature 1 discloses a combustion determination method capable of reducing the time delay in diagnosing the combustion state caused by the time it takes to measure state quantities such as unburned content in ash. Specifically, the O ₂ concentration at the combustion chamber outlet, the pressure inside the combustion chamber, the temperature at the combustion chamber outlet, and the like are measured. In addition, a color image obtained by imaging the combustion state is image-processed, divided into three-component images of red, green, and blue components, and colorimetric system conversion is performed. ). Then, based on the degree of compatibility with typical combustion conditions obtained using the extracted characteristic parameters and process state quantities as inputs to the neural network, judgments such as deterioration of combustion, appropriate combustion, and active combustion are made.

In addition, although it is a different method from Patent Document 1, Patent Documents 2 and 3 disclose a combustion state detection method using a combustion image, and Patent Document 4 uses a model for NOx and unburned content in ash. evaluation method is disclosed.
Further, Patent Document 5 discloses a three-dimensional high-order local autocorrelation feature method as a method for extracting a feature amount of a moving image. Stereoscopic high-order local autocorrelation features are obtained by calculating local autocorrelation features at each point in voxel data (three-dimensional data) in which images are arranged in time series, and converting these local features into voxel data. It is a statistical feature obtained by integrating over the whole (see Patent Document 5).

JP-A-8-49845 Japanese Unexamined Patent Application Publication No. 2013-210108 JP-A-4-315929 JP-A-4-214120 Japanese Patent No. 4368767

A method in which a person predetermines what features (image features) are to be extracted from which part of an image of fuel combustion (hereinafter referred to as a combustion image). It is difficult to formalize (explicitly) an expert's tacit knowledge, for example, it is difficult to describe all the characteristics of In this regard, the present inventors extracted the geometric features contained in the combustion image, and derived the correlation between the extracted features and the combustion state by machine learning. We thought that it would be possible to reproduce the combustion diagnosis method of a person.

In a combustion image, for example, a relatively dark area (dark area) corresponds to a portion in which the combustion state is not good, so it is useful for judging the behavior of the unburned area of the flame. Based on this knowledge, the inventors of the present invention found that if the area (portion) that characterizes the combustion state in the combustion image is highlighted (emphasized) by image processing, the characteristics of the combustion state can be identified from the combustion image after image processing. Since it will be possible to extract more accurately, it will be possible to better derive the correlation between the characteristics of the combustion image (the pattern of the combustion state that appears in the combustion image) and the combustion state, and to evaluate the combustion state more accurately from the combustion image. I found that it is possible. In addition, the method of highlighting features before extracting features from an image as described above is not limited to combustion images, and is based on geometric features included in images obtained by photographing flow fields in which fluid flows, It can also be applied to derive the correlation with the flow state.

In view of the above circumstances, it is an object of at least one embodiment of the present invention to provide an image feature extraction method for extracting geometric features appearing in an image with high accuracy.

(1) An image feature extraction method according to at least one embodiment of the present invention,
For each of a plurality of target images to be processed obtained by imaging a flow field, based on the brightness distribution of the target image, the contrast is obtained by amplifying the difference between the brightness values in the brightness distribution. a preprocessing step of obtaining at least one preprocessed image through enhancement of
a feature amount calculation step of calculating, from the at least one preprocessed image, at least one feature amount obtained by quantifying image features of temporal and spatial changes in the flow field.

According to the configuration (1) above, for each of a plurality of target images, a post-preprocessing image is obtained by amplifying a difference in height between a plurality of luminance values constituting the luminance distribution, and from the post-preprocessing image, Calculate the feature amount. For example, in general, an image is composed of a plurality of pixels, and a threshold is set based on the distribution of luminance values (luminance distribution) of each region when the target image is divided into a plurality of regions, for example, for each pixel. , and after emphasizing the contrast based on the set threshold value, the feature amount is calculated. In this way, by enhancing the contrast for each target image, the characteristics of the flow field (fluid state) appearing in the target image can be further emphasized (emphasized), so the feature amount is calculated from the preprocessed image. By doing so, the features of the flow state can be extracted with higher accuracy.

Therefore, for example, based on the image characteristics of an image (combustion image) obtained by imaging a combustion space (combustion state), which is a type of flow field, changes such as unburned content in ash, NOx concentration, CO concentration, etc. according to the combustion state When evaluating the combustion state based on the image features of the combustion image, such as when estimating the state quantity of the combustion image, if the image features are calculated from the preprocessed image, the combustion image and the combustion state can be more appropriately matched. A correlation can be obtained, and the evaluation accuracy of the combustion state can be improved, such as improving the accuracy of estimating the state quantity.

(2) In some embodiments, in the configuration of (1) above,
The target image is an image that constitutes a partial area of the captured image that is the captured image of the flow field.
According to the configuration (2) above, each target image is an image that configures (forms) a partial region of the captured image (captured image). As a result, if the captured image contains parts that are unrelated to the flow state (such as obstructions that interfere with imaging of the flow state), this prevents the influence of that part on the preprocessing results and prevents the captured image from A portion in which the fluid state appears more conspicuously can be set as the target image, and the contrast can be emphasized more appropriately.

(3) In some embodiments, in the configuration of (2) above,
The partial area of the captured image includes an area positioned at the center of the captured image.
According to the above configuration (3), each target image is an image of a partial area including an area positioned at the center of the captured image (captured image). As a result, it is possible to perform mechanical processing so that the target image includes an area in which the fluid state is appropriately captured in the captured image. Therefore, by performing contrast enhancement on such a target image, the image features of the target image can be appropriately emphasized.

(4) In some embodiments, in the configurations of (1) to (3) above,
The luminance distribution is a distribution of the luminance values of the target image in grayscale.
According to the configuration (4) above, preprocessing is performed on a grayscale (black and white) target image. Generally, in the case of a color image, brightness data exists for each of the three types of RGB, but when the captured image is a color image, it is grayscaled. Since the target image is grayscale, it is possible to reduce the number of brightness information for each pixel, such as using only one type of brightness information, thereby facilitating preprocessing of the target image.

(5) In some embodiments, in the configurations of (1) to (4) above,
The pretreatment step includes
a threshold setting step of setting a threshold of the luminance value corresponding to each of the plurality of target images based on the luminance distribution of each of the plurality of target images;
and a contrast enhancement step of executing the contrast enhancement using the brightness value threshold corresponding to each of the plurality of target images for each of the plurality of target images.
According to the above configuration (5), the threshold value of the brightness value is set for each target image based on the brightness distribution thereof, and contrast enhancement is performed using this threshold value. As a result, the dark region in the target image can be made more conspicuous, and the feature of the fluid state shown in the target image before image processing can be emphasized in the post-preprocessing image.

(6) In some embodiments, in the configuration of (5) above,
The contrast enhancement step includes:
For each of the plurality of target images, a luminance value having a value smaller than the threshold among the plurality of luminance values forming the luminance distribution is changed to a predetermined value smaller than the threshold.
According to the above configuration (6), among the luminance values of each region such as each pixel constituting each target image, the luminance value having a value smaller than a threshold value set according to the target image is set to the minimum value (for example, 0). to a predetermined value such as As a result, the dark region in the target image can be maximally emphasized (emphasized).

(7) In some embodiments, in the configuration of (6) above,
The pretreatment step includes
further comprising, after the contrast enhancement step, a differential image generating step of generating a differential image of the two target images on which the contrast enhancement has been performed;
The differential image generated in the differential image generating step is obtained as the preprocessed image.
With configuration (7) above, by using the preprocessed image as the difference image, the image feature can be calculated from the difference image.

(8) In some embodiments, in the configurations of (1) to (7) above,
The feature amount calculation step includes:
The at least one feature amount is calculated from the at least one preprocessed image by performing processing using a stereoscopic high-order local autocorrelation feature method.
With configuration (8) above, it is possible to appropriately calculate the feature amount of the flow state appearing in a plurality of target images from at least one preprocessed image by the three-dimensional high-order local autocorrelation feature method.

(9) In some embodiments, in the configurations of (1) to (8) above,
The feature amount calculating step is adapted to calculate a plurality of the feature amounts,
The method further includes a feature amount selection step of selecting the at least one feature amount from among the plurality of feature amounts obtained by the feature amount calculation step.
According to the above configuration (9), for example, when a plurality of feature amounts are calculated by a method such as a three-dimensional high-order local autocorrelation feature method that obtains a plurality of feature amounts from a plurality of target images, the calculated plurality of features, for example, One or more desired feature values reflecting changes in the flow state, such as a portion of the quantity, are selected based on a predetermined criterion. As a result, it is possible to extract a feature quantity or the like that accurately reflects the features of the flow state appearing in the plurality of target images.

(10) In some embodiments, in the configurations of (1) to (9) above,
The image features obtained by preparing a plurality of sets of the plurality of target images including the target images different from each other, and performing the preprocessing step and the feature amount calculation step on each of the plurality of sets. It further comprises an averaging processing step of performing an averaging process on each set of the feature amounts for each set.
According to the above configuration (10), the averaging process is performed for each set of feature amounts for each image feature obtained by shifting the time. If the feature amount F changes sharply, it tends to become noise. By performing the averaging process, the noise can be suppressed and the change in the value of the feature F can be appropriately obtained for each image feature.

(11) In some embodiments, in the configuration of (10) above,
The averaging process is a process of calculating a moving average for each set or a process of calculating an interval average for each set.
With configuration (11) above, by performing a moving average or an interval average as the averaging process, it is possible to appropriately obtain changes in the value of the feature amount for each image feature.

(12) In some embodiments, in the configurations of (10) to (11) above,
The feature quantity is for use in estimating the state quantity of the flow state indicated by the plurality of target images,
The method further comprises an averaged feature amount selection step of selecting a feature amount having a high degree of contribution to the estimation of the state amount from among the plurality of averaged feature amounts.
According to the configuration (12) above, one or more feature quantities having a high degree of contribution to the estimation of the state quantity (the degree of contribution to the measured value of the state quantity) are selected by, for example, principal component analysis. For example, by learning the relationship between one or more feature values obtained in this way and the state quantities (measured values, etc.) of the flow state appearing in a plurality of target images from which the feature values are extracted, a plurality of target A relationship between the image feature and the state quantity S, which is particularly closely related to the state quantity of the fluid state appearing in the image, can be obtained. Therefore, based on the relationship, it is possible to accurately estimate the state quantity in the desired target image from the selected image feature.

(13) In some embodiments, in the configurations of (1) to (12) above,
The plurality of target images constitute at least part of captured images obtained by capturing the same range in the incinerator at different times during fuel combustion.
With configuration (13) above, it is possible to extract the characteristics of the combustion state appearing in a plurality of target images.

(14) An image feature extraction device according to at least one embodiment of the present invention,
An image feature extraction device for image feature extraction,
For each of a plurality of target images to be processed obtained by imaging a flow field, based on the brightness distribution of the target image, the contrast is obtained by amplifying the difference between the brightness values in the brightness distribution. a preprocessing unit that obtains at least one preprocessed image through enhancement of
a feature amount calculation unit that calculates at least one feature amount obtained by quantifying image features of temporal and spatial changes in the flow field from the at least one preprocessed image.

According to the configuration of (14) above, the same effect as that of (1) above is achieved.

According to at least one embodiment of the present invention, there is provided an image feature extraction method for extracting geometric features appearing in an image with high accuracy.

FIG. 2 illustrates an image feature extraction method according to an embodiment of the present invention; 1 is a diagram showing an image feature extraction method according to an embodiment of the present invention, comprising a feature amount selection step; FIG. Fig. 3 shows an image feature extraction method according to an embodiment of the present invention, comprising an analysis step (S4); It is a figure which shows the captured image, the post-preprocessing image, and the transition of luminance distribution of these, (a) is a captured image, (b) is a post-preprocessing image which concerns on one Embodiment of this invention. It is a figure which shows the state quantity estimation method which concerns on one Embodiment of this invention. It is a block diagram which shows the function of the state quantity estimation apparatus which concerns on one Embodiment of this invention. 1 is a diagram schematically showing the configuration of a combustion furnace in which an image feature extraction device according to one embodiment of the present invention is installed; FIG. 3 is a block diagram showing functions of an image feature extraction device according to an embodiment of the present invention, which executes processing corresponding to FIG. 2. FIG. 4 is a block diagram showing functions of an image feature extraction device according to an embodiment of the present invention, which executes processing corresponding to FIG. 3. FIG.

Several embodiments of the present invention will now be described with reference to the accompanying drawings. However, the dimensions, materials, shapes, relative arrangements, etc. of the components described as embodiments or shown in the drawings are not intended to limit the scope of the present invention, and are merely illustrative examples. do not have.
For example, expressions denoting relative or absolute arrangements such as "in a direction", "along a direction", "parallel", "perpendicular", "center", "concentric" or "coaxial" are strictly not only represents such an arrangement, but also represents a state of relative displacement with a tolerance or an angle or distance to the extent that the same function can be obtained.
For example, expressions such as "identical", "equal", and "homogeneous", which express that things are in the same state, not only express the state of being strictly equal, but also have tolerances or differences to the extent that the same function can be obtained. It shall also represent the existing state.
For example, expressions that express shapes such as squares and cylinders do not only represent shapes such as squares and cylinders in a geometrically strict sense, but also include irregularities and chamfers to the extent that the same effect can be obtained. The shape including the part etc. shall also be represented.
On the other hand, the expressions "comprising", "comprising", "having", "including", or "having" one component are not exclusive expressions excluding the presence of other components.

FIG. 1 is a diagram showing an image feature extraction method according to one embodiment of the present invention. FIG. 2 is a diagram showing an image feature extraction method according to an embodiment of the present invention, comprising a feature amount selection step. FIG. 3 is a diagram showing an image feature extraction method according to an embodiment of the present invention, comprising an analysis step (S4). 4A and 4B are diagrams showing the captured image Pa, the preprocessed image G, and the transition of the luminance distribution of these according to one embodiment of the present invention, where (a) is the captured image Pa and (b) is the preprocessed This is the back image G.

The image feature extraction method is a method for extracting features of an image (hereinafter referred to as image features) obtained by imaging a flow field (flow field) where fluid flows. Specifically, the flow field includes a combustion chamber (combustion space 8) in a combustion furnace such as a garbage incinerator (see FIG. 7 described later), a boiler, and a gas furnace that burns fuel, and an engine (for example, diesel engine) or gas turbine combustor. Common problems with these are that the difference in flame behavior due to changes in the combustion state in the combustion chamber is minute, making it difficult to identify, and knowledge and experience are required to judge the quality of the combustion state from the appearance of the flame. There are issues. In addition, since the combustion space 8 of the combustor is cleaner than the combustion space 8 of the combustion furnace, it is difficult to photograph the flame due to soot and dust as a problem specific to the combustion furnace. Alternatively, the flow field may be a place where air flows, such as inside a supercharger installed in an engine.

Further, the image feature is the feature of the flow state of the fluid appearing in an image obtained by capturing the flow field (hereinafter, appropriately, captured image Pa). The captured image Pa may be either a color image or a monochrome image, and may be a frame (image) forming a still image or a moving image. For example, if the above image is an image (combustion image) captured inside the incinerator (combustion space 8) during combustion of fuel, the image feature is the feature of the combustion state appearing (shown) in the image. etc.

Then, as shown in FIGS. 1 to 3, the image feature extraction method includes a preprocessing step (S2) and a feature amount calculation step (S3).
The image feature extraction method of the present invention will be described below by taking as an example a case where the above image is a combustion image.

In the preprocessing step (S2), a target image P based on the luminance distribution H of the luminance distribution H, at least one post-enhancement image (hereinafter referred to as post-preprocessing Obtaining image G). That is, in the preprocessing step (S2), image processing is performed to highlight (emphasize) the features of the complicated combustion phenomenon appearing in each of the plurality of images.

More specifically, the plurality of target images P are images of the same range in the incinerator taken at different times during fuel combustion. In general, an image is configured by arranging a plurality of picture elements (pixels) in a grid pattern. Although details will be described later, in some embodiments, as shown in FIG. 1 to FIG. (hereinafter referred to as unit regions R), and for each target image P, a luminance distribution H of luminance values of each of a plurality of unit regions R forming the target image P may be calculated. This luminance distribution H may be a histogram showing the relationship between luminance values and their frequencies. Thereafter, a threshold value L may be set (determined) for each target image P based on this luminance distribution H, and the contrast of each target image P may be enhanced based on the set threshold value L. FIG. In some other embodiments, contrast enhancement may be by other well-known methods, such as multiplying the luminance value of each unit area R by n times.

The unit region R may be composed of, for example, one pixel, which is the minimum unit that constitutes (forms) an image, or may be composed of a plurality of pixels. When each unit region R is composed of a plurality of pixels, the number of pixels included in each of the plurality of unit regions R may be the same, or at least one unit region R may include the other unit regions R may be different from Further, the luminance value of the unit region R in this case is a statistical value obtained by statistically processing a plurality of luminance values, such as an average value of luminance values of a plurality of pixels belonging to the same unit region R. , or an arbitrarily selected luminance value.

Then, by enhancing the contrast of each target image P in this preprocessing step (S2), as shown in FIG. , the relative darkness increases in the post-preprocessing image G (FIG. 4(b)), and the relatively bright portions (hereinafter referred to as high-brightness regions Rh) in the target image P are relatively dark in the post-preprocessing image G. brightness increases. For this reason, by using an estimation model M (described later), for example, unburned content in ash, NOx concentration, CO concentration, etc., from the feature amount F of the combustion state feature (image feature) appearing in the plurality of target images P When estimating the state quantity S, it is possible to extract the feature quantity F from the preprocessed image G with higher accuracy than from the target image P.

In the feature amount calculation step (S3), image features of temporal and spatial changes in the flow field are digitized from at least one preprocessed image G obtained by performing the preprocessing step (S2). It is a step of calculating at least one feature value F. The image feature is an index capable of capturing (quantifying) temporal changes in combustion states (flow states) obtained from a plurality of target images P based on luminance information. Also, the feature quantity F is a numerical representation of the image feature. The brightness information is information obtained from the target image P without detecting the flame itself. Even in this case, the feature quantity F can be obtained. Therefore, it is also possible to facilitate installation of the imaging device 6 (in-furnace camera) for imaging the inside of the furnace.

Specifically, the image features are brightness values, shapes, sizes, colors, shadings, temperatures (temperature distributions), and wavelengths obtained from brightness information appearing in the target image P or obtained based on the target image P. It may be a feature indicating temporal change of at least one piece of information such as (wavelength distribution), or it may be a temporal change of a statistical value such as an average or peak of these information. Alternatively, the image feature may be any of the area, shape, and size of a portion having luminance equal to or greater than a predetermined value or equal to or less than a predetermined value. Note that the temperature, temperature distribution, and the like obtained from the luminance information are obtained by converting the luminance value.

Then, it is possible to perform evaluation corresponding to the magnitude (degree) of the change from the feature amount F of the image feature. This feature amount F is, for example, a three-dimensional high-order local autocorrelation feature amount (CHLAC feature amount) obtained by performing processing by a three-dimensional high-order local autocorrelation feature method, a CNN (Convolution Neural Network), an AE (AutoEncoder), or the like. It may be the amount of the image feature obtained by The CHLAC feature quantity has the advantage of being able to compactly express spatio-temporal variations that appear in the entire target image P. CNN has the advantage of being able to effectively reduce and acquire the image features of the entire target image P by obtaining features through convolution processing, pooling processing, or encoding processing in AE.

In other words, the plurality of target images P are a plurality of images obtained by imaging the same range in the flow field at different times (at intervals of time). Spatial changes such as three-dimensional space can be seen. Therefore, in the feature quantity calculation step (S3), based on a plurality of target images P, the characteristics of temporal and spatial changes in the flow field are discriminated, and the features are quantified to obtain the feature quantity F get

In the embodiment shown in FIGS. 1 to 3, in step S1, a plurality of target images P are acquired by reading them from a storage device or the like (target image acquisition step), and in step S2, the acquired multiple target images P are acquired. , respectively, perform the contrast enhancement described above. Then, in step S3, the feature amount F is calculated based on the preprocessed image G obtained by executing step S2.

According to the above configuration, the post-preprocessing image G is obtained by amplifying the height difference between the plurality of luminance values forming the luminance distribution H for each of the plurality of target images P, and the post-preprocessing image G is obtained. A feature amount F is calculated from G. In this way, by enhancing the contrast for each target image P, the characteristics of the flow field (flow state) appearing in the target image P can be further emphasized (emphasized). By calculating the quantity F, the characteristics of the flow state can be extracted with higher accuracy.

Therefore, for example, when estimating the state quantity S such as the unburned content in the ash, the NOx concentration, the CO concentration, etc., which changes according to the combustion state, based on the image features of an image (combustion image) that captures the combustion state. When the combustion state is evaluated based on the image features of the combustion image, if the image features are calculated from the preprocessed image G, a more appropriate correlation between the combustion image and the combustion state can be obtained. It is possible to improve the evaluation accuracy of the combustion state, such as improving the estimation accuracy of the quantity S.

Next, the target image P described above will be described in more detail.
In some embodiments, each of the plurality of target images P described above is an image that constitutes (forms) at least a partial region of each of the plurality of captured images Pa obtained by imaging the flow field. can be In other words, the captured image Pa may be an image of the flow field itself, and the target image P may be an image forming a part of the captured image Pa. In this case, in the above-described image feature extraction method, in the target image acquisition step (S1 in FIG. 1), a plurality of target images P forming at least a partial region of each of the plurality of captured images Pa are acquired. That is, the target image P may be the captured image Pa itself, or may be an image obtained by cutting out a partial area of the captured image Pa. In the feature extraction, it is not always necessary to target the entire captured image Pa, and the area of the captured image Pa to be the target image P can be selected according to the purpose, such as depending on the state quantity S to be estimated. be.

For example, depending on the imaging conditions, the combustion space 8 to be imaged may be behind an object such as a window frame that hides the combustion space 8 or a shield such as letters. will be occupied by the shield. In addition, from the part of the combustion space 8 where the combustion state is not captured in the captured image Pa because the shielding object is captured together, not only is it not possible to obtain the characteristics of the combustion state, but conversely, noise ( hindrance). Therefore, the target image P may be obtained by cutting out a region in which the shielding object is not captured in the captured image Pa. Furthermore, the target image P may be an image obtained by cutting out a partial area from a partial image (intermediate image) of the captured image Pa obtained by removing a shield or the like from the captured image Pa. When the area to be included in the target image P changes according to the state quantity S to be estimated, the target image P can be obtained by holding an intermediate image and appropriately cutting out a part of the intermediate image. becomes possible.

That is, when the target image P is an image forming a part of the captured image Pa, the target image acquisition step (S1) includes a step of acquiring the captured image Pa or an intermediate image, and a step of acquiring the acquired image. It may comprise a step of cutting out (obtaining) a partial region, and may further include a step of further cutting out a partial region from the image obtained by the cutting step. In the target image acquisition step (S1), the target image P may be acquired by capturing an image of the flow field, or may be stored in a storage device or the like in advance and read. .

In some embodiments, the partial area of the captured image Pa described above may include an area (central area) located in the center of the captured image Pa. This central region may be inside a circle or ellipse having a predetermined radius from the center of the image, a rectangle tangent to the circle or ellipse, or any other shape that includes the center of the image. The center of this image may be the position of the center of gravity of the image. In the combustion space 8, the shielding object is usually at the edge of the captured image Pa, and it is possible to mechanically process so that the target image includes an area in which the flow state is appropriately captured in the captured image Pa. Become.

According to the above configuration, each target image P is an image that constitutes (forms) at least a partial region of the captured image (captured image Pa). As a result, if the captured image contains parts that are unrelated to the flow state (such as obstructions that interfere with imaging of the flow state), this prevents the influence of that part on the preprocessing results and prevents the captured image from A portion in which the fluid state appears more prominently can be set as the target image P, and the contrast can be emphasized more appropriately.

Also, in some embodiments, the target image P described above may be a grayscale (black and white image). In this case, the luminance distribution described above is the distribution of luminance values of the target image P in grayscale. As described above, the target image P may be a color image or a monochrome image. Therefore, if the captured image Pa or the intermediate image described above is a color image (not a grayscale image), it is converted to grayscale (grayscale conversion). Specifically, RGB information may be converted to grayscale by luminance preservation conversion. Each layer of RGB may be separated and one of the layers may be converted to grayscale. Alternatively, grayscaling may be performed using information of a color space other than RGB (such as HSV) that can be converted from RGB information.

In the embodiments shown in FIGS. 1 to 3, the target image P is grayscaled (S1a in FIG. 1) before the preprocessing step (S2) described above. Before calculating the distribution H, the target image P may be grayscaled. Alternatively, the target image acquisition step (S1) may acquire a grayscaled target image P. In this case, a grayscale step may be performed before the target image acquisition step (S1). A grayscale image may be obtained, for example, when the captured image Pa captured as a color image is converted to grayscale and stored.

According to the above configuration, the preprocessing is performed on the grayscale (black and white) target image P. Generally, in the case of a color image, brightness data exists for each of the three types of RGB, but when the captured image Pa is a color image, it is grayscaled. Since the target image P is grayscale, it is possible to reduce the number of brightness information for each pixel, such as using only one type of brightness information, thereby facilitating preprocessing of the target image P. FIG.

Next, the preprocessing step (S2) included in the image feature extraction method described above will be described in detail.
In some embodiments, the preprocessing step (S2 in FIG. 1) comprises a luminance distribution obtaining step (S21), a threshold setting step (S22) and a contrast enhancement step (S23).

The brightness distribution acquisition step (S21) is a step of acquiring (calculating) the brightness distribution H of each of the plurality of target images P. FIG. As described above, the image data has luminance information for each of a plurality of pixels forming an image. Based on the luminance information for each pixel of the target image P (data), are obtained, and a distribution of the plurality of obtained luminance values (luminance distribution H) is obtained.

The threshold setting step (S22) is a step of setting a threshold value L (hereinafter referred to as a threshold value L) for each of the plurality of target images P based on the luminance distribution H of each of the plurality of target images P. is. In some embodiments, as shown in FIGS. 1 to 3, the threshold value L of each target image P is determined by Otsu's threshold discrimination method (Otsu's binarization method). Then, based on the threshold value L of the target image P, a plurality of unit regions R constituting each target image P are classified into high-luminance regions Rh, which are classes having a luminance value greater than the threshold value L, and luminance values equal to or less than the threshold value L. It is classified into two classes, the dark area Rl which is a class having a value. In Otsu's threshold discrimination method, the degree of separation defined by the intra-class variance and the inter-class variance when a brightness histogram (in this embodiment, a histogram) is divided into two classes with an arbitrary brightness value as a boundary is Let the threshold value L be the luminance value when it becomes the maximum.

However, the present invention is not limited to this embodiment. The threshold value L may be determined by another method as long as a plurality of unit regions R forming a part of each target image P can be divided into two regions, ie, a high brightness region Rh and a dark region Rl. For example, in some other embodiments, the threshold value L is the average value, the mode value, the median value, etc. It may be a statistical value of

The contrast enhancement step ( S<b>23 ) is a step of performing the contrast enhancement described above for each of the plurality of target images P using the threshold value L of the luminance value corresponding to each of the plurality of target images P. Specifically, in some embodiments, this contrast enhancement step (S23) is performed such that, for each of the plurality of target images P, A preprocessed image G obtained by preprocessing the target image P (see FIG. 4A) may be generated by changing the luminance value having a small value to a predetermined value smaller than the threshold value L. FIG. The predetermined value may be the minimum luminance value. Alternatively, in the contrast enhancement step, for each of the plurality of target images P, among the plurality of luminance values forming the luminance distribution H, the luminance value having a luminance value larger than the threshold value L is set to a predetermined value larger than the threshold value L. , a preprocessed image G obtained by preprocessing the target image P (see FIG. 4A) may be generated. The predetermined value may be the maximum luminance value. Alternatively, in this contrast enhancement step, for each of the plurality of target images P, among the plurality of luminance values forming the luminance distribution H, the luminance value having a luminance value larger than the threshold value L is set to a predetermined value larger than the threshold value L. , and by changing the luminance value having a luminance value smaller than the threshold L to a predetermined value smaller than the threshold L, the target image P (see FIG. 4A) is preprocessed. A post image G may be generated.

In this way, when the plurality of unit regions R are divided into the high brightness region Rh and the dark region Rl, the brightness value of each unit region R belonging to the dark region Rl is changed to a value smaller than the threshold value L, and , do not change the luminance value of the high-luminance region Rh, or change the luminance value of each unit region R belonging to the high-luminance region Rh to a value larger than the threshold value L, and do not change the luminance value of the dark region Rl. Alternatively, the brightness value of each unit region R belonging to the dark region Rl is changed to a value smaller than the threshold value L, and the brightness value of each unit region R belonging to the high brightness region Rh is changed to a value larger than the threshold value L. By changing, the difference between the luminance values belonging to each of the high-luminance region Rh and the dark region Rl can be increased. As a result, the dark region Rl or the high brightness region Rh of the target image P becomes more conspicuous in the preprocessed image G.

In the embodiment shown in FIGS. 1 to 3, the luminance value of each unit region R belonging to the dark region Rl among the luminance of each of the plurality of unit regions R forming the target image P is set to a predetermined value smaller than the threshold value L. It's supposed to change. At this time, in some embodiments, as shown in FIG. 4, the minimum luminance value is 0, and the luminance value of each unit region R belonging to the dark region Rl is set to 0. Thereby, the dark area Rl in the target image P can be emphasized (made conspicuous) to the maximum.

Referring to FIG. 4, the histogram in FIG. 4A is the distribution (histogram) of the brightness values of each unit region R when the target image P is divided into a plurality of unit regions R, and the above threshold setting step ( The histogram is divided into a high luminance area Rh and a dark area Rl with the threshold L set in S22) as a boundary. Then, the preprocessed image G shown in FIG. Looking at the corresponding histogram, it can be seen that the luminance value of each unit region R belonging to the dark region Rl is 0.

However, the present invention is not limited to this embodiment. In some other embodiments, the contrast enhancement step (S23) may subtract an arbitrary value from each unit region R. Further, in the contrast enhancement step (S23), among the brightness values of the plurality of unit regions R forming the target image P, the brightness value of each unit region R belonging to the high brightness region Rh is changed to a predetermined value larger than the threshold value L. , the predetermined value larger than the threshold value L may be the maximum value of the luminance value, or an arbitrary value is added from the luminance value of each unit region R. Also good.

According to the above configuration, the threshold value L of the brightness value is set for each target image P based on the brightness distribution H thereof, and the threshold value L is used to enhance the contrast. As a result, the dark region Rl in the target image P can be made more conspicuous, and in the post-preprocessing image G, the feature of the fluid state shown in the target image P before image processing can be further emphasized.

Also, in some embodiments, as shown in FIG. 1 (not shown in FIGS. 2-3), the preprocessing step (S2) is performed after the contrast enhancement step (S23) described above. A differential image generating step (S24) of generating a differential image of the executed two target images P may be further provided. In this case, in the preprocessing step (S2), the differential image generated in the differential image generation step (S24) is obtained as the image G after preprocessing. In this way, by using the post-preprocessing image G as the difference image, the image feature can be calculated from the difference image.

Next, the feature amount calculation step (S3) included in the image feature extraction method described above will be described in detail.
In some embodiments, the feature amount calculation step (S3) described above is obtained by performing the preprocessing step (S2) by performing processing by the cubic high-order local autocorrelation feature method (CHLAC) At least one feature amount F is calculated from at least one preprocessed image G. In general, a plurality of feature quantities F can be obtained by the three-dimensional high-order local autocorrelation feature method. Also good. By calculating only the feature amount F that is actually used for estimating the state amount S, the amount of calculation can be reduced, and the time required to calculate the feature amount F can be shortened. As a result, it is possible to appropriately extract the feature amount of the flow state appearing in the plurality of target images P from at least one preprocessed image G by the three-dimensional high-order local autocorrelation feature method.

Next, another step of the image feature extraction method that can be performed after the above feature amount calculation step (S3) will be described.
In some embodiments, as shown in FIG. 2, the above-described feature amount calculation step (S3) calculates a plurality of feature amounts F, for example, by performing processing by the cubic high-order local autocorrelation feature method (CHLAC). As shown in FIG. 2, the image feature extraction method extracts at least one feature value F (image feature) from among the plurality of feature values F obtained in the feature value calculation step (S3) described above. It further comprises a feature quantity selection step (S40) for selecting . That is, one or more feature quantities F suitable for estimating the state quantity S, for example, are selected from a plurality of feature quantities F obtained by, for example, CHLAC. The selection in this feature quantity selection step (S40) may be made based on empirical knowledge. , may also be selected in this step.

According to the above configuration, when a plurality of feature amounts are calculated by a method such as a three-dimensional high-order local autocorrelation feature method that obtains a plurality of feature amounts F from a plurality of target images, for example, the calculated plurality of feature amounts F One or more desired feature quantities F that particularly reflect changes in the flow state, such as a part of , are selected based on a predetermined criterion. As a result, it is possible to extract the feature amount F that accurately reflects the image features of the fluid state appearing in the plurality of target images P.

Further, in some embodiments, as shown in FIG. 3, the image feature extraction method includes, after the feature amount calculation step (S3), a process for selecting an appropriate feature amount F (image feature) according to the purpose. An analysis step (S4) may be further provided. More specifically, in some embodiments, as shown in FIG. 3, the image feature extraction method (analysis step) extracts a plurality of sets of target images P described above, including target images P that are different from each other. For each set of feature values F for each image feature obtained by performing the preprocessing step (S2) and the feature value calculation step (S3) on each of the plurality of sets, An averaging processing step (S41) for executing averaging processing may be further provided. The preparation of the plurality of sets may be completed before the target image acquisition step (S1) described above, or the preprocessing step (S2) and the feature amount calculation step may be performed each time one set is prepared. (S3) does not have to be completed before the target image acquisition step (S1).

For example, when a moving image is captured, a plurality of continuous still images (captured images Pa) forming the captured moving image are obtained. For such continuously obtained images, each time an arbitrary number of target images P are obtained, for example, according to a predetermined rule (continuous, every few frames, etc.), a set of these obtained images are calculated as a plurality of target images P, a plurality of feature amounts F are obtained for each type of image feature. The types of image features referred to here are image features respectively corresponding to a plurality of feature amounts F obtained by performing processing for feature extraction such as CHLAC once. Depending on the type of image feature and the calculation conditions such as the frequency of calculation of the image feature, a component that can be regarded as noise may occur due to the magnitude of variation in the feature amount F. Therefore, noise components are suppressed by performing an averaging process for each set of image features for each type.

This averaging process may be a process of calculating a moving average of a set of feature amounts F for each image feature, or may be a process of calculating an interval average of the set. For the moving average, a moving window that moves with time is used to calculate the average value of a plurality of feature values F for a predetermined period within the range of the moving window. The range of the moving window may include the latest feature amount F by including the range from the current time to the past for a predetermined period. On the other hand, for the section average, a plurality of sections are set along the time axis so that the periods do not overlap, and the average of the plurality of feature amounts F included in each section is calculated for each section. As a result, it becomes possible to appropriately obtain the change in the value of the feature amount F for each image feature, and it is possible to perform real-time processing such as extracting the image feature while acquiring the latest captured image Pa by capturing it. Become. Therefore, it becomes possible to monitor the combustion state (flow state) in real time.

According to the above configuration, the averaging process is performed for each set of feature amounts F for each image feature obtained by shifting the time. If the feature amount F changes sharply, it tends to become noise. By performing the averaging process, the noise can be suppressed and the change in the value of the feature amount F can be appropriately obtained for each image feature.

Further, in some embodiments, the feature amount F described above is used for estimating the state amount S of the combustion state (flow state) indicated by a plurality of target images P, and an image feature extraction method (analysis step) includes an averaged feature quantity selection step (S42) for selecting a feature quantity having a high degree of contribution to the estimation of the state quantity S from among the plurality of feature quantities F averaged by the averaging processing step (S41). , and may be further provided. As will be described later, machine learning is performed using a plurality of data obtained by associating a plurality of feature quantities F corresponding to each of a plurality of image features extracted by this image feature extraction method with the measured value of the state quantity S as learning data. In such a case, the degree of contribution to the measured value of the state quantity S by a plurality of feature quantities F is usually different, and one or more feature quantity F with a high degree of contribution is selected through principal component analysis or the like.

According to the above configuration, one or more feature quantities F having a high degree of contribution to the estimation of the state quantity (the degree of contribution to the measured value of the state quantity) are selected by, for example, principal component analysis. For example, by learning the relationship between one or more feature values F obtained in this manner and state quantities (measured values, etc.) of the flow state appearing in a plurality of target images from which the feature values F are extracted, a plurality of The relationship between the image feature and the state quantity S, which is particularly closely related to the state quantity S of the fluid state appearing in the target image P, can be obtained. Therefore, based on the relationship, it is possible to accurately estimate the state quantity S in the desired target image P from the selected image feature.

One or more feature quantities F extracted by the image feature extraction method described above are used for estimating the state quantity S described above (see FIG. 5). Specifically, from one or more feature quantities F extracted from a plurality of target images P for which the state quantity S is to be estimated, the desired state quantity S in the combustion state appearing in the plurality of target images P is estimated. Using the estimation model M for doing, the corresponding state quantity S is estimated from a plurality of arbitrary target images P. The presumed model M may be obtained and used after being created in advance.

FIG. 5 is a diagram showing a state quantity estimation method according to one embodiment of the present invention. In the embodiment shown in FIG. 5, in step S51, the feature amount F extracted from a plurality of target images P is acquired by the image feature extraction method described above. In this step, the feature amount F may be obtained by performing extraction using an image feature extraction method, or the already extracted feature amount F may simply be obtained. Next, the estimated model M is acquired in step S52. It may be acquired by creating an estimated model M, or may be acquired from an estimated model M that has been created and stored in advance. Then, in step S53, using the estimation model M obtained in step S52, the state quantity S is estimated (calculated) from the feature quantity F obtained in step S51. In other words, the feature quantity F is used as input information and calculated according to the estimation model M, and the state quantity S is obtained as a result. Note that the order of steps S51 and S52 may be reversed.

The state quantity estimating method described above may be executed using a dedicated device (state quantity estimating device 5) as shown in FIG. FIG. 6 is a block diagram showing functions of the state quantity estimation device 5 according to one embodiment of the present invention. As shown in FIG. 6, the state quantity estimating device 5 has functional units (feature quantity acquiring unit 51, estimation model acquiring unit 52, state quantity estimating unit 53) that execute the respective steps described above. The embodiment shown in FIG. 6 is configured to be connected to an image feature extraction device 1 (to be described later) so that a feature amount F extracted by the image feature extraction device 1 (to be described later) is input. The state quantity estimating device 5 (state quantity estimating unit 53 ) is configured to output the estimation result of the state quantity S to the display device 55 by being connected to the display device 55 such as a display. The storage device 5m may be a storage device included in the state quantity estimation device 5, or may be a storage device such as another device connected via a communication network (not shown). Also, the state quantity estimation device 5 and the image feature extraction device 1 (described later) may be composed of one device that operates on the same platform such as the same OS (Operating System).

In addition, the above estimation model M is extracted from a plurality of reference images (past images) used for creating the estimation model M, which are a plurality of images (burning images) taken in the past, by the image feature extraction method described above. Learning data in which one or more feature quantities F are associated with reference state quantity S corresponding to the fuel combustion state shown in the plurality of reference images may be created by machine learning. The reference state quantity is obtained by measuring the unburned content in the ash from samples sampled in synchronism with the imaging timing of a plurality of past images, or by measuring NOx concentration, CO concentration, etc. with a sensor. In addition, machine learning uses one of well-known machine learning techniques (algorithms) such as neural networks and generalized linear models.

In this way, the estimated model M created using the feature quantity F extracted from a plurality of reference images (past images) by the image feature extraction method described above and the measured value of the state quantity S when the reference image was captured. , and one or more feature amounts F extracted from a plurality of target images P to be estimated by the above-described image feature extraction method, thereby estimating the state amount S with high accuracy. can be done.

Next, the image feature extraction device 1 that executes the image feature extraction method described above will be described with reference to FIGS. FIG. 7 is a diagram schematically showing the configuration of the combustion furnace 7 in which the image feature extraction device 1 according to one embodiment of the invention is installed. FIG. 8 is a block diagram showing functions of the image feature extraction device 1 according to one embodiment of the present invention, and executes processing corresponding to FIG. 9 is a block diagram showing the functions of the image feature extraction device 1 according to one embodiment of the present invention, which executes processing corresponding to FIG.

The combustion furnace 7 of the embodiment shown in FIG. 7 is a stoker-type waste incinerator that uses municipal waste or industrial waste as fuel Fg. image extraction. Hereinafter, the present embodiment will be described as an example of a waste incinerator 7 having a combustion space 8 (combustion chamber) for burning waste, but the present invention is not limited to the present embodiment as described above. It may be a combustion furnace 7 having a combustion chamber for burning fuel Fg, such as a gasification furnace for gasifying coal in IGCC. A waste incinerator, a boiler, a gas furnace, etc. are common in that they have an upper portion 7c and a side wall portion 7s, which will be described later. Boiler and gasification furnace shall be read as appropriate.

Referring to the refuse incinerator shown in FIG. 7, in the combustion furnace 7, fuel Fg is pushed into the furnace from a fuel supply port 71 by a fuel push-in device 72, and then on a fire grate 73 (stoker) in the combustion space 8. Drying, combustion, and pre-combustion result in ash (combustion ash), which is discharged from the furnace through an ash discharge port 74 . In addition, the combustion furnace 7 includes a primary combustion space 81 consisting of a main combustion space 81a in which the fuel Fg flames and burns vigorously on a grate 73 and a secondary combustion space 81b in which the fuel Fg burns, and a grate 73. and a secondary combustion space 82 for burning unburned fuel above. Combustion gas A for fuel Fg is supplied into the furnace through a gas supply pipe 77, and is supplied from a gas supply device 76 such as a blower through a first gas flow control valve 77a to the primary combustion space from the lower part of the grate 73. 81, or supplied from the side of the combustion furnace 7 to the secondary combustion space 82 from the gas supply device 76 via the second gas flow control valve 77b. A typical example of the combustion gas A is air, but any combustible gas may be used. may be generated. On the other hand, the exhaust gas E generated by burning the fuel Fg is discharged from the stack 93 through the exhaust gas passage 91 and the exhaust gas treatment device 92 .

Further, as shown in FIG. 7, the combustion furnace 7 is provided with an imaging device 6 for imaging the inside of the furnace. The imaging device 6 is, for example, an in-core camera capable of capturing at least one of a moving image and a still image. More specifically, the imaging device 6 may be, for example, a camera such as a digital camera, a video camera, or an infrared camera capable of detecting specific wavelengths. In the embodiment shown in FIG. 7, the imaging device 6 is installed in the upper part 7c of the combustion furnace 7 located above the primary combustion space 81 (right above in FIG. 7), so as to image the combustion state from directly above. is configured to However, the present invention is not limited to this embodiment, and in some other embodiments, the imaging device 6 may be installed at a position other than the upper portion 7c of the combustion furnace 7, such as the side wall portion 7s of the combustion furnace 7. . For example, the imaging device 6 is installed in the portion forming the secondary combustion space 82 in the side wall portion 7s and the portion further above it (the portion closer to the upper portion 7c) so as to image the lower side where the primary combustion space 81 is located. may be installed obliquely downward.

The target image P (combustion image) in the furnace during combustion of the fuel Fg captured by the imaging device 6 is stored (accumulated) in a storage device connected to the imaging device 6 . In the embodiment shown in FIG. 7, it is configured to be stored in the storage device 1m of the image feature extraction device 1, but in some other embodiments, for example, It may be stored in a storage device.

As shown in FIGS. 8 and 9, the image feature extraction device 1 includes a preprocessing section 2 and a feature amount calculation section 3. FIG. Each of the above functional units will be described.
The image feature extraction device 1 (similar to the state quantity estimation device 5 described later) is composed of a computer, and includes a CPU (processor) (not shown), a memory such as ROM and RAM, and a storage device 1m such as an external storage device. It has Then, the CPU operates (calculates data, etc.) in accordance with the instructions of the program (combustion image processing program) loaded in the memory (main storage device), thereby realizing each of the functional units provided in the image feature extraction device 1. . In other words, the above program is software for causing a computer to implement each functional unit described later, and may be stored in a computer-readable storage medium.

Based on the luminance distribution H of each of the plurality of target images P described above, the preprocessing unit 2 enhances the contrast by amplifying the difference between the luminance values in the luminance distribution H ( A function that obtains at least one preprocessed image G (described above) through light-dark difference enhancement). The details of the processing performed by the preprocessing unit 2 are the same as those performed in the already described preprocessing step (S2), so the details will be omitted. Luminance distribution acquisition unit 21, threshold value setting unit 22, and contrast enhancement, which are functional units that execute the same contents as those performed in each of the distribution acquisition step (S21), threshold setting step (S22), and contrast enhancement step (S23). A portion 23 may be provided. In addition to these functional units (21 to 23), the preprocessing unit 2, as shown in FIG. 8 (not shown in FIG. 9, but similar), has the differential image generation step (S24) described above. A differential image generation unit 24, which is a functional unit that executes the same content as the content performed in 1, may be further provided.

The feature amount calculation unit 3 is a functional unit that calculates, from at least one preprocessed image G, at least one feature amount F obtained by quantifying image features of temporal and spatial changes in the flow field. The details of the processing performed by the feature quantity calculation unit 3 are the same as those performed in the feature quantity calculation step (S3) already described, and thus the details thereof are omitted.

In the embodiment shown in FIGS. 8 and 9, the image feature extraction device 1 further includes a target image acquisition unit 12, which is a functional unit that executes the same content as the target image acquisition step (S1) already described. A plurality of target images P acquired by the target image acquisition unit 12 are input to the preprocessing unit 2 . The preprocessing unit 2 is also connected to the feature amount calculation unit 3 , and is configured such that the preprocessed image G obtained by the preprocessing unit 2 is input to the feature amount calculation unit 3 .

Note that, in some embodiments, as shown in FIGS. 8 and 9, the image feature extraction device 1 may further include a grayscaling unit 14 that grayscales the input image. In the embodiment shown in FIGS. 8 and 9, the target image acquisition unit 12 and the preprocessing unit 2 are connected via the grayscaling unit 14, and multiple target images acquired by the target image acquisition unit 12 are After P is grayscaled by the grayscaling unit 14 , it is configured to be input to the preprocessing unit 2 . Therefore, when the target image acquisition unit 12 acquires a plurality of target images P of color images, the preprocessing unit 2 supplies a plurality of target images P (black and white images) converted to grayscale by the processing by the grayscaling unit 14 . image) will be input. In some other embodiments, the grayscaling unit 14 may acquire the captured image Pa and convert it into a grayscale, and the grayscaled captured image Pa may be input to the target image acquiring unit 12. . That is, the grayscaling unit 14, the target image acquiring unit 12, and the preprocessing unit 2 are connected in a row.

Further, in some embodiments, as shown in FIG. 8, the image feature extraction apparatus 1 includes a feature amount selection function which is a functional unit that executes the same content as the feature amount selection step (S40) described above. You may have the part 40. FIG.

Further, in some embodiments, as shown in FIG. 9, the image feature extraction device 1 includes an analysis unit 4, which is a functional unit that executes the same content as the analysis step (S4) already described. You may have more. Note that the analysis unit 4 may have an averaging processing unit 41 that is a functional unit that executes the same content as that performed in the averaging processing step (S41) already described. Alternatively, in addition to the averaging processing unit 41, the analysis unit 4 includes an averaged feature value selection unit 42 which is a functional unit that executes the same content as that performed in the averaged feature value selection step (S42) already described. may further have

The present invention is not limited to the above-described embodiments, and includes modifications of the above-described embodiments and modes in which these modes are combined as appropriate.

1 image feature extraction device 1m storage device 12 target image acquisition unit 14 grayscaling unit 2 preprocessing unit 21 luminance distribution acquisition unit 22 threshold setting unit 23 contrast enhancement unit 24 difference image generation unit 3 feature amount calculation unit 40 feature amount selection unit 4 analysis unit 41 averaging processing unit 42 average feature amount selection unit 5 state quantity estimation device 5m storage device 51 feature quantity acquisition unit 52 estimation model acquisition unit 53 state quantity estimation unit 55 display device 6 imaging device 7 combustion furnace 7c upper part 7s Side wall 71 Fuel supply port 72 Fuel pushing device 73 Fire grate 74 Ash discharge port 76 Gas supply device 77 Gas supply pipe 77a First gas flow control valve 77b Second gas flow control valve 8 Combustion space 81 Primary combustion space 81a Main combustion space 81b Separate combustion space 82 Secondary combustion space 91 Exhaust gas passage 92 Exhaust gas treatment device 93 Chimney H Luminance distribution L Threshold G Preprocessed image F Feature quantity P Target image R Unit region Rh High luminance region Rl Dark region M Estimation model S State quantity A Combustion gas E Exhaust gas Fg Fuel

Claims (11)

A method for image feature extraction performed by a processor, comprising:
For each of a plurality of target images to be processed obtained by imaging a flow field, based on the brightness distribution of the target image, the contrast is obtained by amplifying the difference between the brightness values in the brightness distribution. a preprocessing step of obtaining at least one preprocessed image through enhancement of
a feature amount calculation step of calculating at least one feature amount obtained by quantifying image features of temporal and spatial changes in the flow field from the at least one preprocessed image;
The image features include multiple types of the image features,
The image features obtained by preparing a plurality of sets of the plurality of target images including the target images different from each other, and performing the preprocessing step and the feature amount calculation step on each of the plurality of sets. An averaging processing step of performing averaging processing for each set on the set of feature amounts for each type of
The pretreatment step includes
a threshold setting step of setting a threshold of the luminance value corresponding to each of the plurality of target images based on the luminance distribution of each of the plurality of target images;
a contrast enhancement step of performing the contrast enhancement using the brightness value threshold corresponding to each of the plurality of target images, for each of the plurality of target images;
The contrast enhancement step includes:
changing, for each of the plurality of target images, a luminance value having a luminance value smaller than the threshold among the plurality of luminance values forming the luminance distribution to a predetermined value smaller than the threshold; An image feature extraction method, characterized in that, among a plurality of luminance values, luminance values having a luminance value equal to or greater than the threshold value are not changed . 2. The image feature extraction method according to claim 1, wherein the target image is an image that constitutes a partial area of a captured image that is an image of the flow field. 3. The image feature extraction method according to claim 2, wherein the partial area of the captured image includes an area positioned at the center of the captured image. The image feature extraction method according to any one of claims 1 to 3, wherein the luminance distribution is a distribution of the luminance values of the target image in grayscale. The pretreatment step includes
further comprising, after the contrast enhancement step, a differential image generating step of generating a differential image of the two target images on which the contrast enhancement has been performed;
2. The image feature extraction method according to claim 1 , wherein the differential image generated in the differential image generating step is obtained as the preprocessed image. The feature amount calculation step includes:
6. The method according to any one of claims 1 to 5 , wherein the at least one feature amount is calculated from the at least one preprocessed image by performing processing by a stereoscopic high-order local autocorrelation feature method. image feature extraction method. 7. The method according to any one of claims 1 to 6 , further comprising a feature amount selection step of selecting said at least one feature amount from said plurality of feature amounts obtained by said feature amount calculation step. image feature extraction method. 2. The image feature extraction method according to claim 1, wherein the averaging process is a process of calculating a moving average for each group or a process of calculating a section average for each group. The feature quantity is for use in estimating the state quantity of the flow state indicated by the plurality of target images,
9. The method according to claim 1 or 8 , further comprising an averaged feature amount selection step of selecting a feature amount having a high degree of contribution to the estimation of the state amount from among the plurality of averaged feature amounts. The image feature extraction method described. 10. The plurality of target images constitute at least part of captured images obtained by capturing the same range in the incinerator at different times during fuel combustion. The image feature extraction method according to the paragraph. An image feature extraction device for image feature extraction,
For each of a plurality of target images to be processed obtained by imaging a flow field, based on the brightness distribution of the target image, the contrast is obtained by amplifying the difference between the brightness values in the brightness distribution. a preprocessing unit that obtains at least one preprocessed image through enhancement of
a feature amount calculation unit that calculates at least one feature amount obtained by quantifying image features of temporal and spatial changes in the flow field from the at least one preprocessed image;
The image features include multiple types of the image features,
preparing a plurality of sets of the plurality of target images including the target images different from each other, and for each of the plurality of sets, each type of the image feature obtained by the preprocessing unit and the feature amount calculation unit; An averaging processing unit that performs an averaging process for each set of the feature amounts of
The pretreatment unit is
a threshold value setting unit that sets a threshold value of the luminance value corresponding to each of the plurality of target images based on the luminance distribution of each of the plurality of target images;
a contrast enhancement unit that performs the contrast enhancement using the brightness value threshold corresponding to each of the plurality of target images, for each of the plurality of target images;
The contrast enhancement unit
changing, for each of the plurality of target images, a luminance value having a luminance value smaller than the threshold among the plurality of luminance values forming the luminance distribution to a predetermined value smaller than the threshold; An image feature extracting apparatus characterized in that, among a plurality of luminance values, a luminance value having a luminance value equal to or greater than the threshold value is not changed .

Images