JP2008076167A - Method of inspecting surface state of object by image processing, and image processing program for it - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately identify the surface state of an object by performing image processing of an image obtained by imaging the surface of the object, and to inspect the surface state of the object without requiring experience, knowledge, or a special device. <P>SOLUTION: The image that is obtained by imaging the surface of the object and has a score, is wavelet-converted horizontally and vertically, and a 2k-dimension feature vector E is determined for each of a plurality of images whose components are values for each resolution m of the ratio E<SP>n</SP><SB>m</SB>of frequency energy Em of the whole image in resolution (m) to the frequency energy sum of the total resolution whose maximal solution is (k), and the ratio E<SP>l</SP><SB>m</SB>of the frequency energy in the inclined direction in the resolution (m) to the frequency energy sum of the total resolution. The determined feature vector is classified according to the score by a support vector machine, and time-by-time model is constructed. The score of an image whose score is unknown is identified with the feature vector E using the constructed identification model. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a method for inspecting the surface of a material and a program therefor, and more particularly to a method for inspecting the surface of a material using image processing effective in evaluating the surface state of the material such as the rust appearance of weathering steel. And an image processing program therefor.

  Weathering steel has the characteristics that rust (protective rust) is formed on the surface of the rust (protective rust) and the rust is stabilized and the corrosion rate is remarkably slowed by repeated repeated drying and drying in the atmosphere. Have. For this reason, if appropriate usage and maintenance are applied, structures such as bridges can be used semi-permanently without painting, and LCC (Life Cycle Cost: total including initial construction costs, maintenance costs, and renewal costs) It is an excellent steel material in terms of cost. The application of weather-resistant steel materials that do not require repainting to steel bridges is increasing year by year. In 2002, it exceeded 100,000 tons and accounted for 15% of all steel bridges. This is advocated by the Ministry of Construction's Civil Engineering Research Institute, and can be said to correspond to the minimum maintenance bridge concept of minimizing LCC in the future without assuming replacement.

  Various studies have been conducted so far on the evaluation of the degree of rust stabilization of weathering steel, and various methods such as appearance inspection, rust thickness measurement, transparent tape test, ion permeation resistance method, etc. have been proposed. Of these methods, methods other than appearance inspection require testing and measurement in direct contact with the target site, and thus a complicated operation such as installing a scaffold is indispensable. When direct contact is impossible or extremely difficult, it is necessary to evaluate only by appearance inspection. However, the appearance inspection is a method of “observing well by visual inspection” and discriminating with a reference photograph showing a standard of rust evaluation, which is inferior in terms of quantitativeness and objectivity, and has a large evaluation variation by the inspector.

  It describes in the following literature about expressing the rust stabilization degree of a weathering steel material.

Non-Patent Document 1 describes an evaluation method that focuses on environmental barrier properties that measure resistance to chloride ions in a rust layer, and Non-Patent Document 2 measures the potential of a steel material and performs electrochemical measurements. Describes a method for evaluating the anticorrosion performance of a rust layer from various characteristics. These are techniques derived from rust chemical reaction theory and have high reliability, but they require knowledge, experience and special equipment for measurement, and have not been used in general daily inspections. Non-Patent Document 3 describes the quantitative evaluation of the degree of coating deterioration of a steel bridge using image processing technology, but the image pattern has irregularities such as rust appearance. It is not suitable for accurately inspecting the surface condition of an object.
Hiroshi Kihira "Evaluation and diagnosis of stable rust formation on weathering steel" (Materials and Environment, vol.48, pp.697-700, 1999) Kazuyuki Kashima et al. “Stability evaluation method of weathering rust layer and its application to actual structures” (Materials and Environment '99, A-107, pp. 25-28, 1999) Hiroshi Fujiwara et al. “Study on Degradation Method of Steel Bridge Coating Film by Image Processing” (Journal of Civil Engineers No. 598 / I-44, pp. 85-96, 1998)

  Many conventional techniques for evaluating the degree of rust stabilization of weathering steel require experience and knowledge, and special equipment is indispensable. A technique using image processing for inspecting the surface state of an object has not actually been used. Therefore, there has been a demand for an effective method for inspecting the surface state of an object by image processing without requiring experience, knowledge, or a special device.

The present invention has been made to solve the above-described problems, and the method for inspecting the surface state of an object according to the present invention is an image of an image obtained by imaging the surface state of an object whose surface state is classified by a score. A method for inspecting a surface state of an object by identifying a score representing a surface state by performing processing, wherein a wavelet is obtained in a horizontal direction and a vertical direction with respect to an image provided with a score obtained by imaging the surface of the object. It performs conversion ratio in an oblique direction of energy to frequency energy sum of the horizontal and vertical directions in the ratio E ⁿ _m and resolution m of frequency energy Em of the entire image at a resolution m for frequency energy sum of all resolutions up to resolution as k A wavelet analysis for obtaining a 2k-dimensional feature vector E whose components are values for each resolution _m with E ^l _m is a plurality of images. What to do for each and classifying the 2k-dimensional feature vector E corresponding to each of the plurality of images according to the score by a support vector machine that classifies the multi-dimensional vectors into two classes by hyperplane Constructing an identification model for identifying a score of an unknown image by its feature vector E, identifying the score from the feature vector E for an image whose score is unknown by the constructed identification model, It consists of You may make it the surface state of the said object be a rust state of a weather-resistant steel material.

Further, an image processing program for executing on a computer to inspect the surface state of an object according to the present invention is provided for each of a plurality of images obtained by imaging the surface of an object and assigned with a score. performs wavelet transform and vertical directions, with respect to the frequency energy sum of the horizontal and vertical directions in the ratio E ⁿ _m and resolution m of frequency energy Em of the entire image at a resolution m for frequency energy sum of all resolutions up to resolution as k A support vector machine that obtains a 2k-dimensional feature vector E whose components are values of resolution m for each of the energy ratios E ¹ _{m in the} oblique direction and classifies a plurality of multi-dimensional feature vectors into two classes by a hyperplane. 2k-dimensional feature vector E corresponding to each of the plurality of images according to the function according to the score An identification model for classifying and identifying a score of an image whose score is unknown by its feature vector E is constructed, and the score is identified from the feature vector E for an image whose default score is unknown by the constructed discrimination model. Thus, the surface state of the object is inspected by identifying the score representing the surface state by performing image processing on an image obtained by capturing the surface state of the object with the surface state classified by the score.

  The present invention performs wavelet analysis on an image obtained by imaging the surface of an object, and constructs an identification model by learning using a support vector machine, so that the surface state of the object is obtained from a feature vector corresponding to the image. Can be easily and accurately identified.

The inspection of the surface state of the object according to the present invention will be described by taking as an example the case of the rust appearance of a weather resistant steel material as the surface state of the object. In this example, image processing is performed on an image obtained by imaging the surface of a weathering steel material with a digital camera, and the rust appearance is appropriately evaluated based on the image processing. This image processing is performed on the surface of the weathering steel material. Wavelet transform is performed on the image obtained by imaging the image, and the score of the rust appearance is identified from the frequency characteristics obtained by the wavelet transform. First, multiresolution analysis using wavelet transform will be described.
[Multi-resolution analysis using wavelet transform]
The wavelet is expressed by Ψ (t) defined as a wave with a mean value of 0 as shown in FIG. An inner product of a base ψ _ab (t) obtained by shifting or enlarging Ψ (t) on the t-axis and an arbitrary signal f (t) is called wavelet transform, and is defined by Expression (1).

ここで、スケールａは拡大縮小を決定する正の実数であり、シフトｂは時間（空間）方向への移動量を決定する実数である。Ψ_ａｂ（ｔ）に関して、

Here, the scale a is a positive real number that determines enlargement / reduction, and the shift b is a real number that determines the amount of movement in the time (space) direction. For Ψ _ab (t)

による直交性の関係があり、また、Ψ_ａｂ（ｔ）は

And Ψ _ab (t) is

のように規格化される。Ψ^＊（ｔ）はΨ（ｔ）の複素共役を示す。Ｗ（ａ，ｂ）はウェーブレット展開係数であり、信号ｆ（ｔ）の中にΨ_ａｂ（ｔ）の成分がどれだけ含まれるかを表すものであることから、ｆ（ｔ）は

It is standardized as follows. Ψ ^* (t) represents a complex conjugate of Ψ (t). W (a, b) is a wavelet expansion coefficient and represents how much component of Ψ _ab (t) is included in the signal f (t), so f (t) is

のように成分分解したものとして表される。ウェーブレット変換は、基底が時間軸を移動し、かつ対象とする任意の信号に対して基底の時間幅を拡大縮小することができるため、局所的な周波数情報が得られ、効率的な時間−周波数解析を行うことができるものである。

It is expressed as a component decomposed as follows. In the wavelet transform, since the base moves on the time axis and the time width of the base can be enlarged / reduced with respect to an arbitrary signal of interest, local frequency information can be obtained and efficient time-frequency It can be analyzed.

  An appropriate orthogonal basis function is a wavelet, and a and b in Equation (1) are discretized as follows.

a = 2 ^−j , b = k · 2 ^−j
(Where j = 0, ± 1, ± 2,...; K = 0, ± 1, ± 2,...)
If the variable t is replaced with an integer value n, the discrete wavelet transform for f [n] for a discrete signal composed of N sampling data is

で示される。

Indicated by

  Discrete wavelet transform is one of multi-resolution analysis. When the original signal is approximated by a linear combination of orthogonal scaling functions, the difference between the approximate function of resolution m and the approximate function of m + 1 coarser by one level is the wavelet. Is the primary combination. That is, the wavelet transform is equivalent to subband decomposition in which a low-frequency signal is divided into two successively while reducing the resolution of the original signal, and a high-frequency signal is output for each resolution.

  FIG. 2 shows a situation where the original image represented by the function f (t) is sequentially divided into two and the high frequency side is output. When the original function f (t) is sequentially decomposed with frequency components, the component image of the highest frequency component is extracted from the original image and is designated as H1, and the remaining component image obtained by subtracting the component image H1 from the original image is denoted as L1. To do. Next, iteratively repeats such that the component image H2 of the next highest frequency component is extracted from the component image L1, and the component image obtained by subtracting H2 from L1 is L2. At each stage of decomposition by frequency components, the resolutions of the component images correspond to each other.

  In such a process, the frequency component of the function f (t) as the one-dimensional information of the original image is decomposed. However, since the image has a two-dimensional variable distribution, the horizontal and vertical variables f ( t) and g (t) are decomposed by frequency components. In this case, first, for example, the horizontal one-dimensional wavelet transform is performed on the image signals of each row, and the image is divided into a high-frequency component image H and a low-frequency component image L. Next, the wavelet transform in the vertical direction is performed on the signal of each column of the component image H divided by the frequency component to divide it into component images of HH and HL. Then, wavelet transform is performed to divide into LH and LL component images. Four frequency components (LL, LH, HL, HH) are calculated by the horizontal division and the vertical division. Each frequency component is a low-frequency component in which LL includes basic information of the original image, LH and HL are enhanced in horizontal and vertical edges, and HH is enhanced in diagonal edges. It has. In the next stage, among the frequency components decomposed into four, the LL component at that stage is further decomposed into four, and the process of decomposing the low frequency side into four in the same manner is repeated. Of the four component images decomposed in the horizontal direction and the vertical direction in this way, the LL component image is further decomposed in the horizontal direction and the vertical direction, and the same is repeated thereafter (FIG. 3).

Assuming that the original image is 2 ¹⁰ × 2 ¹⁰ pixels, the process of performing the decomposition by the frequency component as shown in FIG. 2 in the horizontal direction and the vertical direction, and the information amount of the current image included in one pixel of the component image at each stage Is shown in Table 1.

ウェーブレット展開係数をＷ（ｘ，ｙ）とすると、Ｍ×Ｎの画素からなる画像の解像度ｍにおけるＬＨ成分の周波数エネルギーＥ_ＬＨｍは

When the wavelet expansion coefficient is W (x, y), the frequency energy E _LHm of the LH component at the resolution m of the image composed of M × N pixels is

で、また、画像全体の周波数エネルギーＥ_ｍは

In, also, the frequency energy E _m of the entire image

で算出される。このとき最大解像度をｋとすると、全解像度の周波数エネルギー和に対するＥ_ｍの比Ｅ^ｎ _ｍ（ｎはｎｏｒｍａｌを表す）は

Is calculated by At this time, ^assuming that the maximum resolution is k, the ratio E ⁿ _m (n represents normal) of E _m to the frequency energy sum of all resolutions is

で算出され、画像の周波数エネルギーの分布特性を評価することができる。

The frequency energy distribution characteristics of the image can be evaluated.

The CA corresponding to the center of gravity of the distribution map of E ⁿ _m is

で算出される。Ｅ^ｎ _ｍ、ＣＡは解像度間のエネルギー比であり、複数の画像を評価する場合、画像の明るさや全体的な色合いの違いに影響を受けにくいパラメータである。
〔さび画像の評定〕
画像処理の目標として、耐候性鋼材のさび外観を適切に評価することがある。さび画像によりさび外観を判別するのには個人差が伴うが、熟練技術者は的確に判別し得る。外観耐候性鋼材のさび外観については表２のような外観評点の基準がある。

Is calculated by E ⁿ _m and CA are energy ratios between resolutions, and are parameters that are not easily affected by differences in image brightness or overall hue when evaluating a plurality of images.
[Rust image rating]
A goal of image processing is to appropriately evaluate the rust appearance of weathering steel. Discrimination of the rust appearance from the rust image involves individual differences, but a skilled engineer can accurately discriminate. Appearance Weather-resistant steel materials have rust appearance standards as shown in Table 2.

ウェーブレット変換による多重解像度解析を用いた画像処理の目標としては、このようなさび画像を画像処理して得られた結果により表２に示されるような外観評点と同程度にさび外観を判別できるようにすることがある。目視外観基準にあるさびの平均外観粒径は同一スケールのさび画像におけるさびの粗密さ、周波数として表現できると考えられる。表２における評点１（異常）や評点２（要観察）のさび外観はさび粒子とは別に層状ないしうろこ状の剥離形状をもち、これらのさび画像は低周波領域のエネルギーが高くなると考えられる。

The goal of image processing using multi-resolution analysis by wavelet transform is to be able to discriminate the rust appearance to the same extent as the appearance score as shown in Table 2 based on the results obtained by image processing such a rust image. It may be. It is considered that the average appearance particle diameter of rust in the visual appearance standard can be expressed as rust roughness and frequency in rust images of the same scale. The rust appearance of grade 1 (abnormal) and grade 2 (needed to be observed) in Table 2 has a layered or scaly peeled shape separately from the rust particles, and these rust images are considered to have high energy in the low frequency region.

By calculating the mean value and standard deviation of the distribution center of gravity CA of E ⁿ _m described above for the rust image for each maximum resolution, and by linear approximation, it can be seen that E ⁿ _m and CA can roughly express the roughness of the rust quantitatively. . E ⁿ _m and CA is a characteristic quantity that focuses on energy of the entire image at each resolution. Therefore, it is suitable for the evaluation of rust images with a rating of 3 to 5 in Table 2, but is appropriate for the evaluation of layered or scaly rust like the rust images with a rating of 1 and 2, which are not uniform. Absent. For this reason, it is preferable to use a pattern recognition technique in order to increase the accuracy.
[Distinction using pattern recognition technique]
For non-uniform images such as delamination rust or scaly rust such as grades 1 and 2, an index that expresses local distribution characteristics is required, and a new feature value for each resolution is required.

Ｅ^ｌ _ｍ（ｌはｌｏｃａｌを表す）を用いる。Ｅ^ｌ _ｍは水平方向及び鉛直方向の周波数エネルギー和に対する斜め方向のエネルギーの比であり、さび画像のテクスチャとしての不均一性に着目した特徴量である。式（８）及び式（１０）から、さび画像の特徴ベクトルＥを

E ^l _m (l represents local) is used. E ^l _m is the ratio of energy in an oblique direction with respect to the frequency energy sum of the horizontal and vertical directions, a feature quantity which focuses on non-uniformity of the texture of rust image. From equation (8) and equation (10), the feature vector E of the rust image is

で定義する。このような特徴量と外観評点との相関をよくするために、非線形性を考慮したパターン認識の手法を取り入れる。この手法について、
・特徴ベクトルＥを学習パターンとした外観評点の識別モデルをサポートベクトルマシンにより構築する。
・構築された識別モデルによりさび画像の識別を行い、その結果により識別モデルを用いたさび画像の識別の有効性について検証する。
という段階に分けて説明する。
（ａ）サポートベクトルマシンによる識別モデルの構築
サポートベクトルマシン（Ｓｕｐｐｏｒｔ Ｖｅｃｔｏｒ Ｍａｃｈｉｎｅ；以下、ＳＶＭ）は、２クラスの分類問題を解くために作られた学習機械である。式（１１）における特徴ベクトルＥは２ｋ次元の特徴量空間に分布するサポートベクトルとして示されるが、このような２クラスのパターン分布には、超平面によって２つに分けられる線形分離可能な場合と、分けることができない非線形な場合とが存在する。ＳＶＭはこの両方に適用可能で、マージン最大化と呼ばれる手法に基づいた汎化能力の高いものと言える。

Define in. In order to improve the correlation between the feature quantity and the appearance score, a pattern recognition method taking into account nonlinearity is adopted. About this technique,
-An appearance score identification model using the feature vector E as a learning pattern is constructed by a support vector machine.
・ Identify the rust image using the constructed identification model, and verify the effectiveness of identifying the rust image using the identification model.
This will be described in stages.
(A) Construction of Identification Model by Support Vector Machine A support vector machine (hereinafter referred to as “SVM”) is a learning machine created to solve a two-class classification problem. The feature vector E in the equation (11) is shown as a support vector distributed in a 2k-dimensional feature amount space. However, in such a two-class pattern distribution, a linearly separable case divided into two by a hyperplane is possible. There are non-linear cases that cannot be separated. SVM can be applied to both, and can be said to have a high generalization ability based on a technique called margin maximization.

The learning function X = (x ₁ ,..., X _d ) ^{t in} the general linear classifier has a learning function X of dimension d,

のように表される。ここで、ｗ^ｉは識別器の重みを表し、ｂはバイアス項と呼ばれるパラメータである。この識別器のＰ（Ｘ）＝０を満たす点の集合（識別面）は、（ｄ−１）次元の超平面となる。ここで、図４（ａ）のように異なるクラスの学習パターンが（ｄ−１）次元の超平面（図では直線）によって分離できたとすると、学習パターンを完全に識別する超平面は無数に存在する。識別の最終目標は、無数にある未知のテストパターンを正確に識別することであり、ＳＶＭでは２つのクラスの中央部を通る（マージンを最大にする）超平面が最も識別性に優れているものとして、図４（ｂ）のように、これを探索する。

It is expressed as Here, w ⁱ represents the weight of the discriminator, and b is a parameter called a bias term. A set of points satisfying P (X) = 0 (discrimination plane) of this discriminator becomes a (d-1) -dimensional hyperplane. Here, if learning patterns of different classes can be separated by (d-1) -dimensional hyperplanes (straight lines in the figure) as shown in FIG. 4A, there are innumerable hyperplanes that completely identify the learning patterns. To do. The ultimate goal of identification is to accurately identify a myriad of unknown test patterns. In SVM, the hyperplane that passes through the middle of the two classes (maximizing the margin) has the best discrimination. As shown in FIG. 4B, this is searched.

In the case of a feature space incapable of linear separation, a discriminant function such as Expression (12) cannot obtain a sufficient discriminating performance. Therefore, the feature vector XεR ^d is mapped to a higher-dimensional space R ^q using a nonlinear function Φ (X). Thus, if linear identification is performed in the q-dimensional Euclidean space of the mapping destination, it is substantially equivalent to performing nonlinear identification in the original space (FIG. 5).

Using any number of nonlinear functions φ _j (X) (j = 1,..., H) that output scalars, the function Φ (X) is

のように定義する。Φ（Ｘ）を新たなパターンとみなせば、線形ＳＶＭと同様に最適な識別関数が定式化されるが、Φ（Ｘ）を含む非線形関数は高次元のベクトル演算を含み、膨大な計算量を必要とする。そのため、もとの空間で定義されるカーネル関数Ｋ（Ｘ，Ｙ）を、

Define as follows. If Φ (X) is regarded as a new pattern, an optimal discriminant function is formulated as in linear SVM, but a nonlinear function including Φ (X) includes high-dimensional vector operations and requires a large amount of calculation. I need. Therefore, the kernel function K (X, Y) defined in the original space is

として導入する。カーネル関数は、もとの空間における学習パターンＸ及びＹの位置関係を写像先で内積として反映し、識別に重要な役割をもつもので、例えば多項式型やガウシアン型のカーネル関数がある。簡易に高次元への写像を可能にするものとして、

Introduce as. The kernel function reflects the positional relationship between the learning patterns X and Y in the original space as an inner product at the mapping destination and plays an important role in identification. For example, there are polynomial type and Gaussian type kernel functions. As something that allows easy mapping to higher dimensions,

の形のＲＢＦ関数がある。ここで、γ（γ＞０）はカーネルパラメータである。

There is an RBF function of the form Here, γ (γ> 0) is a kernel parameter.

  In this way, by learning using the SVM, a large number of rust images each given a score according to the appearance score criteria shown in Table 2 for the case of a rust image are represented by the feature vector E corresponding to each rust image. Build an identification model to be able to identify which grade it belongs to. For this purpose, as shown in FIGS. 4A and 4B, feature models E whose scores are known are classified using SVM to create an identification model that can be classified into five scores. Here, since the SVM is a learning machine that solves the two-class classification problem, the image of the target score and the other images are classified one-by-one for each of the appearance scores of all classes (two classes). (Repeat classification)

By using the identification model constructed through the learning process using SVM in this way, it is possible to automatically identify the rust image by using the identification model and the image of which score the unidentified rust image belongs to. can do. FIG. 6 shows a flow of constructing the identification model and identifying the score of the rust image.
(B) Identification of a rust image using an identification model and effectiveness of identification of a rust image using the identification model based on the identification model As a result, the construction of an identification model for identifying the score of a rust image by a flow as shown in FIG. It is executed on the rust image, the rust image is identified using the constructed identification model, and the result is compared with the result obtained by the appearance evaluation separately to verify the effectiveness of the identification model.

The rust images used in the analysis were taken from a 6-year-old unpainted bridge with a digital camera with 3 million effective pixels.
Y = 0.299R + 0.587G + 0.114B
Pre-processing is performed in which R, G, and B are converted to a gray scale according to an expression of red, green, and blue. An identification model is constructed with 558 rust images and their appearance scores as all patterns, and a part of them as a learning pattern. Table 3 shows the appearance scores of all patterns. For the learning pattern, in order to verify the robustness of the identification model due to the restriction on the number of patterns, the ratio α of the learning pattern with respect to all the image data is changed from 10% to 50%, and the matching rates are compared. Further, in order to prevent bias in learning, identification under each condition is performed on 30 identification models based on randomly selected learning patterns, and the matching ratios are averaged and evaluated. Evaluation of the relevance ratio for this identification model is performed in the form of the flow of FIG.

When the ratio α of the learning patterns among all the patterns is changed in the range of 10% to 50%, the change in the matching rate is as shown in FIG. 8A, and the maximum resolution (k) is 5-8. Going about the case. Further, in FIG. 8 (b), the maximum resolution (k) is 8, shows a feature vector _{^{(E, E n m, E}} l m) changes in the adaptation rate when changing the ratio α of the learning pattern Yes. From the results shown in FIGS. 8 (a) and 8 (b), it is clear that the matching rate improves as the ratio of the learning pattern increases. Since the total number of patterns is constant, if the number of learning patterns increases, the number of remaining patterns to be identified passes, but the relevance rate is improved by learning. Further, according to the result shown in FIG. 8B, the matching rate is better when E is used as a feature vector than when E ⁿ _m and E ^l _m are used alone, and the parameter is effective for identification. It turns out that it becomes.

  The number of feature vector dimensions (d) and the number of learning patterns (n) greatly affect the precision, and generally n> 2 (d + 1) is necessary to obtain a statistically reliable result in a two-class classification problem. It is considered necessary to meet. In the results shown in FIG. 8A, it can be said that the difference in the maximum resolution (k = 5, 6, 7, 8) does not significantly affect the precision. If the maximum resolution is lowered, the number of dimensions of the feature vector can be reduced. Even if the number of learning patterns is the same, the reliability is improved and the time required for wavelet transform and SVM learning is shortened. Considering that the error with respect to the frequency energy in the low frequency region is large in the wavelet transform, it can be said that 5 to 6 is reasonable for the maximum resolution.

  Table 4 shows that the maximum resolution is 6 and the ratio (α) of the learning pattern is 50%, learning is performed using the feature vector E for the learning pattern, and an identification model is constructed. % Of patterns are identified, and a learning model is selected at random, and an identification model is constructed and evaluated 30 times, and the total of the identification results obtained is shown in comparison with the appearance rating. In Table 4, it is shown that the result in the diagonal frame in which the identification result by the identification model constructed by SVM matches the appearance score is correctly identified, and the ratio is the matching rate of the numbers in parentheses. The highest relevance rate is 94% with a score of 1 and 72% of a score of 3 with the lowest, and a good relevance rate is obtained as a whole by discrimination by a learning model.

  In the description here, from the aspect of constructing an identification model and verifying its effectiveness, a part (α%) of all image patterns is used as a learning pattern, and the remaining (100 -Α)% of patterns were identified, and the effectiveness of the identification model was confirmed by the matching result that matched the identification result or the correct score. In this way, wavelet analysis on rust images and SVM were used. By building an identification model by learning, it is possible to automatically identify the score of the rust image. In order to automatically perform identification, it is necessary to construct a highly accurate identification model. To that end, it is necessary to accumulate as many and various image patterns as possible.

  Although the present invention has been described with respect to an example of evaluating a rust appearance based on a rust image on the surface of a weather resistant steel material, a wavelet analysis here, construction of an identification model by learning using SVM, and automatic identification by that The method of performing is not particularly limited to the evaluation of the rust appearance of weathering steel, and even when evaluating the surface condition of general materials or the image of the ground / vegetation condition etc. Applicable as the state is classified by score.

It is a figure which shows the example of the waveform of a wavelet. It is a figure explaining sequential decomposition | disassembly by a frequency component, reducing the resolution of an image by wavelet transformation. It is a figure which shows roughly the process which performs the wavelet transformation of a horizontal direction and a vertical direction with respect to an image, and is decomposed | disassembled. (A), (b) It is a figure which shows roughly the classification | category of the feature vector by a support vector machine. It is a figure which images the mapping to the high-dimensional space of the feature space. It is a flowchart which shows the process of construction of the identification model by this invention, and the identification using it. It is a figure which shows the process which verifies the effectiveness of the identification model constructed | assembled by this invention. It is a figure which shows the relevance rate when changing the ratio of the learning pattern with respect to all the patterns in the case of construction | assembly of an identification model, (a) changes the maximum resolution, (b) shows changing the feature vector.

Claims (3)

A method for inspecting a surface state of an object by identifying a score representing the surface state by performing image processing on an image obtained by imaging the surface state of the object whose surface state is classified by a score,
The image obtained by imaging the surface of the object is subjected to wavelet transform in the horizontal direction and the vertical direction with respect to the image, and the frequency energy Em of the entire image at the resolution m with respect to the frequency energy sum of all resolutions with the maximum resolution as k. wavelet analysis for obtaining the 2k-dimensional feature vector E to the value of each resolution m and components of the ratio E ^l _m of energy in an oblique direction with respect to the frequency energy sum of the horizontal and vertical directions in the ratio E ⁿ _m and resolution m of Doing for each of the plurality of images;
A support vector machine that classifies a plurality of multi-dimensional vectors into two classes according to a hyperplane classifies the 2k-dimensional feature vector E corresponding to each of the plurality of images according to the score, and determines the score of an image whose score is unknown. Building an identification model for identification by feature vector E;
Identifying the score from the feature vector E for an image with an unknown score by the constructed identification model;
A method for inspecting a surface state of an object characterized by comprising:   The method for inspecting the surface state of an object according to claim 1, wherein the surface state of the object is a rust state of a weather-resistant steel material. The entire image at resolution m with respect to the sum of frequency energy of all resolutions, with wavelet transform performed in the horizontal and vertical directions for each of a plurality of images obtained by imaging the surface of the object and given a score. the frequency energy Em of the ratio E ⁿ _m and 2k dimension whose value is the component for each resolution m of the ratio E ^l _m of diagonal energy to frequency energy sum of the horizontal and vertical directions at resolution m feature vectors E The score is unknown by classifying the 2k-dimensional feature vector E corresponding to each of the plurality of images according to the score by the function of the support vector machine that classifies the multi-dimensional feature vectors into two classes by the hyperplane. An identification model for identifying the score of an image by its feature vector E is constructed. A score that represents the surface state by performing image processing on an image obtained by capturing the surface state of the target object by identifying the score from the feature vector E of the image whose score is unknown by the identification model. An image processing program for executing on a computer to inspect the surface state of an object based on identification.

Images