CN108710919A - A kind of crack automation delineation method based on multi-scale feature fusion deep learning - Google Patents

Abstract

The present invention discloses the automatic delineation method in crack based on multi-scale feature fusion deep learning, and method includes：Crack Analysis On Multi-scale Features figure is obtained using learning strategy is migrated；The convolutional layer and bilinear interpolation for being 1 × 1 using convolution kernel size successively merge the characteristic pattern of multiple scales and finally obtain the fusion feature of multidimensional；The Pixel Information of multidimensional fusion feature is further merged using continuous multiple dimensioned full convolutional network, realizes the prediction for each pixel class in image.The present invention may learn the FRACTURE CHARACTERISTICS of multiple scales, the relationship between pixel in the relationship and regional extent between different scale feature respective pixel is fully considered, realize that the crack of quick high accuracy is delineated automatically, it can be suitable for the Crack Detection all types of complex environments.

Description

An automatic crack delineation method based on multi-scale feature fusion deep learning

technical field

The invention relates to the field of structure detection and evaluation, in particular to a method for automatic detection of structural surface cracks by using images, video screens and the like.

Background technique

Structural cracks are one of the most common diseases in the field of civil engineering, which can cause great harm to the durability and safety of structures. Therefore, cracks are one of the main evaluation indicators for the health of various types of structures. At present, the detection of cracks is still dominated by manual detection, which requires manual marking to outline the cracks, and then analyzes the length, width, and type of cracks. This detection method requires the use of scaffolding and other equipment, which is labor-intensive, low in safety, and low in detection efficiency.

Although the crack detection technology based on digital image processing has been gradually applied to structural crack detection, the traditional automatic crack delineation methods based on image processing, such as edge detection, threshold segmentation, and simple neural network, have poor applicability, and the detection effect is largely Relying on manual intervention, the anti-interference ability for a large number of stains in complex images in engineering, artificial marking interference, etc. is poor.

In recent years, with the astonishing success of deep learning in the field of image recognition, deep learning technology has gradually been applied to crack detection. In foreign countries, Chen, Zhiqiang et al. applied deep learning to crack recognition based on image processing, and used Faster RCNN to automatically detect cracks of different scales in the image. In China, Li Hui and others used the depth-restricted Boltzmann machine to realize the automatic detection of fatigue cracks in steel structures. In China, the patent document with the publication number CN106910186A discloses a bridge crack detection and positioning method based on CNN deep learning. Capable of outputting pixel-accurate crack binarized images. The patent document with publication number CN107133960A discloses an image crack segmentation method based on fully convolutional neural network (Fully Convolutional Networks, FCN), which has the disadvantage of being insensitive to the details of cracks, and does not fully consider the relationship between pixels and pixels. , ignoring the spatial regularization step used in the usual pixel classification based segmentation methods.

In general, crack status is one of the main indicators for evaluating the health status of various types of structures. Manually drawing cracks is time-consuming and laborious. The traditional crack detection technology based on digital image processing has poor applicability and is difficult to apply in complex engineering environments. The crack detection of learning has achieved certain results, but there is still a lot of room for improvement in practical engineering applications.

Contents of the invention

In view of the deficiencies in the existing technology, the present invention intends to disclose an automatic crack delineation method based on multi-scale feature fusion and deep learning, which can realize crack detection and positioning completely relying on deep learning itself, and complete high-precision automatic delineation.

The idea of the present invention is:

Based on Feature pyramid networks (FPN) and migration learning strategy, the present invention provides a deep learning network that realizes the location and segmentation of cracks in images, can automatically identify cracks in images, and integrates the layers of the deep learning network The learned features of different scales and dimensions are used to realize crack prediction accurate to each pixel by using the subsequent multi-layer full convolution layer.

In order to solve the problems of the technologies described above, the present invention adopts the following technical solutions:

An automatic crack delineation method based on multi-scale feature fusion deep learning, including:

(1) Crack qualitative detection and feature extraction method based on migration learning

The camera is used to collect images of cracks on the surface of different structures, and a database of crack images is constructed. In the field of image recognition, there are a large number of pre-trained deep learning models of various types of 1000-category image database (ImageNet database), such as GoogLeNet, VGG-16, ResNet 101, etc. The ability of these deep learning models for image classification has surpassed that of humans. . Replace the last layer of the pre-trained model that has saved image classification information, i.e. replace the number of classification outputs with 2 (crack, non-crack). Using the constructed crack image database, combined with the size of the input image of the pre-training model, the crack image training set and test set were constructed through image cropping, rotation, scaling and other operations. The database contains two types of crack images and non-crack images, and the crack image marks is 1, and non-cracked images are marked as 0. The pre-trained model parameters already include the prior knowledge of image qualitative classification, and use the built crack image database for transfer learning, train the model until the network converges, and finally obtain a high-quality crack image qualitative classification detection model.

When the trained deep learning model qualitatively detects the crack image, the output of each layer of the deep learning network is used to extract the features of different dimensions of the input crack image at each scale, and the feature extraction of the input image is completed. Due to the large number of layers in the deep learning network, it is not necessary to extract the output of all layers as the features of different scales of the image. The size of the output feature map of each layer of the deep learning network is often from large to small. For layers with the same output feature map size, only take Output the last layer of the feature map network of this size as the feature extraction layer, so that the crack feature map of each scale with the proportional reduction of the feature map scale can be obtained. Taking GoogLeNet as an example, the input image size is 224×224 pixels, then the size of the feature maps extracted in sequence is: 64 first-order feature maps of 112×112 size, 192 second-order feature maps of 56×56 size, 480 A third-order feature map of size 28×28, 832 fourth-order feature maps of size 14×14, and 1024 fifth-order feature maps of size 7×7.

(2) Multi-scale deep learning feature fusion method

On the basis of the existing trained deep learning model to extract multiple scale crack feature images, the features of each scale are fused layer by layer, and finally a multi-dimensional crack feature map is obtained. Among them, the size of the high-order feature image is small but the dimension is large, and the size of the low-order feature is large but the dimension is small. The fusion of features of different scales must be unified in both scale and dimension. The specific method is:

① When fusing multiple scale features, the adjacent two-level features are fused layer by layer. First, the higher-level two-scale features are fused into one scale feature, and the lower-level features are fused with the fused high-level feature image. feature image;

② When fusing adjacent two-order feature images, first unify the two-order features in dimension, and use the convolution layer with a convolution kernel size of 1×1 to achieve dimensionality reduction of high-order features, so that the dimension of high-order features Reduced to the same dimension as the low-level features. Taking GoogLeNet's 4th and 5th-order feature fusion as an example, the 4th-order feature is 832 14×14-sized fourth-order feature maps, and the fifth-order is 1024 7×7-sized fifth-order feature maps. The 5th-order feature map is followed by a convolution layer with a convolution kernel of 1×1 size, and the control output dimension is 832, so that the fifth-order feature map is reduced to 832 feature maps of 7×7 size;

③ After the high-order feature dimension reduction realizes the unification of dimensions, it is also necessary to unify the scale of feature images. The dimensionally reduced high-order feature image is enlarged to the same size as the low-order feature image by using bilinear interpolation. Also take GoogLeNet's 4th and 5th-order feature fusion as an example. After the 5th-order feature is reduced to 832, bilinear interpolation is used to scale the 7×7 feature map to 14×14 size. The first-order feature becomes 832 feature maps of 14×14 size;

④ When the dimensions and sizes of the adjacent two-level feature images are uniform, the fusion of the two scale features is realized by accumulating the pixel values of the corresponding pixels of the corresponding dimensions pixel by pixel. Taking GoogLeNet's 4th and 5th-order feature fusion as an example, when the dimension and size of the 5th-order feature are unified with the 4th-order feature, the pixel values of the 4th-order features are accumulated pixel by pixel in 832 dimensions, and the fusion 832 feature maps of size 14×14 are obtained.

⑤ After the selected multi-level features are fused, the convolutional layer with a convolution kernel size of 1×1 can be used to achieve dimensionality enhancement or dimensionality reduction in the final fusion feature dimension to adapt to the computer's operating capabilities.

By analogy, the fusion of multi-level features is completed by means of layer-by-layer fusion. The patent of the present invention also includes the method of merging low-order features with high-order features through a 1×1 convolutional layer to achieve dimensionality enhancement.

(3) Continuous multi-scale fully convolutional layer crack prediction method

After obtaining the multi-level and different scale feature fusion features, the subsequent multi-scale continuous full convolution layer is used to realize the crack prediction at the pixel level. The specific method is as follows:

① Build 4 consecutive multi-scale full convolutional layers with convolutional layers, nonlinear activation layers, and batch normalization layers, which do not include pooling layers.

② Among them, the size of the convolution kernel of the first full convolutional layer is a small scale, which is used for the linear combination of information of each pixel in different dimensions, such as a convolution kernel of 1×1 size or a volume of 3×3 size The product kernel, followed by the batch normalization layer and the nonlinear activation layer, the first layer of full convolution layer keeps the output dimension unchanged; the size of the convolution kernel of the second layer of full convolution layer is a large scale, which is used for each pixel The integration of pixel information of the point and its surrounding large-scale area, such as 21×21 size, etc., follows the batch normalization layer and nonlinear activation layer, and the second layer of full convolution layer keeps the dimension of the output unchanged. The structure of the third full convolutional layer is the same as that of the first full convolutional layer; the fourth full convolutional layer adopts a convolution kernel size of 1×1, and the output dimension is reduced to 1 dimension, and then the Sigmoid function is used to obtain each The final predicted value corresponding to pixels.

③The image obtained in the previous step is not the binary image of the crack. The image directly predicted by the network needs to be scaled by bilinear interpolation to the same size as the input image. A threshold of 0.5 is used to discriminate the scaled image, the pixel value greater than 0.5 is the crack pixel, and the pixel value less than 0.5 is the background pixel.

(4) The multi-scale feature fusion deep learning adopts the following method for sample training, specifically:

① First, use the migration learning strategy, combined with the constructed crack database, to train the network for qualitative crack identification, and use deep learning to extract crack features of different scales;

② In the training set and test set of crack qualitative identification, some crack images were extracted and manually sketched (the crack pixel value is 1, the background pixel value is 0), and the training set and test set of multi-scale feature fusion deep learning network were built. Among them, the label scale of the training set and the test set is changed according to the final output size of the network, and the size of the training set and the test set label is changed by bilinear interpolation, and the label is binarized with 0.3 as the threshold;

③Set the loss function of the training network to the cross-entropy cost function (cross-entropy cost function), and its expression is:

Among them, y is the label (0 or 1) of the corresponding pixel, and a is the actual output of the network.

④Using migration learning, the weight of the feature extraction layer of the deep learning network for qualitative crack identification is assigned to the multi-scale feature fusion deep learning network for automatic crack delineation, and the stochastic gradient descent algorithm is used to continuously update the weight of the network and reduce the value of the loss function. Gradually converge.

(5) Crack detection and positioning and automatic crack drawing

After the network training is completed, the automatic delineation of cracks is divided into two steps:

① Crack qualitative deep learning network realizes crack location detection. For the case where the image size is larger than the input of the deep learning detection model, the sliding window scanning detection method is used to scan and detect the image. The windows of each scanning detection are staggered in the x and y directions of the image by 1 /2 input image size, the approximate location of the crack is achieved according to the classification of each window detection image.

② On the basis of the crack location in the previous step, the image of each detected crack window is cut out, and used as the input of the multi-scale feature fusion deep learning network for automatic crack delineation, so as to output the automatic delineation of cracks in each window. The end result is automated delineation of cracks within the entire image.

Compared with the prior art, the beneficial effects of the present invention are as follows: the development equipment of the present invention can be widely used in the detection of cracks on the surface of structures, and it only takes a few seconds to complete the automatic outline of high-precision cracks:

(1) This technical solution can realize fast, automatic and accurate delineation of structural surface cracks. Compared with traditional methods: the algorithm proposed in the present invention does not require manual intervention to adjust coefficients such as threshold segmentation, and completely relies on the network itself to identify cracks. It has a wide range of applications. high degree of automation;

(2) Deep learning is used to extract fracture features at various scales, fully considering the relationship between pixels corresponding to different scale features and the relationship between pixels in the region, and can quickly achieve high-precision fracture region segmentation.

(3) This technology can be applied to the detection of cracks on the surface of various types of structures. It can effectively eliminate noise interference in different complex environments. It only takes a few seconds to complete the high-precision automatic detection of cracks in an image. It can be equipped with To the UAV or handheld detection platform to meet the needs of real-time image acquisition and real-time analysis.

Description of drawings

Fig. 1 is a schematic diagram of the crack automatic delineation method based on multi-scale feature fusion deep learning described in the present invention;

Fig. 2 is the GoogLeNet network schematic diagram in the specific embodiment of the present invention;

Fig. 3 is the 3rd, 4th, 5th order feature schematic diagrams that the fusion GoogLeNet network in the specific embodiment of the present invention extracts;

Fig. 4 is a schematic diagram of a continuous multi-scale fully convolutional layer in a specific embodiment of the present invention;

Fig. 5 is a comparative analysis diagram of crack detection under different complex environmental backgrounds in the example of the present invention.

Detailed ways

In order to make the purpose, content and advantages of the present invention more understandable, the implementation of the present invention will be introduced in detail below, but the protection scope of the present invention should not be limited thereby.

Embodiment: a kind of method for automatic delineation of cracks based on deep learning of multi-scale feature fusion, the method includes the following steps:

(1) Crack qualitative detection and feature extraction method based on migration learning,

(2) Multi-scale deep learning feature fusion method,

(3) Continuous multi-scale fully convolutional layer crack prediction method,

(4) The multi-scale feature fusion deep learning adopts the following method for sample training,

(5) Crack detection and location and crack automatic outline.

The details are as follows: Step 1. The deep learning model with an input of 224*224*3 pixels is the public GoogLeNet model. The output layer of the original network is 1000 categories. The output layer of the modified network is divided into 2 categories, namely cracks and non-cracks. Two, as shown in Figure 2. The 224*224*3 image crack training set was constructed, including 64,000 crack images and 6.4 non-crack images, and the image test set was constructed, including 16,000 crack images and 16,000 non-crack images. Using the strategy of migration learning, the weights of the open-source 1000-category classification pre-training model are assigned to the model with a modified classification number of 2 categories, and the stochastic gradient descent algorithm is used to update the weight of the network until the network converges. After the network is trained, the feature image sizes output by each layer of the network have the following five types: 112×112, 56×56, 28×28, 14×14, 7×7. Take the highest-level layer that outputs feature images of each size as the output layer of this size feature, and finally get: 112×112 size first-order features of 64 dimensions, 56×56 size second-order features of 192 dimensions, 480 dimensions 28 ×28 size third-order features, 832 dimensions 14×14 size fourth-order features, 1024 dimensions 7×7 size fifth-order features.

Step 2. Take the fusion of the 3rd, 4th, and 5th order feature images as an example, as shown in FIG. 3 . The fifth-order feature map has a size of 7×7×1024. First, the feature map is converted to a size of 7×7×832 through a convolution layer with a convolution kernel of 1×1 size, and then the feature map is converted into 14×14×832 in size; fuse the converted 5th-order features and 4th-order features, and accumulate pixel by pixel in the corresponding dimension to obtain the fused 4th-order feature map with a size of 14×14×832; Similarly, use the convolution layer with a convolution kernel of 1×1 size and the bilinear interpolation layer to convert the fused fourth-order feature map into a size of 28×28×480, and then perform the corresponding dimension one by one with the third-order features. The accumulation of pixels results in a fused third-order feature map with a size of 28×28×480; then, in order to reduce the burden of computer training, the data is further reduced in dimension, and the convolution layer with a convolution kernel of 1×1 size is used to fuse The dimension of the third-order feature map is reduced to 28×28×192, and the feature map is enlarged to 112×112×192 by two bilinear interpolation to complete the fusion of the 3rd, 4th, and 5th-order features.

Step 3: Build a multi-scale fully convolutional layer to achieve the purpose of predicting each pixel, as shown in Figure 4. The size of the final fusion feature obtained in step 2 is 112×112×192, followed by the first layer of small-scale full convolution layer with a convolution kernel size of 1×1, followed by a batch normalization layer and a nonlinear activation layer , the output dimension of this layer is 192 dimensions, and the output size remains unchanged; followed by the second layer of large-scale full convolution layer, the convolution kernel size is 21×21, followed by batch normalization layer and nonlinear activation layer, The dimension of the output of this layer is 112×112×192; the size of the convolution kernel of the third layer of small-scale full convolution layer is 1×1, followed by the batch normalization layer and the nonlinear activation layer. The dimension of the layer output is 192 dimensions, and the output dimension remains unchanged; followed by the fourth fully convolutional layer, which is the last fully convolutional layer, the convolution kernel size is 1×1, and the output dimension is 1, that is The output size of this layer is 112×112×1; the last layer is followed by the sigmoid function layer to realize pixel-by-pixel crack prediction.

Step 4. Build the training set and test set of the multi-scale feature fusion deep learning network, and train the network. Extract the crack images in the deep learning qualitative identification network crack training set and test set built in step 1, and manually outline the cracks to create a crack label file (the crack pixel is 1, and the non-crack pixel is 0). Because the output size of the network is set to 112×112, the artificially drawn crack labels are bilinearly interpolated to 112×112, and the binarized labels are obtained with 0.3 as the threshold. Set the loss function of the training network as the cross-entropy cost function, and use the migration learning strategy to assign the weight of the feature extraction layer of the crack qualitative identification deep learning network to the multi-scale feature fusion deep learning network for automatic crack delineation, and use the stochastic gradient descent algorithm to continuously update The weight of the network reduces the value of the loss function so that the network gradually converges.

Step 5. After training the two networks, for a crack image with an input of 4000×6000 pixels, first use the crack qualitative deep learning network to realize the crack location detection, and use the sliding window scanning detection method to scan the detection image. Each scan The detected windows are staggered by 112 pixels in the x and y directions in the image, and the cracks are roughly positioned according to the classification of the detected images of each window; the 224×224 (3-channel) size image of each detected crack window is cut out, As the input of the multi-scale feature fusion deep learning network for automatic crack delineation, the image size directly predicted by each window is 112×112 pixels, bilinearly interpolated to 224×224, and divided with a threshold of 0.5, so as to output each The cracks in each window can be automatically drawn (binary image), and finally the automatic drawing of cracks in the whole image can be completed.

Application example:

The present invention will be further described below through specific examples, but the protection scope of the present invention should not be limited thereby.

Crack images of different complexity are collected under different structural surfaces and different environmental backgrounds, and the invention is used to realize automatic detection of cracks and perform precision analysis.

(1) Deep learning scanning detection

Use the trained 224*224*3 input GoogLeNet deep learning model to scan and detect the image, and some output results are shown in the first column of Figure 5.

(2) Using multi-scale feature fusion deep learning network to realize automatic delineation of cracks

Cut out the 224×224 (3-channel) size image of each detected crack window, and use it as the input of the multi-scale feature fusion deep learning network for automatic crack delineation, so as to output the automatic delineation of cracks in each window, after bilinear Interpolation and threshold segmentation realize the binarization of the crack image in each window, and finally complete the automatic delineation of cracks in the entire image. Part of the output results are shown in the second column of Figure 5.

(3) Compared with traditional methods

In each window where cracks are detected in the first step, the results of adaptive threshold segmentation and Canny operator detection are visually compared. Some output results are shown in the third and fourth columns of Figure 5.

(4) Accuracy analysis

The binary image of cracks manually drawn is compared with the result of the binary image of cracks output by the method proposed by the invention. Divide the image into equally spaced 7*7 pixel squares, a total of 857*571 (rounded) squares. Comparing the manual and automatic drawing of cracks in each grid to count the detection accuracy.

TP (True Positive): The number of squares with cracks in both manual and automatic sketching cracks;

TN (True Negative): Count the number of squares with no cracks in manual sketching and automatic sketching;

FP (False Positive): Count the number of squares with no cracks in manual drawing and cracks in automatic drawing;

FN (False Negative): Count the number of squares with cracks drawn manually and without cracks automatically drawn;

Set up three indicators:

Precision=TP/(TP+FP), recall=TP/(TP+FN), F-measure=2×precision×recall/(precision+recall).

The method proposed by the invention can maintain high accuracy rate, high precision rate and high recall rate in different complex environments.

Claims (6)

1. A method for automatic delineation of cracks based on multi-scale feature fusion deep learning, characterized in that, the method comprises the following steps: (1) Crack qualitative detection and feature extraction method based on migration learning; (2) Multi-scale deep learning feature fusion method; (3) Continuous multi-scale fully convolutional layer crack prediction method; (4) The multi-scale feature fusion deep learning adopts the following method for sample training; (5) Crack detection and location and crack automatic outline. 2. The crack automatic delineation method based on multi-scale feature fusion deep learning according to claim 1, characterized in that, said step 1) crack qualitative detection and feature extraction method based on migration learning, specifically as follows, Use the camera to collect images of cracks on the surface of different structures, construct a database of crack images, use the constructed crack image database, combine the size of the input image of the pre-training model, and build a training set and test set of crack images through image cropping, rotation, scaling, etc. Operations, the database There are two types of cracked images and non-cracked images. The cracked image is marked as 1, and the non-cracked image is marked as 0. The pre-trained model parameters have included the prior knowledge of image qualitative classification, and the crack image database is used for migration. Learning, training the model until the network converges, and finally obtain a high-quality crack image qualitative classification detection model; when the trained deep learning model qualitatively detects the crack image, use the output of each layer of the deep learning network to extract the input crack image at various scales The features of different dimensions of the input image are extracted to complete the feature extraction of the input image. The size of the output feature map of each layer of the deep learning network is often from large to small. For the layer with the same output feature map size, only the last one of the output feature map network of this size The layer is used as the feature extraction layer, so that the feature maps of fractures at various scales can be obtained in which the scale of the feature map is reduced proportionally. 3. The crack automatic delineation method based on multi-scale feature fusion deep learning according to claim 1, characterized in that, the step 2) multi-scale deep learning feature fusion method is specifically as follows, ① When fusing multiple scale features, the adjacent two-level features are fused layer by layer. First, the higher-level two-scale features are fused into one scale feature, and the lower-level features are fused with the fused high-level feature image. feature image; ② When fusing adjacent two-order feature images, first unify the two-order features in dimension, and use the convolution layer with a convolution kernel size of 1×1 to achieve dimensionality reduction of high-order features, so that the dimension of high-order features Reduced to the same dimension as the low-level features; ③After the high-order feature dimensionality reduction realizes the unification of dimensions, it is necessary to realize the unification of the feature image scale, and the high-order feature image after dimensionality reduction is enlarged to the same size as the low-order feature image by using bilinear interpolation; ④ When the dimensions and sizes of the adjacent two-level feature images are uniform, the fusion of the two scale features is realized by accumulating the pixel values of the corresponding pixels of the corresponding dimensions pixel by pixel; ⑤ After the selected multi-level features are fused, the convolutional layer with a convolution kernel size of 1×1 can be used to achieve dimensionality enhancement or dimensionality reduction in the final fusion feature dimension to adapt to the computer's operating capabilities; By analogy, the fusion of multi-level features is completed by means of layer-by-layer fusion. 4. The crack automatic delineation method based on multi-scale feature fusion deep learning according to claim 1, characterized in that, the step 3) continuous multi-scale full convolution layer crack prediction method, specifically as follows, ① Build 4 consecutive multi-scale full convolutional layers with convolutional layers, nonlinear activation layers, and batch normalization layers, which do not include pooling layers; ② Among them, the size of the convolution kernel of the first full convolutional layer is a small scale, which is used for the linear combination of information of each pixel in different dimensions, such as a convolution kernel of 1×1 size or a volume of 3×3 size The product kernel, followed by the batch normalization layer and the nonlinear activation layer, the first layer of full convolution layer keeps the output dimension unchanged; the size of the convolution kernel of the second layer of full convolution layer is a large scale, which is used for each pixel The integration of pixel information of the point and its surrounding large-scale area, such as 21×21 size, etc., follows the batch normalization layer and nonlinear activation layer, and the second layer of full convolution layer keeps the dimension of the output unchanged. The structure of the third full convolutional layer is the same as that of the first full convolutional layer; the fourth full convolutional layer adopts a convolution kernel size of 1×1, and the output dimension is reduced to 1 dimension, and then the Sigmoid function is used to obtain each The final predicted value corresponding to pixels; ③The image obtained in the previous step is not the binary image of the crack. The image directly predicted by the network needs to be scaled by bilinear interpolation to the same size as the input image, and the scaled image is discriminated with a threshold of 0.5. The pixel value greater than 0.5 is the crack pixel, and the pixel value less than 0.5 is the background pixel. 5. The crack automatic delineation method based on multi-scale feature fusion deep learning according to claim 1, characterized in that, the multi-scale feature fusion deep learning in the step 4) adopts the following method for sample training, specifically as follows, ① First, use the migration learning strategy, combined with the constructed crack database, to train the network for qualitative crack identification, and use deep learning to extract crack features of different scales; ②In the training set and test set of crack qualitative identification, some crack images were extracted and manually sketched (the crack pixel value is 1, the background pixel value is 0), and the training set and test set of multi-scale feature fusion deep learning network were built. , the label scale of the training set and the test set changes with the final output size of the built network, the size of the training set and the test set label is changed by bilinear interpolation, and the label is binarized with 0.3 as the threshold; ③Set the loss function of the training network to the cross-entropy cost function (cross-entropy cost function), and its expression is: Among them, y is the label (0 or 1) of the corresponding pixel, and a is the actual output of the network; ④Using migration learning, the weight of the feature extraction layer of the deep learning network for qualitative crack identification is assigned to the multi-scale feature fusion deep learning network for automatic crack delineation, and the stochastic gradient descent algorithm is used to continuously update the weight of the network and reduce the value of the loss function. Gradually converge. 6. The crack automatic delineation method based on multi-scale feature fusion deep learning according to claim 1, characterized in that, said step 5) crack detection and positioning and crack automatic delineation, specifically as follows, after the network training is completed, the crack is realized Automatic sketching is divided into two steps: ① Crack qualitative deep learning network realizes crack location detection. For the case where the image size is larger than the input of the deep learning detection model, the sliding window scanning detection method is used to scan and detect the image. The windows of each scanning detection are staggered in the x and y directions of the image by 1 /2 input image size, according to the classification of each window detection image to achieve the approximate location of cracks; ② On the basis of the crack location in the previous step, the image of each detected crack window is cut out, and used as the input of the multi-scale feature fusion deep learning network for automatic crack delineation, so as to output the automatic delineation of cracks in each window. The end result is automated delineation of cracks within the entire image.