Multi-scale feature extraction and recognition of slope damage in high fill channel based on Gabor-SVM method

Abstract

In order to make the feature data more linearly separable for support vector machine(SVM) classifier, a set of different scale and direction parameter was proposed for improving the recognition effect of cement slope damage in high fill channels. The Gabor wavelet was used to extract the multi-scale and multi-directional features of the high-fill channel’s abrupt features. Then SVM algorithm was utilized to perform damage classification and level recognition. To compare the recognition effect of the Gabor-SVM method, histogram-SVM, grayscale symbiotic matrix -SVM, canny-SVM algorithm were adopted to identify the damage degree of cement surface in the same environment, and these damage recognition rates are compared with Gabor-SVM’s. The experimental results show that the damage recognition model, based on Gabor-SVM, tends to better stable value when the wavelet takes 6 scale and 12 directions. The recognition rate of the normal slope is 0.98, while the recognition rate of the crack, hole, and broken slope are 0.63, 0.88 and 0.90, respectively. Overall, the damage recognition model, based on Gabor-SVM, has better recognition effect, and it will provide technical support for finding potential leakage hazards in the high fill channel of South-to-North Water Diversion Project.

Keywords

Key words: Middle route of South-to-North Water Diversion Project Feature extraction of multiple scales and directions damage recognition Gabor wavelet support vector machine

1 Introduction

High fill channels, which are generally considered to be more than 6 m above the ground, are widely distributed in the middle route of the South-to-North Water Transfer Project in China [1 –3]. Due to the high filling height, wide distribution range and complicated engineering geological conditions, the lining panel is cracked or the slope of the channel slope is damaged [4]. The high fill channel once leakage for a long time with bursts, which will directly affect the life property safety of the residents along the road. Therefore, high fill channel section of the seepage monitoring technology is one of the key technology in long distance water transfer project safety monitoring.

However, leakage information is difficult to be effectively perceived and detected. There are many methods for high fill channel leakage detection, including geological radar detection method, resistivity method, ultrasonic detection method and image processing detection method [5 –7]. The algorithm for recognition and control of nonlinear systems includes model of seepage flow, multi-source data fusion, multi-objective inverse modeling, adaptive fuzzy control, and so on [8 –10].

When the high fill channel is damaged, structural deformation or physical diagenesis will occur on the slope of the high fill channel, which will bring about the slope texture and corner changes [11]. Therefore, the digital image processing of machine vision can detect the slope texture and corner changes of the high fill channel section. It provides feature extraction technique support for finding structural damage detection and helps the leak detection of the high fill channel section [12].

The extraction of image features is one of the key links in the identification of cement surface damage in high fill channels. Image feature extraction methods usually include binarization, canny edge detection, and gray level co-occurrence matrix [13 –15]. Although the general image detection method can have advantages in large-area detection, it is difficult to effectively identify the characteristics of the sudden change of the leakage region.

Multi-scale feature vector reflects not only the global features of the whole image, but also the sub-bands of different characteristics such as edges and texture [16]. Therefore it has more identifying information. Gabor wavelet transform method is sensitive to the edge features of images, which can provide good direction and scale selection characteristics, and has good adaptability to illumination changes [17]. The support vector machine(SVM) classifier can obtain discriminate functions based on a limited set of data samples, and still obtain small errors for independent test sets, and has strong generalization ability [18, 19]. The SVM has been used in pipeline leak and leak detection [20].

Therefore, the study of detection model for cement surface damage, based on Gabor-SVM in high fill channel, was proposed. The Gabor wavelet was used to extract the multi-scale and multi-directional features of the high-fill channel’s abrupt features, to obtain the good features of the high-fill channel local space and frequency domain information. Then SVM algorithm was utilized to perform damage classification and level judgment, in order to obtain higher leakage level and recognition effect.

2 Materials and methods

2.1 Materials

The high fill channel model, based on the South-North Water Transfer (Open Channel), was built in the laboratory. The slope image is acquired with a floating bin loaded with a high-definition and movable camera. Due to the limitations of experimental conditions, the 520 images data were collected in the experiment, which was derived from the image of the cement slope above the water surface.

The damage level of the cement slope in the high fill channel is defined as crack, hole, broken and normal, as shown in Fig. 1. The definition of these four types of damage level is based on the features of structure or physics, which will bring about the average variance and average mean characteristics of the slope texture and corner changes.

Fig. 1

Cracks, holes, broken, normal cement surface.

The crack in the rock is a planar discontinuity formed by structural deformation. The crack has the characteristics of rapid change, and the variance and mean of the amplitude have lower characteristics. The brightness of the crack is darker than the brightness of the background of the image, and has certain linear characteristics, such as strips and sawtooth strips.

The hole refers to the damage of the hole in the cement panel. The diameter of the hole is generally 2.5mm∼10 mm, and the depth is 1cm∼5 cm. The variance and mean of the hole are higher, and the shape is blocky, round, elliptical, etc.

The broken is composed of cracks with multiple linear features. The width of the broken joint is wider than that of the crack, and the crack is divided into three or more pieces. And its variance and mean of the amplitude are high. If the cement slope has no above three kinds of damage, it is called normal cement surface (referred to as normal).

2.2 Gabor wavelet filter methods

In general, the image is affected by various noises. Therefore, it needs to be preprocessed before extracting its features. These processing methods include image enhancement, median filtering, and grayscale processing. The Gabor wavelet has good characteristics in extracting the local space and frequency domain information of the target, and the basis function of the two-dimensional Gabor wavelet can well describe the sensory characteristics of most simple visual neurons in the mammalian primary vision system [21, 22]. The expression of the two-dimensional Gabor wavelet function is as follows [23]. Then, Gabor wavelet is used to extract the texture features of the image, as well as the image convolution, commonly used amplitude, and phase to represent the texture features. $g (x, y) = [\frac{1}{2 π σ_{x} σ_{y}}] exp [\frac{1}{2} [\frac{x^{2}}{σ_{x}^{2} σ_{y}^{2}}] + 2 π jWx]$ (1)

Where W is the frequency bandwidth of the filter, which is the complex modulation frequency of the Gaussian function. σ_x and σ_y represent the standard deviation of the Gaussian function along two different coordinate axes. The Fourier transform of Equation (1) is the formula G (u, v). $G (u, v) = exp {- \frac{1}{2} [\frac{(u - W)^{2}}{σ_{u}^{2}} + \frac{v^{2}}{σ_{v}^{2}}]}$ (2) where: ${\begin{matrix} σ_{u} = \frac{1}{2 π σ_{x}} \\ σ_{v} = \frac{1}{2 π σ_{y}} \end{matrix}$ (3)

Studies by neurophysics have shown that W = π/2(Hz) conforms to the human visual system. However, the value of W in digital image processing needs to be re-adjusted to accommodate optimal machine vision processing. Taking g (x, y) as the mother wavelet, a proper self-similar set of filters, called Gabor wavelet, can be obtained by performing appropriate scale expansion and rotation transformation on g (x, y). $g_{mn} (x, y) = a^{- m} g (x^{'}, y^{'}), a > 1, m, n \in Z$ (4)

In Equation (4): ${\begin{matrix} x^{'} = a^{- m} (x cos θ + y sin θ) \\ y^{'} = a^{- m} (- x sin θ + y cos θ) \end{matrix}$ (5)

Where m, n is the scale and direction of the wavelet, m = 0, 1, . . . S–1, n = 0, 1, . . . K–1. S is the total number of scales, k is the total number of directions, a^-m is the scale parameter, θ = nπ/K is the direction parameter.

2.3 SVM classifier

SVM has many advantages in solving small sample, nonlinear and high-dimensional pattern recognition, and can be extended and applied to other machine learning problems such as function fitting [24, 25]. In addition, the SVM can find the global optimal solution to the convex quadratic programming problem. The SVM can solve the problem of linear indivisibility of data through slack variables and kernel function techniques. Suppose the collection of two types of data samples is: {(x_i, y_i) , i = 1, 2, . . . , n}, x_i ∈ R^d, y_i ∈ {-1, 1}, the general representation of linear discriminate functions is f (x) = wx + b, then the classification equation corresponding to f (x) = wx + b is: $wx + b = 0$ (6)

The classification interval here is: $d = \frac{2}{∥ w ∥}$ , which to maximize the interval between classifications. w refers to weight, ∥w∥ refers to Euclidean distance.

The Lagrange function is defined as follows: $L (w, b, α) = \frac{1}{2} {∥ w ∥}^{2} - \sum_{i = 1}^{n} α_{i} [y_{i} ({wx}_{i} + b) - 1]$ (7)

Find the minimum value of Equation (7): ${\begin{matrix} \frac{\partial L}{\partial w} = 0 \Rightarrow w = \sum_{i = 1}^{n} α_{i} y_{i} x_{i} \\ \frac{\partial L}{\partial b} = 0 \Rightarrow \sum_{i = 1}^{n} α_{i} y_{i} = 0 \\ \frac{\partial L}{\partial w} \Rightarrow α_{i} [y_{i} ({wx}_{i} + b) - 1] = 0 \end{matrix}$ (8)

Solving the above optimal classification surface problem is transformed into a dual problem for solving convex quadratic programming optimization under the constraints of Equations (7) and (8): ${\begin{matrix} max \sum_{i = 1}^{n} α_{i} - \frac{1}{2} \sum_{i = 1}^{n} \sum_{j = 1}^{n} α_{i} α_{j} y_{i} y_{j} (x_{i}, x_{j}) \\ s . t . α_{i} ⩾ 0, i = 1, . . ., n \\ \sum_{i = 1}^{n} y_{i} α_{i} = 0 \end{matrix}$ (9) Where a is a Lagrange multiplier. Equation (9) is a problem in which there is a unique solution and it is a quadratic function. If the solution obtained is the optimal solution, then there are: $w^{*} = \sum_{i = 1}^{n} α_{i}^{*} y_{i} x_{i}$ (10)

In the above formula, the sample with a not zero is the support vector. Therefore, the linear combination of support vectors is the weight coefficient vector of the optimal classification plane. To solve the classification domain value c under the constraint of b, then the optimal classification surface function for the above problem is: $g (x) = sgn {wx + b} = sgn {\sum_{i = 1}^{n} α_{i}^{*} y_{i} (x_{i} x) + b^{*}}$ (11)

2.4 Gabor-SVM fusion architecture

The Gabor-SVM fusion architecture is shown in Fig. 2. First, a convolution operation is performed on IM()m*m, which is the training sample image. And the Gabor wavelet filter G_F (x, y) is designed with parameters. Second, the convolved matrix is partitioned and squared. The multiple variances and the mean are combined into a one-dimensional array and normalized to obtain the characteristics of the magnitude. Finally, after the SVM is trained, the sample decision function Model() is obtained. Before the pattern recognition, the picture to be tested also needs to be convolved, divided, squared and averaged with the Gabor filter, then merged and normalized. The extracted variance and mean features are classified in the decision function of the SVM classifier to obtain the classification result.

Fig. 2

Gabor-SVM image processing architecture.

In Fig. 2, IM()m*m represents the training image and im()m*m represents the test image. R()m*m represents the feature set after the training image and gabor wavelet convolution operation. Correspondingly, r()m*m represents the feature set after the test image and the gabor wavelet convolution operation.MM(j) _i *i represents the set after block processing of the R()m*m feature set. And mm(j) _i *i represents the set after block processing of the r()m*m feature set.M() j*j represents the set of merged for each MM(j) _i *i. While m() j*j represents the set of merged for each MM(j) _i *i.M’() j*j represents the set of normalized for each M() j*j. Correspondingly, m’() j*j represents the set of normalized for each m() j*j. Model () is the decision function of SVM.

3 Experiments and results

3.1 Analyze the digital features

According to the processing framework of Fig. 2, Gabor filtering is performed on of slope images of the crack, hole, broken and normal, and the average variance and average mean characteristics of each scale and each direction are obtained. Table 1 shows the mean and variance of various damage level when the frequency bandwidth W is π/2. It can be seen that the mean and variance of the holes are larger, and the mean and variance of the broken are larger than the crack and normal, and the mean and variance of the normal cement face are generally the smallest. The more obvious numerical relationship is, the more favorable it is to find the linear relationship of various damage conditions. It is also seen from Table 1 that when W = π/2, the variance and mean of the four cases are not very obvious, and further adjustment is needed.

Table 1
Multi-scale and multi-directional amplitude characteristics when the frequency bandwidth is π/2

Scale\Direction 1 2 3 4 5 6 7 8 9 10 11 12 13

(a) Mean

1 Hole 11.2 9.85 8.19 8.51 8.94 9.13 9.23 9.35 9.47 9.58 9.68 9.77 9.85

Broken 15.7 15.9 13.0 13.6 13.8 14.0 14.1 14.2 14.2 14.3 14.3 14.4 14.5

Crack 8.31 8.85 9.71 9.31 9.12 9.12 9.20 9.24 9.14 9.24 9.14 9.08 9.03

Normal 11.4 11.9 11.7 11.7 11.8 11.7 11.6 11.4 11.3 11.2 11.1 11.0 11.0

2 Hole 8.73 8.24 6.61 6.95 7.37 7.65 7.74 7.79 7.85 7.92 8.00 8.06 8.12

Broken 15.4 15.9 12.0 12.0 12.5 13.3 13.7 13.8 13.9 14.0 14.0 14.1 14.2

Crack 7.34 7.67 9.77 8.35 7.69 7.78 7.96 7.99 7.94 7.86 7.79 7.74 7.69

Normal 9.32 10.1 8.65 8.93 9.14 9.24 9.24 9.20 9.18 9.17 9.15 9.16 9.17

3 Hole 7.52 6.78 5.97 5.79 5.93 6.13 6.21 6.25 6.31 6.41 6.54 6.66 6.77

Broken 14.3 15.9 11.0 10.8 10.8 11.8 12.6 13.0 13.2 13.3 13.3 13.3 13.3

Crack 6.54 7.95 10.2 8.20 7.27 7.26 7.41 7.49 7.49 7.43 7.46 7.28 7.21

Normal 7.88 8.15 6.67 7.02 7.53 7.57 7.47 7.44 7.45 7.45 7.47 7.48 7.49

4 Hole 7.58 6.26 5.29 5.35 5.63 5.79 5.85 5.93 6.04 6.19 6.37 6.53 6.69

Broken 14.0 14.7 8.91 9.93 10.4 11.7 12.4 12.5 12.6 12.6 12.6 12.7 12.7

Crack 6.15 7.53 10.2 7.97 7.66 7.66 7.63 7.66 7.66 7.63 7.58 7.51 7.44

Normal 6.51 7.02 6.07 5.34 5.15 4.29 5.35 5.38 5.43 5.48 5.54 5.59 5.65

5 Hole 7.87 7.98 5.87 6.13 5.89 5.96 6.18 6.34 6.45 6.57 6.67 6.75 6.81

Broken 13.4 15.1 8.34 8.91 9.65 10.5 11.2 11.7 11.9 12.0 12.0 12.0 12.0

Crack 6.35 8.51 10.1 6.44 5.80 6.36 6.76 6.95 7.05 7.10 7.13 7.14 7.14

Normal 5.97 6.77 4.74 4.83 4.00 4.00 4.21 4.30 4.35 4.39 4.45 4.51 4.59

6 Hole 9.00 7.31 6.00 5.91 6.18 6.34 6.62 6.81 6.88 6.90 6.93 6.98 7.07

Broken 12.6 13.7 7.67 6.73 8.55 9.69 10.1 10.3 10.4 10.6 10.8 10.9 11.1

Crack 6.28 8.81 9.99 6.95 5.42 5.23 5.70 6.16 6.42 6.50 6.51 6.48 6.46

Normal 5.40 6.81 3.35 4.14 3.91 3.52 3.36 3.29 3.23 3.20 3.20 3.23 3.29

7 Hole 12.2 11.1 8.52 8.45 8.59 8.72 9.05 9.40 9.76 10.1 10.5 10.7 10.9

Broken 11.1 13.5 6.33 6.01 7.58 9.02 9.57 9.59 9.60 9.69 9.80 9.90 10.0

Crack 7.37 9.43 8.27 5.50 4.54 4.03 4.12 4.38 4.81 5.37 5.92 6.37 6.72

Normal 6.65 7.85 3.08 3.25 3.51 3.42 3.34 3.37 3.63 4.06 4.57 5.05 5.45

(b) variance

1 Hole 6.18 5.53 4.39 4.73 5.10 5.16 5.18 5.22 5.26 5.29 5.32 5.34 5.37

Broken 11.1 11.7 7.90 8.80 9.01 8.92 8.87 8.93 9.05 9.21 9.36 9.49 9.59

Crack 5.24 6.08 7.82 6.68 6.12 6.10 6.19 6.27 6.30 6.30 6.30 6.28 6.19

Normal 6.03 6.51 6.22 6.06 6.05 6.13 6.08 6.02 5.97 5.95 5.94 5.96 5.98

2 Hole 4.80 4.69 3.78 4.11 4.22 4.31 4.37 4.43 4.43 4.41 4.39 4.38 4.38

Broken 11.0 12.2 7.34 8.24 8.04 8.47 8.85 9.17 9.39 9.52 9.60 9.65 9.70

Crack 4.87 5.67 9.14 6.49 5.50 5.53 5.69 5.82 5.87 5.88 5.86 5.83 5.79

Normal 5.21 5.44 4.52 4.59 4.84 4.95 4.93 4.86 4.80 4.77 4.78 4.80 4.82

3 Hole 4.58 4.16 3.74 3.62 3.78 3.99 4.09 4.11 4.12 4.14 4.16 4.18 4.20

Broken 9.76 12.1 6.67 7.09 7.39 8.28 8.80 9.00 7.04 9.05 9.08 9.11 9.12

Crack 4.15 5.55 9.52 6.67 5.32 5.12 5.32 5.54 5.64 5.65 5.60 5.51 5.41

Normal 4.17 4.32 3.51 3.73 3.77 3.67 3.61 3.60 3.64 3.68 3.71 3.73 3.75

4 Hole 5.56 3.75 3.28 3.40 3.48 4.45 4.58 4.72 4.88 4.02 4.16 4.29 4.41

Broken 7.92 10.7 5.64 6.08 6.85 7.73 8.09 8.07 7.94 7.84 7.81 8.81 8.83

Crack 3.84 5.29 8.37 6.15 4.91 4.91 5.26 5.52 5.70 5.76 5.74 5.67 5.58

Normal 3.27 3.48 3.00 2.41 2.46 2.51 2.53 2.55 2.57 2.62 2.68 2.73 2.77

5 Hole 4.67 4.59 3.45 3.77 3.64 3.60 3.69 3.79 3.90 3.95 3.98 4.03 4.10

Broken 6.82 9.34 4.77 4.67 5.14 6.12 6.56 6.46 6.24 6.12 6.07 6.05 6.03

Crack 3.41 5.10 6.40 4.08 3.24 3.51 4.14 4.65 4.91 5.00 4.98 4.91 4.81

Normal 3.30 3.47 2.12 1.86 1.88 1.91 1.96 2.05 2.10 2.14 2.18 2.25 2.31

6 Hole 3.71 3.43 2.77 3.23 2.97 3.30 3.46 3.51 3.59 3.66 3.69 3.70 3.67

Broken 4.85 6.84 3.46 3.36 3.84 4.86 5.24 5.14 5.02 4.88 4.80 4.78 4.78

Crack 2.83 3.99 4.71 3.46 2.53 2.41 2.73 2.94 3.12 3.27 3.38 3.44 3.47

Normal 2.73 3.03 1.43 1.23 1.46 1.51 1.49 1.49 1.47 1.43 1.40 1.39 1.41

7 Hole 5.41 4.79 3.97 4.38 3.72 3.87 4.14 4.26 4.39 4.56 4.80 5.04 5.24

Broken 4.54 5.77 2.69 2.37 3.36 3.61 3.70 3.79 3.76 3.75 3.85 4.02 4.19

Crack 2.99 4.31 2.91 2.20 1.51 1.60 1.71 1.86 2.06 2.32 2.53 2.71 2.86

Normal 2.18 3.51 1.09 1.23 1.11 1.00 1.07 1.13 1.42 1.76 1.98 2.09 2.14

Scale\Direction	1	2	3	4	5	6	7	8	9	10	11	12	13
(a) Mean
1	Hole	11.2	9.85	8.19	8.51	8.94	9.13	9.23	9.35	9.47	9.58	9.68	9.77	9.85
	Broken	15.7	15.9	13.0	13.6	13.8	14.0	14.1	14.2	14.2	14.3	14.3	14.4	14.5
	Crack	8.31	8.85	9.71	9.31	9.12	9.12	9.20	9.24	9.14	9.24	9.14	9.08	9.03
	Normal	11.4	11.9	11.7	11.7	11.8	11.7	11.6	11.4	11.3	11.2	11.1	11.0	11.0
2	Hole	8.73	8.24	6.61	6.95	7.37	7.65	7.74	7.79	7.85	7.92	8.00	8.06	8.12
	Broken	15.4	15.9	12.0	12.0	12.5	13.3	13.7	13.8	13.9	14.0	14.0	14.1	14.2
	Crack	7.34	7.67	9.77	8.35	7.69	7.78	7.96	7.99	7.94	7.86	7.79	7.74	7.69
	Normal	9.32	10.1	8.65	8.93	9.14	9.24	9.24	9.20	9.18	9.17	9.15	9.16	9.17
3	Hole	7.52	6.78	5.97	5.79	5.93	6.13	6.21	6.25	6.31	6.41	6.54	6.66	6.77
	Broken	14.3	15.9	11.0	10.8	10.8	11.8	12.6	13.0	13.2	13.3	13.3	13.3	13.3
	Crack	6.54	7.95	10.2	8.20	7.27	7.26	7.41	7.49	7.49	7.43	7.46	7.28	7.21
	Normal	7.88	8.15	6.67	7.02	7.53	7.57	7.47	7.44	7.45	7.45	7.47	7.48	7.49
4	Hole	7.58	6.26	5.29	5.35	5.63	5.79	5.85	5.93	6.04	6.19	6.37	6.53	6.69
	Broken	14.0	14.7	8.91	9.93	10.4	11.7	12.4	12.5	12.6	12.6	12.6	12.7	12.7
	Crack	6.15	7.53	10.2	7.97	7.66	7.66	7.63	7.66	7.66	7.63	7.58	7.51	7.44
	Normal	6.51	7.02	6.07	5.34	5.15	4.29	5.35	5.38	5.43	5.48	5.54	5.59	5.65
5	Hole	7.87	7.98	5.87	6.13	5.89	5.96	6.18	6.34	6.45	6.57	6.67	6.75	6.81
	Broken	13.4	15.1	8.34	8.91	9.65	10.5	11.2	11.7	11.9	12.0	12.0	12.0	12.0
	Crack	6.35	8.51	10.1	6.44	5.80	6.36	6.76	6.95	7.05	7.10	7.13	7.14	7.14
	Normal	5.97	6.77	4.74	4.83	4.00	4.00	4.21	4.30	4.35	4.39	4.45	4.51	4.59
6	Hole	9.00	7.31	6.00	5.91	6.18	6.34	6.62	6.81	6.88	6.90	6.93	6.98	7.07
	Broken	12.6	13.7	7.67	6.73	8.55	9.69	10.1	10.3	10.4	10.6	10.8	10.9	11.1
	Crack	6.28	8.81	9.99	6.95	5.42	5.23	5.70	6.16	6.42	6.50	6.51	6.48	6.46
	Normal	5.40	6.81	3.35	4.14	3.91	3.52	3.36	3.29	3.23	3.20	3.20	3.23	3.29
7	Hole	12.2	11.1	8.52	8.45	8.59	8.72	9.05	9.40	9.76	10.1	10.5	10.7	10.9
	Broken	11.1	13.5	6.33	6.01	7.58	9.02	9.57	9.59	9.60	9.69	9.80	9.90	10.0
	Crack	7.37	9.43	8.27	5.50	4.54	4.03	4.12	4.38	4.81	5.37	5.92	6.37	6.72
	Normal	6.65	7.85	3.08	3.25	3.51	3.42	3.34	3.37	3.63	4.06	4.57	5.05	5.45
(b) variance
1	Hole	6.18	5.53	4.39	4.73	5.10	5.16	5.18	5.22	5.26	5.29	5.32	5.34	5.37
	Broken	11.1	11.7	7.90	8.80	9.01	8.92	8.87	8.93	9.05	9.21	9.36	9.49	9.59
	Crack	5.24	6.08	7.82	6.68	6.12	6.10	6.19	6.27	6.30	6.30	6.30	6.28	6.19
	Normal	6.03	6.51	6.22	6.06	6.05	6.13	6.08	6.02	5.97	5.95	5.94	5.96	5.98
2	Hole	4.80	4.69	3.78	4.11	4.22	4.31	4.37	4.43	4.43	4.41	4.39	4.38	4.38
	Broken	11.0	12.2	7.34	8.24	8.04	8.47	8.85	9.17	9.39	9.52	9.60	9.65	9.70
	Crack	4.87	5.67	9.14	6.49	5.50	5.53	5.69	5.82	5.87	5.88	5.86	5.83	5.79
	Normal	5.21	5.44	4.52	4.59	4.84	4.95	4.93	4.86	4.80	4.77	4.78	4.80	4.82
3	Hole	4.58	4.16	3.74	3.62	3.78	3.99	4.09	4.11	4.12	4.14	4.16	4.18	4.20
	Broken	9.76	12.1	6.67	7.09	7.39	8.28	8.80	9.00	7.04	9.05	9.08	9.11	9.12
	Crack	4.15	5.55	9.52	6.67	5.32	5.12	5.32	5.54	5.64	5.65	5.60	5.51	5.41
	Normal	4.17	4.32	3.51	3.73	3.77	3.67	3.61	3.60	3.64	3.68	3.71	3.73	3.75
4	Hole	5.56	3.75	3.28	3.40	3.48	4.45	4.58	4.72	4.88	4.02	4.16	4.29	4.41
	Broken	7.92	10.7	5.64	6.08	6.85	7.73	8.09	8.07	7.94	7.84	7.81	8.81	8.83
	Crack	3.84	5.29	8.37	6.15	4.91	4.91	5.26	5.52	5.70	5.76	5.74	5.67	5.58
	Normal	3.27	3.48	3.00	2.41	2.46	2.51	2.53	2.55	2.57	2.62	2.68	2.73	2.77
5	Hole	4.67	4.59	3.45	3.77	3.64	3.60	3.69	3.79	3.90	3.95	3.98	4.03	4.10
	Broken	6.82	9.34	4.77	4.67	5.14	6.12	6.56	6.46	6.24	6.12	6.07	6.05	6.03
	Crack	3.41	5.10	6.40	4.08	3.24	3.51	4.14	4.65	4.91	5.00	4.98	4.91	4.81
	Normal	3.30	3.47	2.12	1.86	1.88	1.91	1.96	2.05	2.10	2.14	2.18	2.25	2.31
6	Hole	3.71	3.43	2.77	3.23	2.97	3.30	3.46	3.51	3.59	3.66	3.69	3.70	3.67
	Broken	4.85	6.84	3.46	3.36	3.84	4.86	5.24	5.14	5.02	4.88	4.80	4.78	4.78
	Crack	2.83	3.99	4.71	3.46	2.53	2.41	2.73	2.94	3.12	3.27	3.38	3.44	3.47
	Normal	2.73	3.03	1.43	1.23	1.46	1.51	1.49	1.49	1.47	1.43	1.40	1.39	1.41
7	Hole	5.41	4.79	3.97	4.38	3.72	3.87	4.14	4.26	4.39	4.56	4.80	5.04	5.24
	Broken	4.54	5.77	2.69	2.37	3.36	3.61	3.70	3.79	3.76	3.75	3.85	4.02	4.19
	Crack	2.99	4.31	2.91	2.20	1.51	1.60	1.71	1.86	2.06	2.32	2.53	2.71	2.86
	Normal	2.18	3.51	1.09	1.23	1.11	1.00	1.07	1.13	1.42	1.76	1.98	2.09	2.14

Table 2 shows the variance of various damage cases when the frequency bandwidths of W = π/4 and W = π/8. It can be seen from Table 2 that when the frequency bandwidth is W = π/4, the mean and variance of the holes are larger on the 6th and 7th scales, the mean and variance of the broken are larger than the crack and normal, and the average and variance of the normal cement surface are generally The smallest. Similarly, when the frequency bandwidth W = π/8, the 4th and 5th scales also have such numerical characteristics. From the numerical value, the amplitude of the variance is the same as the amplitude of W = π/4, but the selection of the scale is different. Therefore, a moderate frequency bandwidth of W = π/4 is taken here as feature extraction. In the 6-scale and 7-scale of the frequency bandwidth of W = π/4, the 6-scale has a linear relationship in multiple directions and the difference spacing of each case is 0.88, and the linear relationship is more obvious. Therefore, the research is carried out in multiple directions at 6 scales to find the most obvious scale and direction parameter groups. This parameter group is used for classification and identification to verify the reliability of feature extraction analysis.

In Table 2(b), the amplitude variance of the 6 scale in the 5∼13 direction is shown in detail on the number axis. Among these scale and direction parameter groups, 6/12 is easier to distinguish between various types of damage in terms of eigenvalue stability and value range. Therefore, the 6-scale 12 direction is taken here as the experimental parameter set.

Table 2

Multi-scale and multi-directional amplitude characteristics

Scale\Direction		1	2	3	4	5	6	7	8	9	10	11	12	13
(a) The frequency bandwidth is the variance of W =π/4
1	Hole	4.58	4.16	3.74	3.62	3.78	3.99	4.09	4.11	4.12	4.14	4.16	4.18	4.20
	Broken	9.76	12.1	6.67	7.09	7.39	8.28	8.80	9.00	9.04	9.05	9.08	9.11	9.12
	Crack	4.15	5.55	9.52	6.67	5.32	5.12	5.32	5.54	5.64	5.65	5.60	5.51	5.41
	Normal	4.17	4.32	3.51	3.73	3.77	3.67	3.61	3.60	3.64	3.68	3.71	3.73	3.75
2	Hole	5.56	3.75	3.28	3.40	3.48	3.45	3.58	3.72	3.88	4.02	4.16	4.29	4.41
	Broken	7.92	10.7	5.64	6.08	6.85	7.73	8.09	8.07	7.94	7.84	7.81	7.81	7.83
	Crack	3.84	5.59	8.37	6.15	4.91	4.91	5.26	5.52	5.70	5.76	5.74	5.67	5.58
	Normal	3.27	3.48	3.00	2.41	2.46	2.51	2.53	2.55	2.57	2.62	2.68	2.73	2.77
3	Hole	4.67	4.59	3.45	3.77	3.64	3.60	3.69	3.79	3.90	3.95	3.98	4.03	4.10
	Broken	6.82	9.34	4.77	4.67	5.14	6.12	6.56	6.46	6.24	6.12	6.07	6.05	6.03
	Crack	3.41	5.10	6.40	4.08	3.24	3.51	4.14	4.65	4.91	5.00	4.98	4.91	4.81
	Normal	3.30	3.47	2.12	1.86	1.88	1.91	1.96	2.05	2.10	2.14	2.18	2.25	2.31
4	Hole	3.71	3.43	2.77	3.30	2.97	3.30	3.46	3.51	3.59	3.66	3.69	3.70	3.67
	Broken	4.85	6.84	3.46	3.36	3.84	4.86	5.24	5.14	5.02	4.88	4.80	4.78	4.78
	Crack	2.83	3.99	4.71	3.23	2.53	2.41	2.73	2.94	3.12	3.27	3.38	3.44	3.47
	Normal	2.73	3.03	1.43	1.23	1.46	1.51	1.49	1.49	1.47	1.43	1.40	1.39	1.41
5	Hole	5.41	4.79	3.97	4.38	3.72	3.87	4.14	4.26	4.39	4.56	4.80	5.04	5.24
	Broken	4.54	5.77	2.69	2.37	3.36	3.61	3.70	3.79	3.76	3.75	3.85	4.02	4.19
	Crack	2.99	4.31	2.91	2.20	1.51	1.60	1.74	1.86	2.06	2.32	2.53	2.71	2.86
	Normal	2.17	3.51	1.09	1.23	1.11	1.00	1.07	1.13	1.42	1.76	1.98	2.09	2.14
6	Hole	9.45	8.48	4.01	3.36	3.94	4.04	3.29	4.92	6.47	7.22	7.62	7.86	8.03
	Broken	6.05	7.37	3.25	1.48	2.47	2.76	2.32	3.76	4.66	5.15	5.40	5.56	5.67
	Crack	3.45	6.00	2.78	1.25	2.02	2.11	1.33	2.45	2.72	2.85	2.97	3.07	3.15
	Normal	2.98	4.79	1.71	0.71	1.50	1.55	1.10	1.68	1.85	1.97	2.08	2.18	2.27
7	Hole	5.99	5.18	4.32	5.71	3.20	4.39	6.09	6.31	6.33	6.32	6.31	6.31	6.31
	Broken	4.17	4.83	3.23	3.73	2.53	2.82	3.30	3.54	3.71	3.85	3.97	4.06	4.14
	Crack	3.44	3.47	2.62	2.68	1.84	1.76	1.86	2.03	2.20	2.37	2.52	2.65	2.76
	Normal	2.20	3.82	1.74	1.99	1.45	1.45	1.45	1.51	1.60	1.68	1.75	1.80	1.85
(b) The frequency bandwidth is the variance of W =π/8
1	Hole	4.67	4.60	3.45	3.77	3.64	3.60	3.69	3.79	3.90	3.95	3.98	4.03	4.10
	Broken	6.82	9.34	4.77	4.67	5.14	6.12	6.56	6.46	6.24	6.12	6.07	6.05	6.03
	Crack	3.41	5.10	6.40	4.08	3.24	3.51	4.14	4.65	4.91	5.00	4.98	4.91	4.81
	Normal	3.30	3.47	2.12	1.86	1.88	1.91	1.96	2.05	2.10	2.14	2.18	2.25	2.31
2	Hole	3.71	3.43	2.77	3.23	2.97	3.30	3.46	3.51	3.59	3.66	3.69	3.70	3.67
	Broken	4.85	6.84	3.46	3.36	3.84	4.86	5.24	5.14	5.02	4.88	4.80	4.78	4.78
	Crack	2.83	3.99	4.71	3.46	2.53	2.41	2.73	2.94	3.12	3.27	3.38	3.44	3.47
	Normal	2.73	3.03	1.43	1.23	1.46	1.51	1.49	1.49	1.47	1.43	1.40	1.39	1.41
3	Hole	5.41	4.79	3.97	4.38	3.72	3.87	4.14	4.26	4.39	4.56	4.80	5.04	5.24
	Broken	4.54	5.77	2.69	2.37	3.36	3.61	3.70	3.79	3.76	3.75	3.85	4.02	4.19
	Crack	2.99	4.31	2.91	2.20	1.51	1.60	1.74	1.86	2.06	2.32	2.54	2.71	2.86
	Normal	2.18	3.51	1.09	1.23	1.11	1.00	1.07	1.13	1.42	1.76	1.98	2.09	2.14
4	Hole	9.45	8.58	4.01	3.36	3.94	4.04	3.29	4.92	6.47	7.22	7.62	7.86	8.03
	Broken	6.05	7.37	3.25	1.48	2.47	2.76	2.32	3.76	4.66	5.15	5.40	5.56	5.67
	Crack	3.45	6.00	2.78	1.25	2.02	2.11	1.33	2.45	2.72	2.85	2.97	3.07	3.15
	Normal	2.98	4.79	1.72	0.71	1.50	1.55	1.10	1.68	1.85	1.97	2.08	218	2.27
5	Hole	5.99	5.18	4.32	5.71	3.20	4.39	6.09	6.31	6.33	6.32	6.31	6.31	6.31
	Broken	4.17	4.83	3.23	3.73	2.53	2.82	3.30	3.54	3.71	3.85	3.97	4.06	4.14
	Crack	3.44	3.47	2.62	2.68	1.84	1.76	1.86	2.03	2.20	2.37	2.52	2.65	2.76
	Normal	2.20	3.82	1.74	1.99	1.45	1.45	1.45	1.51	1.60	1.68	1.75	1.80	1.85
6	Hole	3.90	3.18	3.91	0.70	3.10	4.13	4.11	4.06	4.02	4.00	3.98	3.97	3.96
	Broken	3.41	3.45	2.78	0.58	2.15	2.67	2.78	2.87	2.94	3.00	3.06	3.10	3.14
	Crack	2.64	2.02	1.66	0.54	1.63	1.86	2.02	2.13	2.22	2.29	2.34	2.38	2.42
	Normal	2.40	2.15	1.56	0.53	1.53	1.71	1.82	1.92	1.99	2.06	2.11	2.15	2.18
7	Hole	1.67	1.07	1.92	2.14	2.18	2.17	2.16	2.15	2.15	2.14	2.13	2.12	2.11
	Broken	1.43	1.52	1.60	1.48	1.51	1.56	1.60	1.64	1.67	.1.69	1.70	1.71	1.72
	Crack	1.31	0.96	0.88	0.94	1.06	1.16	1.23	1.29	1.33	1.35	1.38	1.39	1.40
	Normal	1.19	1.05	0.88	0.87	0.96	1.05	1.12	1.17	1.21	1.24	1.26	1.27	1.28

3.2 Recognition result

Combined with the above analysis, a total of 520 images were collected, 360 of which (90 for each type of damage) were applied to Gabor-SVM algorithm training.160 (40 for each type of damage) images were used for testing. Here, the parameters of 1∼7 scale and 1∼13 direction are selected for Gabor wavelet feature extraction. The recognition rate is shown in Table 3.

Table 3
Multi-directional and multi-scale recognition rates when the frequency bandwidth is π/4

Scale\Direction 1 2 3 4 5 6 7 8 9 10 11 12 13

(a) Normal

1 1.00 1.00 0.85 0.85 0.88 0.83 0.88 0.93 0.83 0.80 0.85 0.78 0.83

2 0.98 0.95 0.85 0.88 0.93 0.85 0.88 0.88 0.83 0.80 0.88 0.88 0.88

3 0.78 0.93 0.88 0.88 0.90 0.93 0.95 0.85 0.88 0.88 0.90 0.95 0.90

4 0.75 0.98 0.90 0.93 0.88 0.93 0.93 0.93 0.93 0.95 0.93 0.95 0.95

5 0.83 0.83 0.88 0.93 0.98 0.95 0.95 1.00 0.95 0.93 0.98 0.98 0.98

6 0.80 0.40 0.78 0.80 0.68 0.70 0.63 0.68 0.68 0.78 0.73 0.98 0.95

7 0.68 0.18 0.83 0.58 0.78 0.68 0.48 0.48 0.58 0.73 0.70 0.95 0.95

(b) Crack

1 0.00 0.28 0.30 0.33 0.53 0.38 0.53 0.55 0.55 0.63 0.65 0.65 0.63

2 0.03 0.13 0.33 0.35 0.43 0.60 0.58 0.58 0.65 0.58 0.58 0.60 0.70

3 0.08 0.28 0.48 0.35 0.53 0.58 0.45 0.60 0.63 0.60 0.65 0.63 0.68

4 0.05 0.33 0.48 0.48 0.55 0.63 0.50 0.55 0.63 0.58 0.65 0.63 0.70

5 0.08 0.38 0.53 0.48 0.50 0.60 0.50 0.50 0.58 0.63 0.63 0.63 0.63

6 0.00 0.58 0.25 0.25 0.35 0.38 0.48 0.25 0.33 0.28 0.23 0.63 0.65

7 0.20 0.55 0.25 0.45 0.30 0.38 0.63 0.45 0.48 0.33 0.33 0.58 0.55

(c) Hole

1 0.13 0.05 0.53 0.73 0.75 0.65 0.58 0.70 0.65 0.73 0.65 0.70 0.60

2 0.23 0.35 0.40 0.60 0.75 0.50 0.60 0.68 0.68 0.73 0.70 0.58 0.65

3 0.50 0.35 0.50 0.55 0.68 0.70 0.68 0.68 0.78 0.85 0.78 0.73 0.75

4 0.50 0.25 0.73 0.65 0.83 0.83 0.78 0.83 0.88 0.85 0.85 0.78 0.83

5 0.50 0.48 0.78 0.78 0.85 0.88 0.80 0.88 0.88 0.85 0.88 0.85 0.88

6 0.25 0.28 0.48 0.30 0.28 0.38 0.43 0.50 0.43 0.45 0.50 0.88 0.88

7 0.35 0.30 0.43 0.33 0.45 0.45 0.30 0.43 0.38 0.33 0.55 0.83 0.83

(d) broken

1 0.25 0.05 0.10 0.40 0.55 0.53 0.45 0.53 0.65 0.80 0.60 0.65 0.68

2 0.15 0.28 0.13 0.23 0.53 0.58 0.58 0.63 0.78 0.83 0.80 0.73 0.75

3 0.13 0.13 0.18 0.33 0.60 0.60 0.73 0.80 0.85 0.95 0.88 0.88 0.85

4 0.15 0.38 0.20 0.48 0.70 0.60 0.85 0.85 0.95 0.98 0.90 0.93 0.88

5 0.23 0.28 0.30 0.68 0.78 0.70 0.85 0.88 0.95 0.95 0.90 0.90 0.95

0.95 0.53 0.03 0.13 0.18 0.43 0.35 0.28 0.25 0.40 0.40 0.20 0.90 0.93

7 0.25 0.35 0.30 0.35 0.38 0.35 0.53 0.38 0.40 0.43 0.25 0.93 0.93

Scale\Direction	1	2	3	4	5	6	7	8	9	10	11	12	13
(a) Normal
1	1.00	1.00	0.85	0.85	0.88	0.83	0.88	0.93	0.83	0.80	0.85	0.78	0.83
2	0.98	0.95	0.85	0.88	0.93	0.85	0.88	0.88	0.83	0.80	0.88	0.88	0.88
3	0.78	0.93	0.88	0.88	0.90	0.93	0.95	0.85	0.88	0.88	0.90	0.95	0.90
4	0.75	0.98	0.90	0.93	0.88	0.93	0.93	0.93	0.93	0.95	0.93	0.95	0.95
5	0.83	0.83	0.88	0.93	0.98	0.95	0.95	1.00	0.95	0.93	0.98	0.98	0.98
6	0.80	0.40	0.78	0.80	0.68	0.70	0.63	0.68	0.68	0.78	0.73	0.98	0.95
7	0.68	0.18	0.83	0.58	0.78	0.68	0.48	0.48	0.58	0.73	0.70	0.95	0.95
(b) Crack
1	0.00	0.28	0.30	0.33	0.53	0.38	0.53	0.55	0.55	0.63	0.65	0.65	0.63
2	0.03	0.13	0.33	0.35	0.43	0.60	0.58	0.58	0.65	0.58	0.58	0.60	0.70
3	0.08	0.28	0.48	0.35	0.53	0.58	0.45	0.60	0.63	0.60	0.65	0.63	0.68
4	0.05	0.33	0.48	0.48	0.55	0.63	0.50	0.55	0.63	0.58	0.65	0.63	0.70
5	0.08	0.38	0.53	0.48	0.50	0.60	0.50	0.50	0.58	0.63	0.63	0.63	0.63
6	0.00	0.58	0.25	0.25	0.35	0.38	0.48	0.25	0.33	0.28	0.23	0.63	0.65
7	0.20	0.55	0.25	0.45	0.30	0.38	0.63	0.45	0.48	0.33	0.33	0.58	0.55
(c) Hole
1	0.13	0.05	0.53	0.73	0.75	0.65	0.58	0.70	0.65	0.73	0.65	0.70	0.60
2	0.23	0.35	0.40	0.60	0.75	0.50	0.60	0.68	0.68	0.73	0.70	0.58	0.65
3	0.50	0.35	0.50	0.55	0.68	0.70	0.68	0.68	0.78	0.85	0.78	0.73	0.75
4	0.50	0.25	0.73	0.65	0.83	0.83	0.78	0.83	0.88	0.85	0.85	0.78	0.83
5	0.50	0.48	0.78	0.78	0.85	0.88	0.80	0.88	0.88	0.85	0.88	0.85	0.88
6	0.25	0.28	0.48	0.30	0.28	0.38	0.43	0.50	0.43	0.45	0.50	0.88	0.88
7	0.35	0.30	0.43	0.33	0.45	0.45	0.30	0.43	0.38	0.33	0.55	0.83	0.83
(d) broken
1	0.25	0.05	0.10	0.40	0.55	0.53	0.45	0.53	0.65	0.80	0.60	0.65	0.68
2	0.15	0.28	0.13	0.23	0.53	0.58	0.58	0.63	0.78	0.83	0.80	0.73	0.75
3	0.13	0.13	0.18	0.33	0.60	0.60	0.73	0.80	0.85	0.95	0.88	0.88	0.85
4	0.15	0.38	0.20	0.48	0.70	0.60	0.85	0.85	0.95	0.98	0.90	0.93	0.88
5	0.23	0.28	0.30	0.68	0.78	0.70	0.85	0.88	0.95	0.95	0.90	0.90	0.95
0.95	0.53	0.03	0.13	0.18	0.43	0.35	0.28	0.25	0.40	0.40	0.20	0.90	0.93
7	0.25	0.35	0.30	0.35	0.38	0.35	0.53	0.38	0.40	0.43	0.25	0.93	0.93

It can be seen from Table 3 that the normal slope image recognition rate is generally high, and most of them are distributed between 0.8 and 1.0. The image recognition rate of the slope image of the crack grows steadily from low to high, but the overall recognition rate is low, and most of it is distributed between 0.50 and 0.65. The image recognition rate of the slope image of the hole fluctuates greatly, which is obviously related to the change of the scale. The recognition rate of broken slope image has a certain relationship with the size of the scale. from low to high to low, the fluctuation is large, and the recognition rate is mostly between 0.80 and 0.95.

By comparing the recognition rate, it is found that the scale and direction with a mean of 0.85 are 4/9, 5/11, 6/12, 6/13 (scale/direction) and other parameter groups, 6 scale / 12 direction (6/12). The extraction effect of the feature is in the optimal range, which corresponds to the analysis conclusion of the feature of Section 4.2. Therefore, the 6-scale and 12-direction are the best identified parameter sets. In the 6 scale / 12 direction, the recognition rate of hole, broken, crack, and normal is 0.88, 0.90, 0.63, and 0.98, respectively.

3.3 Comparisons of various feature extraction methods

In the same environment, the experiments of histogram-SVM, gray level co-occurrence matrix-SVM, canny algorithm-SVM, and Gabor-Euclidean distance algorithm are also carried out, and the error analysis of Gabor-SVM recognition effect is carried out, such as Table 4 shows. From the recognition rate and discriminant effect of Table 4, the histogram-SVM, grayscale co-occurrence matrix-SVM, canny algorithm-SVM, and Gabor-Euclidean distance algorithm have higher overall recognition rate of slope image. The slope image recognition rate of cracks and holes is low. Compared with other methods, the Gabor-SVM method has better overall crack identification.

Table 4
Identification rate of multiple algorithm processing and recognition rate of not less than 0.7

Methods Normal Crack Hole Broken

(a) Recognition rate of multiple algorithms

Histogram-SVM 0.98 0 0.25 0.65

Gray level co-occurrence matrix-SVM 0.55 0.25 0.13 0.68

Canny algorithm-SVM 0.88 0.30 0 0.75

Gabor-Euclidean distance 0.73 0.15 0.25 0.90

Gabor-SVM 0.98 0.63 0.88 0.90

(b) Discrimination table with recognition rate not less than 0.7

Histogram-SVM ✓ × × ×

Gray level co-occurrence matrix-SVM × × × ×

Canny algorithm-SVM ✓ × × ✓

Gabor-Euclidean distance ✓ × ✓ ✓

Gabor-SVM ✓ × ✓ ✓

Methods	Normal	Crack	Hole	Broken
(a) Recognition rate of multiple algorithms
Histogram-SVM	0.98	0	0.25	0.65
Gray level co-occurrence matrix-SVM	0.55	0.25	0.13	0.68
Canny algorithm-SVM	0.88	0.30	0	0.75
Gabor-Euclidean distance	0.73	0.15	0.25	0.90
Gabor-SVM	0.98	0.63	0.88	0.90
(b) Discrimination table with recognition rate not less than 0.7
Histogram-SVM	✓	×	×	×
Gray level co-occurrence matrix-SVM	×	×	×	×
Canny algorithm-SVM	✓	×	×	✓
Gabor-Euclidean distance	✓	×	✓	✓
Gabor-SVM	✓	×	✓	✓

In Table 4, from the two methods of Gabor-SVM and Gabor-Euclidean distance, the normal and broken pictures processed by Gabor-European distance algorithm have better classification, but the crack and hole slope image processed by Gabor-European distance algorithm The recognition rate is low, reflecting the instability of its identification and classification. At the same time, it can be seen that the Gabor-SVM model has a higher recognition rate for the hole, normality and broken of the cement slope, but the recognition effect on the crack is not satisfactory. The reason is that the slit width of the crack is low in contrast with the background image, and after multi-directional and multi-scale division, the stripe of the crack is refined, and the probability of recognition error increases.

4 Conclusions

High-fill channel cement slope damage monitoring model, based on Gabor-SVM, was designed in this research. The characteristics of damage levels are obvious in the 6th scales and the 12th directions of Gabor wavelet filter, and its recognition effect is the best. The model of Gabor-SVM algorithm has the highest recognition effect on the normal condition of the canal slope, but the recognition effect on the crack condition is not ideal. The algorithm model of Gabor-SVM also needs retrain and test In the actual work.

Overall, the image detection method based on Gabor-SVM, realizes the monitoring of high fill slope damage. In addition, the image data in this paper is derived above the water surface, and the underwater image has not been acquired. The next step of this work will get the underwater slope image information of the high fill channel, to improve the recognition rate and reliability of the Gabor-SVM method.

Footnotes

Acknowledgments

This research work is supported by the Open Research Foundation of Key Laboratory of Sediments in Chinese Ministry of Water Resources (2017001), Scientific Research Fund of Henan Provincial Education Department (15A510003), Henan Province science and technology research program (172102210050). The authors would like to thank the editor and anonymous referees for their comments.

References

Cui

, Chen

J.-S.

and Wang

L.-L.

, Design scheme of seepage flow monitoring in high fill section of south-to-north water diversion project, West-china Exploration 25(2) (2013), 39–41. [DOI:10.3969/j.issn.1004-5716.2013.02.014].

Y.-J.

, Li

X.-Y.

, Wilson

, et al., Water loss by evaporation from china’s south-north water transfer project, Ecological Engineering 95 (2016), 206–215. [DOI:10.1016/j.ecoleng.2016.06.086].

Z.-G.

, Shen

L.-P.

, Li

M.-X.

, et al., Discussion on the strengthening measures of high fill channel in the middle route of south-to-north water diversion project, Yangtze River 44(16) (2013), 63–66. [DOI:10.3969/j.issn.1001-4179.2013.16.018].

Liu

, Wu

, Yin

, et al., Evaluation and change of Danjiangkou reservoir dynamic capacity, Chinese Journal of Applied Ecology 29(8) (2018), 2658–2666 (in Chinese). [DOI:10.13287/j.1001-9332.201808.013].

Gade

S.O.

, Weiss

, Peter

M.A.

, et al., Relation of electromagnetic emission and crack dynamics in epoxy resin materials, Journal of Nondestructive Evaluation 33(4) (2014), 711–723. [DOI:10.1007/s10921-014-0265-5].

Lin

Ch.-H.

, Lin

Ch.-P.

, Hung

Y.-Ch.

, et al., Application of geophysical methods in a dam project: Life cycle perspective and Taiwan experience, Journal of Applied Geophysics 158 (2018), 82–92. [DOI:10.1016/j.jappgeo.2018.07.012].

Bernstone

and Heyden

, Image analysis for monitoring of crack growth in hydropower concrete structures, Measurement 42(6) (2009), 878–893. [DOI:10.1016/j.measurement.2009.01.007].

Liu

M.-T.

, Wang

, Qin

Z.-N.

, et al., Research on monitoring model of high fill canal leakage based on SBKF-PNN fusion, Journal of Basic Science and Engineering 27(2) (2019), 878–893. [DOI:10.16058/j.issn.1005-0930.2019.02.005].

Sun

, Mou

, Qiu

, et al., Adaptive fuzzy control for non-triangular structural stochastic switched nonlinear systems with full state constraints, IEEE Transactions on Fuzzy Systems 11(09) (2018), 1–1. [doi:10.1109/TFUZZ.2018.2883374].

10.

Qiu

, Sun

, Wang

, et al., Observer-Based fuzzy adaptive event-triggered control for pure-feedback nonlinear systems with prescribed performance, IEEE Transactions on Fuzzy Systems 1 (2019), 1–1. [doi:10.1109/TFUZZ.2019.2895560].

11.

Wang

Y.X.

, On reasons for damages of the cement concrete pavement[J], Shanxi Architecture 35(5) (2009), 275–276. [DOI:10.3969/j.issn.1009-6825.2009.05.174].

12.

Loh

C.-H.

and Chao

S.-H.

, Centralized vs. pattern-level feature extraction for structural damage detection, Procedia Engineering 79 (2014), 479–489. [DOI:10.1016/j.proeng.2014.06.369].

13.

Yang

, An algorithm for estimating the precision of scene guidance based on gabor wavelet, Computer Engineering 29(18) (2003), 72–73. [DOI:10.3969/j.issn.1000-3428.2003.18.030].

14.

Yang

J.-H.

, Yao

J.-P.

, Fan

, et al., Research on veneer surface crack image processing methods, Forestry Machinery & Woodworking Equipment 45(4) (2017), 29–31. [DOI:10.16640/j.cnki.37-1222/t.2017.15.209].

15.

Sun

, Xing

, Xie

, et al., An adaptive threshold-based Canny algorithm for crack detection, Micro-computer & Its Applications 36(5) (2017), 35–37. [DOI:10.19358/j.issn.1674-7720.2017.05.012].

16.

Tian

, Chen

and Wang

, A classification method for epileptic electroencephalogram based on wavelet multi-scale analysis and particle swarm optimization algorithm, Neuroquantology 16(4) (2018).

17.

Hadizadeh

, Multi-resolution local Gabor wavelets binary patterns for gray-scale texture description, Elsevier Science Inc 65 (2015), 163–169. [DOI:10.1016/j.patrec.2015.07.038].

18.

Belcher

, Camp

and Krzhizhanovskaya

V.V.

, Crack detection in earth dam and levee passive seismic data using support vector machines, International Conference on Computational Science 80 (2016), 577–586. [DOI:10.1016/j.procs.2016.05.339].

19.

Xia

, Lu

, Yang

, et al., A hybrid method based on extreme learning machine and k-nearest neighbor for cloud classification of ground-based visible cloud image, Neurocomputing 160 (2015), 238–249. [DOI:10.1016/j.neucom.2015.02.022].

20.

Mandal

S.K.

, Chan

F.T.S.

and Tiwari

M.K.

, Leak detection of pipeline: An integrated approach of rough set theory and artificial bee colony trained SVM, Expert Systems with Applications 39(3) (2012), 3071–3080. [DOI:10.1016/j.eswa.2011.08.170].

21.

Xing

, Yang

and Guo

, Face recognition based on gabor enhanced marginal fisher model and error correction SVM International Conference on Advances in Neural Networks, Springer-Verlag (2011), 296–304. [DOI:10.1007/978-3-642-21090-7_35].

22.

Zhang

and Shen

, An approach of using gabor wavelets for texture feature extraction, Journal of Image & Graphics 15(002) (2010), 247–254. [DOI:10.11834/jig.20100210].

23.

Chaudhari

and Gulati

R.M.

, Script identification using gabor feature and SVM classifier, Procedia Computer Science 79 (2016), 85–92. [DOI:10.1016/j.procs.2016.03.012].

24.

Panja

and Pal

N.R.

, MS-SVM: minimally spanned support vector machine, Applied Soft Computing 64 (2018), 356–365. [DOI:10.1016/j.asoc.2017.12.017].

25.

Liu

, Wen

, Gao

, et al., SVM based multi-label learning with missing labels for image annotation, Pattern Recognition 78 (2018), 307–317. [DOI:10.1016/j.patcog.2018.01.022].