Composite material surface defect detection system and method

Abstract

The invention relates to the technical field of composite material surface defect detection, in particular to a composite material surface defect detection system and method.The method comprises the steps that collected composite material surface images are input into a convolutional neural network model, a support vector machine model and a random forest model; each model independently carries out defect detection on an input composite material surface image, outputs a probability group containing the probability of each defect category, generates an evaluation label for the probability group, ensures that the evaluation module can detect whether data loss occurs in the transmission process, then obtains the defect category probability output by each model, and determines whether the data loss occurs or not. And respectively carrying out weighted fusion on the probability of each category by using a weighted average formula, and determining the category of the surface defect of the composite material based on a set threshold value. The capturing capability of the system on different types of defect characteristics in composite material surface defect detection is guaranteed, and the beneficial technical effect of accurately and efficiently detecting the composite material surface defects is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of composite material surface defect detection, and in particular to a composite material surface defect detection system and method. Background Art

[0002] Composite materials have been widely used in high-end manufacturing fields such as aerospace, automobile manufacturing, shipbuilding and wind power generation due to their excellent properties such as light weight, high strength and corrosion resistance. However, due to the complex structure of composite materials, surface defects such as cracks, scratches, bubbles, delamination and fiber dislocation are prone to occur during their manufacturing and use. These defects not only affect the appearance of the material, but may also seriously affect the mechanical properties of the material and even cause catastrophic failures. Therefore, the development of an efficient and accurate surface defect detection system has important application value and theoretical significance.

[0003] A search of the prior art revealed that a composite material surface defect detection method based on an improved SegNet segmentation network is disclosed in a Chinese invention patent with publication number CN117132553A. This method changes the anti-pooling upsampling in the SegNet network decoder to bilinear interpolation upsampling, avoiding the sparsity in the anti-pooling operation, and changes the pooling index connection in the SegNet network to a skip connection, supplementing the information lost in the encoder downsampling process, and adding a bottom-up channel to the SegNet network to achieve multi-scale semantic fusion, so that each layer of the decoder obtains richer multi-scale semantic information;

[0004] However, in actual production life, the training data may have imbalanced categories or distribution changes. A single model may perform poorly for samples of a minority category. For different types of defects in composite surface defect detection, a single model may find it difficult to capture all patterns and features, and the generalization ability of the model is difficult to guarantee. Summary of the invention

[0005] The purpose of the present invention is to provide a composite material surface defect detection system and method in view of the problems existing in the background technology.

[0006] The technical solution of the present invention is a method for detecting surface defects of composite materials, comprising the following specific implementation steps:

[0007] Step 1: The image acquisition module transmits the acquired composite material surface image to the image preprocessing module, which performs denoising, normalization and contrast enhancement operations on the composite material surface image, and transmits the processed image data to the surface defect detection module a, the surface defect detection module b and the surface defect detection module c respectively;

[0008] Step 2: The surface defect detection module a inputs the preprocessed image into the convolutional neural network model, and the model outputs the defect category probability group P CNN , and generate an evaluation label L for it Ⅰ , which will then contain the probability group P CNN and the evaluation label L Ⅰ The evaluation data packet I is transmitted to the evaluation module;

[0009] The surface defect detection module b extracts feature vectors from the image, inputs the feature vectors into the support vector machine model, and the support vector machine outputs the defect category probability group P SVM , and generate an evaluation label L for it Ⅱ , which will then contain the probability group P SVM and the evaluation label L Ⅱ The evaluation data packet II is transmitted to the evaluation module;

[0010] The surface defect detection module c extracts feature vectors from the image, inputs the feature vectors into the random forest model, and the random forest classifies the feature vectors and outputs the defect category probability group P RF , and generate an evaluation label L for it Ⅲ , which will then contain the probability group P RF and the evaluation label L Ⅲ The evaluation data packet III is transmitted to the evaluation module;

[0011] Step 3: The evaluation module detects whether data loss occurs during data transmission. If data loss does not occur, the evaluation module obtains the probability group P CNN , probability group P SVM , probability group P RF The probability of each category is weighted and fused, and the defect category in the composite material surface image is confirmed based on the set threshold;

[0012] Step 4: The evaluation module annotates the defects detected in the composite material surface image and outputs a list containing the defect categories and their probabilities detected in the image.

[0013] Preferably, the surface defect detection module b uses Canny edge detection to extract edge information of the composite material surface image, uses local binary patterns to extract texture information of the composite material surface image, and uses Fourier transform to extract shape information of the composite material surface image, and converts the extracted edge information, texture information and shape information into feature vectors.

[0014] Preferably, the convolutional neural network adopts the VGG16 network architecture.

[0015] Preferably, the evaluation label L Ⅰ The generation method is as follows:

[0016] S4-1. Select large prime numbers p1 and q1 and calculate auxiliary generation parameters:

[0017] P A1Ⅰ =p1×q1;

[0018] P A1Ⅱ =(p1-1)(q1-1);

[0019] S4-2, select label generation factor θ;

[0020] Among them, 1<θ<P A1Ⅱ , and θ and P A1Ⅱ mutual prime;

[0021] S4-3. Calculate label evaluation factor ε = θ -1 mod P A1Ⅱ ;

[0022] S4-4. Generate evaluation label L Ⅰ =(p CNN,1 ||...||p CNN,i ||...||p CNN,n ) ε mod P A1Ⅰ ;

[0023] Among them, || is the string connection symbol, p CNN,i Indicates the probability of whether the i-th defect exists in the composite surface image.

[0024] Preferably, the evaluation label L Ⅱ The generation method is as follows:

[0025] S5-1. Select large prime numbers p2 and q2 and calculate auxiliary generation parameters:

[0026] P A2Ⅰ =p2×q2;

[0027] P A2Ⅱ =(p2-1)(q2-1);

[0028] S5-2, select label generation factor μ;

[0029] Among them, 1<μ<P A2Ⅱ , and μ and P A2Ⅱ mutual prime;

[0030] S5-3. Calculate label evaluation factor

[0031] S5-4. Generate evaluation labels

[0032] Among them, pSVM,i Indicates the probability of whether the i-th defect exists in the composite surface image.

[0033] Preferably, the evaluation label L Ⅱ The generation method is as follows:

[0034] S6-1. Select large prime numbers p3 and q3 and calculate auxiliary generation parameters:

[0035] P A3Ⅰ =p3×q3;

[0036] P A3Ⅱ =(p3-1)(q3-1);

[0037] S6-2, select label generation factor κ;

[0038] Among them, 1<κ<P A3Ⅱ , and κ and P A3Ⅱ mutual prime;

[0039] S6-3. Calculate label evaluation factor σ = κ -1 mod P A3Ⅱ ;

[0040] S6-4. Generate evaluation label L Ⅲ =(p RF,1 ||...||p RF,i ||...||p RF,n ) σ mod P A3Ⅰ ;

[0041] Among them, p RF,i Indicates the probability of whether the i-th defect exists in the composite surface image.

[0042] Preferably, the evaluation module detects whether data loss occurs during data transmission as follows:

[0043] S7-1. Calculate the following evaluation parameters:

[0044] P E1 =p CNN,1 ||...||p CNN,i ||...||p CNN,n ;

[0045] P E1 '=L Ⅰ θ mod P A1Ⅰ ;

[0046] P E2 =p SVM,1 ||...||p SVM,i||...||p SVM,n ;

[0047] P E2 '=L Ⅱ μ mod P A2Ⅰ ;

[0048] P E3 =p RF,1 ||...||p RF,i ||...||p RF,n ;

[0049] P E3 '=L Ⅲ σ mod P A3Ⅰ ;

[0050] S7-2, if equation P E1 =P E1 '、P E2 =P E2 '、P E3 =P E3 'If both are true at the same time, it means that there is no data loss during the transmission process.

[0051] The technical solution of the present invention is a composite material surface defect detection system, which is applicable to the above-mentioned composite material surface defect detection method, comprising:

[0052] An image acquisition module, used for acquiring images of the surface of the composite material;

[0053] An image preprocessing module is used to perform denoising, normalization, and contrast enhancement preprocessing operations on the collected composite material surface images;

[0054] Surface defect detection module a, uses a convolutional neural network model to automatically extract high-level image features and perform defect detection;

[0055] The surface defect detection module b uses a support vector machine model to perform defect detection using features extracted from the image;

[0056] The surface defect detection module c uses a random forest model to classify and detect defects based on the features extracted from the image;

[0057] The evaluation module is used to analyze and evaluate the prediction results output by the surface defect detection module a, the surface defect detection module b and the surface defect detection module c, and output the detected surface defect categories of the composite material.

[0058] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0059] The present invention proposes a composite material surface defect detection system and method. The method performs denoising and normalization processing on the collected composite material surface image through an image acquisition module, and inputs the processed image into a convolutional neural network (CNN) model, a support vector machine (SVM) model and a random forest (RF) model respectively. The convolutional neural network model, the support vector machine model and the random forest model independently perform defect detection on the input composite material surface image, output a probability group containing the probability of each defect category, and generate an evaluation label for the probability group to ensure that the evaluation module can detect whether data loss occurs during data transmission. After that, the evaluation module obtains the defect category probability output by each model, and performs weighted fusion on the probability of each category respectively, and confirms the composite material surface defect category based on a set threshold. The present invention ensures that the composite material surface defect detection system can capture different types of defect features in composite material surface defect detection, and achieves the beneficial technical effect of accurately and efficiently detecting composite material surface defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a structural diagram of a composite material surface defect detection system proposed by the present invention;

[0061] Figure 2 This is the network structure diagram of the convolutional neural network model in the surface defect detection module a. DETAILED DESCRIPTION

[0062] Embodiment 1, as Figure 1 As shown, a composite material surface defect detection system proposed by the present invention includes: an image acquisition module, an image preprocessing module, a surface defect detection module a, a surface defect detection module b, a surface defect detection module c and an evaluation module;

[0063] The image acquisition module uses a high-resolution imaging device (such as an industrial camera, an optical scanner or an X-ray imaging device) to acquire images of the composite surface;

[0064] The image preprocessing module is used to perform denoising, normalization, and contrast enhancement preprocessing steps on the collected composite material surface images to ensure the accuracy of subsequent detection;

[0065] The surface defect detection module a uses a convolutional neural network (CNN) model to automatically extract high-level image features and perform defect detection;

[0066] The surface defect detection module b uses a support vector machine (SVM) model to classify defects using low-level features extracted from the image;

[0067] The surface defect detection module c uses a random forest (RF) model to classify and detect defects based on the features extracted from the image;

[0068] The evaluation module is used to analyze and evaluate the prediction results output by the surface defect detection module a, the surface defect detection module b and the surface defect detection module c, and output the detected surface defect categories of the composite material.

[0069] Embodiment 2, a composite material surface defect detection method proposed by the present invention is applicable to a composite material surface defect detection system proposed in Embodiment 1, and its specific implementation steps are as follows:

[0070] S1, the image acquisition module shoots the surface of the composite material based on the configured high-definition industrial camera, collects the surface image of the composite material, and transmits the collected image data to the image preprocessing module;

[0071] S2, the image preprocessing module performs denoising, normalization and contrast enhancement operations on the collected composite material surface image, and transmits the processed image data to the surface defect detection module a, the surface defect detection module b and the surface defect detection module c respectively;

[0072] S3, surface defect detection module a, surface defect detection module b and surface defect detection module c respectively use built-in defect detection models to detect the received image data, and transmit the output probability of each defect category to the evaluation module;

[0073] The detection processes of the surface defect detection module a, the surface defect detection module b and the surface defect detection module c on the received image data are as follows:

[0074] (1) The surface defect detection module a inputs the preprocessed image into the convolutional neural network (CNN) model. The model extracts image features through multi-layer convolution and pooling, and finally outputs the probability distribution of defect categories through the fully connected layer and the sigmoid layer. The cross entropy is used as the loss function, and the optimizer uses Adam (adaptive moment estimation) to output the defect category probability group P CNN ={p CNN,1 , ..., p CNN,i , ..., p CNN,n};

[0075] Among them, p CNN,i Indicates the probability of the existence of the i-th type of defect in the composite surface image;

[0076] Afterwards, the surface defect detection module a is the defect category probability group P CNN Generate evaluation labels. The generation process is as follows:

[0077] Select large prime numbers p1 and q1 and calculate the auxiliary generation parameters:

[0078] P A1Ⅰ =p1×q1;

[0079] P A1Ⅱ =(p1-1)(q1-1);

[0080] Select the label generation factor θ(1<θ<P A1Ⅱ , and θ and P A1Ⅱ mutually prime);

[0081] Calculate the label evaluation factor ε = θ -1 mod P A1Ⅱ ;

[0082] Generate evaluation label L Ⅰ =(p CNN,1 ||...||p CNN,n ) ε mod P A1Ⅰ ;

[0083] Among them, || is the string connection symbol;

[0084] Output evaluation data packet {P CNN ,θ,P A1Ⅰ , L Ⅰ} and transmit it to the evaluation module;

[0085] It should be noted that if Figure 2 As shown, CNN adopts the VGG16 network architecture, and its network structure is as follows:

[0086] Input layer: input size is 224×224×32;

[0087] Convolutional layer 1: 64 3×3 convolution kernels, stride 1, ReLU activation function, output size 224×224×64;

[0088] Convolutional layer 2: 64 3×3 convolution kernels, stride 1, ReLU activation function, output size 224×224×64;

[0089] Pooling layer 1: max pooling, pooling window size is 2×2, stride is 2, output size is 112×112×64;

[0090] Convolutional layer 3: 128 3×3 convolution kernels, stride 1, ReLU activation function, output size 112×112×128;

[0091] Convolutional layer 4: 128 3×3 convolution kernels, stride 1, ReLU activation function, output size 112×112×128;

[0092] Pooling layer 2: maximum pooling, pooling window size is 2×2, stride is 2, output size is 56×56×128;

[0093] Convolutional layer 5: 256 3×3 convolution kernels, stride 1, ReLU activation function, output size 56×56×256;

[0094] Convolutional layer 6: 256 3×3 convolution kernels, stride 1, ReLU activation function, output size 56×56×256;

[0095] Convolutional layer 7: 256 3×3 convolution kernels, stride 1, ReLU activation function, output size 56×56×256;

[0096] Pooling layer 3: maximum pooling, pooling window size is 2×2, stride is 2, output size is 28×28×256;

[0097] Convolutional layer 8: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 28×28×512;

[0098] Convolutional layer 9: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 28×28×512;

[0099] Convolutional layer 10: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 28×28×512;

[0100] Pooling layer 4: maximum pooling, pooling window size is 2×2, stride is 2, output size is 14×14×512;

[0101] Convolutional layer 11: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 14×14×512;

[0102] Convolutional layer 12: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 14×14×512;

[0103] Convolutional layer 13: 512 3×3 convolution kernels, stride 1, ReLU activation function, output size 14×14×512;

[0104] Pooling layer 5: maximum pooling, pooling window size is 2×2, stride is 2, output size is 7×7×512;

[0105] Fully connected layer 1: The number of neurons is 4096, and the ReLU activation function is used;

[0106] Fully connected layer 2: The number of neurons is 4096, and the ReLU activation function;

[0107] Output layer: Use Sigmoid activation function to output defect category probability group P CNN ={p CNN,1 , ..., p CNN,i , ..., p CNN,n};

[0108] (2) The surface defect detection module b extracts feature vectors from the image and inputs the feature vectors into the SVM classifier. The SVM maps the features to a high-dimensional space through the kernel function and finds the optimal hyperplane for classification. It outputs the probability of each defect category and classifies them according to the probability, outputting the defect category probability group P SVM ={p SVM,1 , ..., p SVM,i , ..., p SVM,n};

[0109] Among them, p SVM,i Indicates the probability of the existence of the i-th type of defect in the composite surface image;

[0110] Afterwards, the surface defect detection module b is the probability group P SVM Generate evaluation labels. The generation process is as follows:

[0111] Select large prime numbers p2 and q2, and calculate the auxiliary generation parameters:

[0112] P A2Ⅰ =p2×q2;

[0113] P A2Ⅱ =(p2-1)(q2-1);

[0114] Select the label generation factor μ(1<μ<P A2Ⅱ , and μ and P A2Ⅱ mutually prime);

[0115] Calculate label evaluation factor

[0116] Generate evaluation labels

[0117] Output evaluation data packet {P SVM ,μ,P A2Ⅰ , L Ⅱ} and transmit it to the evaluation module;

[0118] It should be noted that the training and detection process of the SVM model is as follows:

[0119] Feature extraction: Canny edge detection is used to extract edge information of composite surface images, local binary pattern (LBP) is used to extract texture information of composite surface images, and Fourier transform is used to extract shape information of composite surface images;

[0120] Generate feature vector: Convert the extracted features into feature vectors, which represent specific information in the composite surface image;

[0121] Feature vector input: The feature vector extracted from the composite surface image is input into the SVM classifier;

[0122] Kernel function mapping: Select the radial basis function kernel (RBF), SVM maps the feature vector to a high-dimensional space in order to find the optimal hyperplane for classification;

[0123] Finding the optimal hyperplane: SVM uses an optimization algorithm to find a hyperplane that can most clearly separate the defect features of different categories. This hyperplane will maximize the classification interval (that is, the interval between categories is the largest);

[0124] The optimization algorithm is as follows:

[0125]

[0126] st:y i (w T x i +b)≥1

[0127] Where: w is the normal vector of the hyperplane, b is the deviation term, x i is the training sample, y i is the sample category, and the constraint (st) ensures that the correctly classified samples are on one side of the hyperplane, and the distance between the sample and the hyperplane is at least 1;

[0128] Prediction category: SVM classifies the input feature vector according to the optimal hyperplane found, and outputs the defect category probability group P SVM ={p SVM ,1,...,p SVM ,i,...,p SVM,n};

[0129] (3) The surface defect detection module c extracts feature vectors from the image and inputs the feature vectors into the random forest (RF) classifier. RF classifies the feature vectors through a decision tree and outputs a defect category probability group P through voting. RF ={p RF ,1,...,p RF,i,...,p RF,n};

[0130] Among them, p RF,i Indicates the probability of the existence of the i-th type of defect in the composite surface image;

[0131] Afterwards, the surface defect detection module c is the probability group P RF Generate evaluation labels. The generation process is as follows:

[0132] Select large prime numbers p3 and q3 and calculate the auxiliary generation parameters:

[0133] P A3Ⅰ =p3×q3;

[0134] P A3Ⅱ =(p3-1)(q3-1);

[0135] Select the label generation factor κ(1<κ<P A3Ⅱ , and κ and P A3Ⅱ mutually prime);

[0136] Calculate the label evaluation factor σ = κ -1 modP A3Ⅱ ;

[0137] Generate evaluation label L Ⅲ =(p RF,1 ||...||p RF,n ) σ modP A3Ⅰ ;

[0138] Output evaluation data packet {P RF ,σ,P A3Ⅰ , L Ⅲ} and transmit it to the evaluation module;

[0139] It should be noted that the training and testing process of the RF model is as follows:

[0140] Feature extraction: Canny edge detection is used to extract edge information of composite surface images, local binary pattern (LBP) is used to extract texture information of composite surface images, and Fourier transform is used to extract shape information of composite surface images;

[0141] Generate feature vector: convert the extracted features into feature vector;

[0142] Decision tree classification: Each decision tree in the random forest will classify based on the input feature vector, and each tree will give a predicted category, that is, the type of defect it believes is in the image;

[0143] Integrated prediction: Random forest uses the probability average method to vote on the results of all decision trees to determine the defect category and output the defect category probability group P RF ={p RF,1 , ..., p RF,i , ..., p RF,n};

[0144] Among them, p RF,i Indicates the probability of the existence of the i-th type of defect in the composite surface image;

[0145] S4, the evaluation module evaluates the evaluation data packets generated by the surface defect detection module a, the surface defect detection module b and the surface defect detection module c, and detects whether data loss occurs during the data transmission process. The detection process is as follows:

[0146] (1) From the received evaluation data packets {P CNN ,θ,P A1Ⅰ , L Ⅰ}、Evaluation data package {P SVM ,μ,P A2Ⅰ , L Ⅱ} and the evaluation data package {P RF ,σ,P A3Ⅰ , L Ⅲ} to extract the following parameters:

[0147] Auxiliary generation parameter P A1Ⅰ , label generation factor θ, evaluation label L Ⅰ , P CNN ={p CNN,1 , ..., p CNN,i , ..., p CNN,n};

[0148] Auxiliary generation parameter P A2Ⅰ , label generation factor μ, evaluation label L Ⅱ , P SVM ={p SVM,1 , ..., p SVM,i , ..., p SVM,n};

[0149] Auxiliary generation parameter P A3Ⅰ , label generation factor σ, evaluation label L Ⅲ , P RF ={p RF,1 , ..., p RF,i , ..., p RF,n};

[0150] (2) Calculate the following evaluation parameters:

[0151] P E1=p CNN,1 ||...||p CNN,n ;

[0152] P E1 '=L Ⅰ θ mod P A1Ⅰ ;

[0153] P E2 =p SVM,1 ||...||p SVM,n ;

[0154] P E2 '=L Ⅱ μ mod P A2Ⅰ ;

[0155] P E3 =p RF,1 ||...||p RF,n ;

[0156] P E3 '=L Ⅲ σ mod P A3Ⅰ ;

[0157] (3) If the equation P E1 =P E1 '、P E2 =P E2 '、P E3 =P E3 'If both are true, it means that there is no data loss during the transmission process, and then execute S5;

[0158] S5. The evaluation module obtains the defect category probabilities output by the surface defect detection module a, the surface defect detection module b, and the surface defect detection module c, and uses the weighted average formula to perform weighted fusion on the probability of each category:

[0159] p final,i (y|X)=α·p CNN,i (y|X)+β·p SVM,i (y|X)+γ·p RF,i (y|X);

[0160] Where: α, β, γ are weighting coefficients, p fimal,i is the final predicted probability of each type of defect;

[0161] S6, the evaluation module sets the threshold T, if the predicted probability p fimal,i ≥T, then the i-th type defect is considered to exist;

[0162] If the predicted probability pfimal,i <T, it is considered that the i-th type of defect does not exist;

[0163] The evaluation module then labels each detected defect in the composite surface image and outputs a list containing all defect categories and their probabilities detected in the image.

[0164] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for detecting surface defects of composite materials, characterized in that: The specific implementation steps include the following: Step 1: The image acquisition module transmits the acquired composite material surface image to the image preprocessing module, which performs denoising, normalization and contrast enhancement operations on the composite material surface image, and transmits the processed image data to the surface defect detection module a, the surface defect detection module b and the surface defect detection module c respectively; Step 2: The surface defect detection module a inputs the preprocessed image into the convolutional neural network model, and the model outputs the defect category probability group P CNN , and generate an evaluation label L for it Ⅰ , which will then contain the probability group P CNN and the evaluation label L Ⅰ The evaluation data packet I is transmitted to the evaluation module; The surface defect detection module b extracts feature vectors from the image, inputs the feature vectors into the support vector machine model, and the support vector machine outputs the defect category probability group P SVM , and generate an evaluation label L for it Ⅱ , which will then contain the probability group P SVM and the evaluation label L Ⅱ The evaluation data packet II is transmitted to the evaluation module; The surface defect detection module c extracts feature vectors from the image, inputs the feature vectors into the random forest model, and the random forest classifies the feature vectors and outputs the defect category probability group P RF , and generate an evaluation label L for it Ⅲ , which will then contain the probability group P RF and the evaluation label L Ⅲ The evaluation data packet III is transmitted to the evaluation module; Step 3: The evaluation module detects whether data loss occurs during data transmission. If data loss does not occur, the evaluation module obtains the probability group P CNN , probability group P SVM , probability group P RF The probability of each category is weighted and fused, and the defect category in the composite material surface image is confirmed based on the set threshold; Step 4: The evaluation module annotates the defects detected in the composite material surface image and outputs a list containing the defect categories and their probabilities detected in the image.

2. A composite material surface defect detection method according to claim 1, characterized in that: The surface defect detection module b uses Canny edge detection to extract edge information of the composite material surface image, uses local binary patterns to extract texture information of the composite material surface image, and uses Fourier transform to extract shape information of the composite material surface image, and converts the extracted edge information, texture information and shape information into feature vectors.

3. A composite material surface defect detection method according to claim 1, characterized in that: The convolutional neural network adopts the VGG16 network architecture.

4. A composite material surface defect detection method according to claim 1, characterized in that: Evaluation label L Ⅰ The generation method is as follows: S4-1. Select large prime numbers p1 and q1 and calculate auxiliary generation parameters: P A1Ⅰ =p1×q1; P A1Ⅱ =(p1-1)(q1-1); S4-2, select label generation factor θ; Among them, 1<θ<P A1Ⅱ , and θ and P A1Ⅱ mutual prime; S4-3. Calculate label evaluation factor ε = θ -1 mod P A1Ⅱ ; S4-4. Generate evaluation label L Ⅰ =(p CNN,1 ||...||p CNN,i ||...||p CNN,n ) ε mod P A1Ⅰ ; Among them, || is the string connection symbol, p CNN,i Indicates the probability of whether the i-th defect exists in the composite surface image.

5. A composite material surface defect detection method according to claim 1, characterized in that: Evaluation label L Ⅱ The generation method is as follows: S5-1. Select large prime numbers p2 and q2 and calculate auxiliary generation parameters: P A2Ⅰ =p2×q2; P A2Ⅱ =(p2-1)(q2-1); S5-2, select label generation factor μ; Among them, 1<μ<P A2Ⅱ , and μ and P A2Ⅱ mutual prime; S5-3. Calculate label evaluation factor S5-4. Generate evaluation labels Among them, p SVM,i Indicates the probability of whether the i-th defect exists in the composite surface image.

6. A composite material surface defect detection method according to claim 1, characterized in that: Evaluation label L Ⅱ The generation method is as follows: S6-1. Select large prime numbers p3 and q3 and calculate auxiliary generation parameters: P A3Ⅰ =p3×q3; P A3Ⅱ =(p3-1)(q3-1); S6-2, select label generation factor κ; Among them, 1<κ<P A3Ⅱ , and κ and P A3Ⅱ mutual prime; S6-3. Calculate label evaluation factor σ = κ -1 mod P A3Ⅱ ; S6-4. Generate evaluation label L Ⅲ =(p RF,1 ||...||p RF,i ||...||p RF,n ) σ mod P A3Ⅰ ; Among them, p RF,i Indicates the probability of whether the i-th defect exists in the composite surface image.

7. A composite material surface defect detection method according to claim 1 or 4 or 5 or 6, characterized in that: The evaluation module detects whether data loss occurs during data transmission as follows: S7-1. Calculate the following evaluation parameters: P E1 =p CNN,1 ||...||p CNN,n ; P E1 '=L Ⅰ θ mod P A1Ⅰ ; P E2 =p SVM,1 ||...||p SVM,n ; P E2 '=L Ⅱ μ mod P A2Ⅰ ; P E3 =p RF,1 ||...||p RF,n ; P E3 '=L Ⅲ σ mod P A3Ⅰ ; S7-2, if equation P E1 =P E1 '、P E2 =P E2 '、P E3 =P E3 'If both are true at the same time, it means that there is no data loss during the transmission process.

8. A composite material surface defect detection system, which is applicable to a composite material surface defect detection method according to any one of claims 1 to 7, characterized in that: include: An image acquisition module, used for acquiring images of the surface of the composite material; An image preprocessing module is used to perform denoising, normalization, and contrast enhancement preprocessing operations on the collected composite material surface images; Surface defect detection module a, uses a convolutional neural network model to automatically extract high-level image features and perform defect detection; The surface defect detection module b uses a support vector machine model to perform defect detection using features extracted from the image; The surface defect detection module c uses a random forest model to classify and detect defects based on the features extracted from the image; The evaluation module is used to analyze and evaluate the prediction results output by the surface defect detection module a, the surface defect detection module b and the surface defect detection module c, and output the detected surface defect categories of the composite material.

Images