Paper: automatic detection of pixel-level cracks on dam surface based on deep participation network

Received: 2020.0 1. 17.

Most dam cracks can only be classified and roughly located. Pixel level semantic segmentation detection can provide more accurate and intuitive detection results. A dam surface crack detection algorithm based on depth convolution network is proposed. Firstly, the UAV is used to collect data, then the collected images are preprocessed (including cutting and manual labeling), and finally the designed CDDS network structure is trained, verified and tested.

Compare segnetu-netfcn based on resnet 152. ?

Dam is an important water conservancy structure of hydropower station. The safe operation of the dam is of great significance to the hydropower station. Cracks caused by structural deformation, earthquake and water flow have a serious impact on the dam and threaten the safe operation of hydropower stations. Therefore, the regular health assessment of dam structure, especially the detection of dam cracks, becomes particularly important.

According to the structural characteristics and crack strength of the dam cracks, people can evaluate and monitor the structural health of the dam. The traditional detection task of dam cracks is usually manual detection, but it is inefficient, time-consuming and laborious, and wastes a lot of manpower costs, so it is very necessary to detect cracks automatically and efficiently.

The crack detection algorithm based on computer vision has been widely studied. Most of these methods use traditional image processing techniques and machine learning methods to identify some simple structural damage.

These methods use manually extracted features to extract features from images, and then evaluate whether the extracted features represent defects. However, the results of the above methods are inevitably influenced by subjective factors.

Convolutional neural network (CNN) has made great progress in the field of image classification and recognition, and the crack detection algorithm based on CNN also shows good performance.

Characteristics of dam cracks:

Repair marks, large noise, complex background texture, no structure, uneven distribution, random crack position, blurred background and other shortcomings.

A pixel-level dam surface crack detection method is proposed, and feature extraction is carried out by using depth convolution network. Based on the positioning characteristics of shallow convolution layer and the abstract characteristics of deep convolution layer, multi-scale convolution cascade fusion and multi-dimensional loss value calculation are carried out to realize pixel-level segmentation of crack defects, which solves the obvious crack detection problem of dam surface with the advantages of high precision and high efficiency, eliminates possible safety hazards and ensures the safety of dam surface. The experimental results show that this method is the best method to detect pixel-level cracks on the dam surface.

Semantic segmentation

PSPNet [42], ICNet [43], Deeplabv3[44], UNet [45] and SegNet [46]

Semantic segmentation networks are usually divided into coding networks and decoding networks.

Coding network:

Convolution layer: used to extract the features of the input image.

Pool layer: reduce the scale of feature map and reduce the calculation burden.

Decoding network:

Deconvolution layer (deconvolution layer): Up-sampling restores the feature map with the same size as the input image, and outputs the prediction result.

The codec network structure is highly symmetrical: sparse feature maps and dense feature maps are used at the same time.

In order to integrate sparse and dense characteristics, a hop module is used to connect codec network.

Coding network:

15 convolution layer: 3*3? Step size 1

4 Pool Layer: 2*2 Step 2

Decoding network:

15 deconvolution 1* 1

4 pool layer

Dropout and BN are used to prevent over-fitting.

Skip branch

4. 1* 1 Convolution and deconvolution.

Each branch calculates the branch loss, and the cascade of four branch losses is a part of the total loss.

The input and output image sizes of the skipped branch remain unchanged.

The number of channels of convolution kernel must be equal to the number of channels of input tensor.

The maximum convolution kernel size of the downsampling matrix is 2*2, and the step size is 2.

Deconvolution is also called transposed convolution.

The feature map recovered by up-sampling is the same size as the input image.

Up-sampling methods: deconvolution and interpolation.

Deconvolution method: fill the outermost layer with zero-filling tensor, then deconvolution with deconvolution kernel and trim the first and last lines.

1000 5472*3648 pictures

Use LEAR software to mark manually.

504 data sets were obtained, 404 for training, 50 for verification and 50 for testing.

Using TensorFlow to build on Linux system.

Perform training, validation and testing on HP workstations equipped with 8 GB GPU.

The virtual python environment of CDDS network was established by Anaconda.

Evaluation index:

Precision precision indicates the probability that the basic authenticity of the sample is also cracked in all the samples predicted to be cracked.

Recall indicates the probability that one sample is predicted to be cracked among all samples marked as cracked. When there is a big gap between positive and negative samples, it is unreasonable to evaluate the performance only by accuracy or recall. TPR represents the probability that all samples marked as cracks are correctly predicted as cracks. TNR represents the probability of being correctly predicted as the background in all samples with labels as the background.

F- measure is a comprehensive index considering the influence of recall and accuracy.

IoU is a common method to evaluate the positioning accuracy in the field of target detection. IoU represents the ratio of the intersection of the prediction result and the actual ground value to the intersection of the union set.

The image of dam surface cracks can be divided into background and cracks. The number of background pixels is much larger than the number of cracked pixels. Usually, we will calculate the background arrears and crack the arrears at the same time, and then take the average of the two as the final arrears. IoU value is determined by background pixels, which can not accurately express the positioning accuracy of cracks.

Use three learning rates: 10 4, 10 5, 10 6.

Use softmax function to calculate probability.

Use dice loss to calculate network loss.

Crack skeleton extraction: fast thinning algorithm

Call OpenCV library for calculation.

Calculate the crack area, length and width.

Other fracture data sets are used for supplementary verification. On the test data set, the IOU and F metrics of the proposed CDDS network reach 67.45438+0% and 80. 14% respectively.

A little.