CN110956591B - Dam crack image data enhancement method based on depth convolution generation countermeasure network - Google Patents

Abstract

The invention discloses a dam crack image data enhancement method based on a depth convolution generation countermeasure network, which comprises the following steps: (1) using dam real crack image data as input to generate a dam crack image; (2) and (4) judging the dam crack image generated in the step (1). On the basis of basic data enhancement operation, the method generates a crack image by using an image generator (DGM) which generates a countermeasure network based on depth convolution, judges the quality of the generated image by using an image discriminator (DDM) which generates the countermeasure network based on depth convolution, generates a new dam crack sample data set according to original small-sample dam crack data, and meets the requirements on the class and data volume of a training sample in the dam crack image identification technology.

Description

Dam crack image data enhancement method based on depth convolution generation countermeasure network

Technical Field

The invention belongs to the technical field of engineering safety monitoring, and particularly relates to a dam crack image data enhancement method based on a depth convolution generation countermeasure network.

Background

The hydraulic structure is influenced by various factors such as illumination, water flow impact, temperature change and the like in the long-term operation process, so that the problems of function degradation and surface crack generation are easy to occur, and the problems can influence the normal operation of the dam body. Dam crack detection is an important way for timely finding potential safety hazards of the dam. The traditional dam crack image recognition mainly depends on human vision, but the efficiency and the accuracy of crack detection recognition are reduced in long-time work. With the development of digital image processing technology, the traditional manual image crack detection method is gradually replaced by a crack detection method based on machine vision, and becomes one of important nondestructive detection methods in the field of structure crack detection.

The crack detection method based on machine vision requires a large number of image samples to be trained to obtain higher detection accuracy. However, the data volume of real fracture images generated by an actual production environment is limited, and more images which can be used for fracture image recognition training need to be obtained through a data enhancement method. Common data enhancement methods comprise turning, rotating, zooming, cutting, translating, random data erasing and the like, but the enhancement effect is not obvious, the enhanced image quality is not high, and the requirements on the type and the data volume of a training sample in the dam crack image identification technology cannot be met.

Disclosure of Invention

The purpose of the invention is as follows: in order to overcome the problems in the prior art, the invention provides a dam crack image data enhancement method for generating a countermeasure network based on depth convolution.

The technical scheme is as follows: in order to achieve the purpose, the invention provides a dam crack image data enhancement method based on a depth convolution generation countermeasure network, which comprises the following steps:

(1) using dam real crack image data as input to generate a dam crack image;

(2) and (4) judging the dam crack image generated in the step (1).

Further, the specific steps of generating the dam crack image in the step (1) are as follows:

(1.1) inputting a picture of a dam real crack subjected to basic data enhancement operation;

(1.2) performing convolution operation, setting parameters of a convolution layer to be 3 multiplied by 64, completing missing values by using 0, and using a rectification Linear Unit, namely a Rectified Linear Unit, ReLu as an activation function;

(1.3) sequentially using 12 same residual blocks to train and learn, removing a BN layer of each residual block to simplify a network structure, and using a Skip Connection mode to train, so that the problems of gradient explosion and gradient disappearance in the training process are solved;

and (1.4) after entering the convolutional layer, using Skip Connection, and finally outputting a generated picture.

Further, the specific steps of judging the generated dam crack image in the step (2) are as follows:

(2.1) inputting a picture to be distinguished;

(2.2) extracting picture features by using 5 layers of convolution layers, wherein the convolution layers adopt convolution kernels with the size of 4 multiplied by 4, and the number of channels in each layer is increased by 2 times, so that the local receptive field is increased;

(2.3) performing dimensionality reduction using a convolution layer with a convolution kernel of 1 × 1;

(2.4) training in a Skip Connection manner using 3 convolutional layers;

(2.5) performing a flattening (Flatten) operation on the image data;

and (2.6) outputting a judgment result by using full Connected Layer functions (namely Fully Connected Layer, FC and Sigmoid function), wherein the judgment result is the probability value that the picture is a real picture.

According to the dam crack image data enhancement method based on the deep convolution generation countermeasure network, the residual error network (ResNet) and the deep convolution generation countermeasure network (DCGAN) are combined, a simplified residual error block network structure is constructed, the number of characteristic diagram channels is increased, performance reduction caused by network deepening is relieved, and the quality of dam crack data enhancement images is improved. Specifically, a simplified residual block structure is established based on the concept of scaling convolution in the aspect of a network structure, the number of characteristic channels of an image quality discriminator is increased in the aspect of network width, the parameters of an original discriminator are improved, the network stability is improved, the number of residual blocks of an image generator part is increased in the aspect of network depth, so that the network hierarchy is deepened, and the performance loss caused by deepening of the network depth is reduced by utilizing a residual network. On the basis of carrying out basic data enhancement operation on dam crack image data, dam real crack image data is used as input, a dam crack image is generated by adopting an image generator (DGM) for generating a countermeasure network based on depth convolution, and the quality of the image generated by the DGM is judged by utilizing an image discriminator (DDM) for generating the countermeasure network based on depth convolution.

Has the advantages that: compared with the prior art, the invention has the following advantages:

compared with the prior art, the dam crack image data enhancement method based on the deep convolution generation countermeasure network processes the original image on the basis of the existing basic data enhancement operations such as rotation, scaling, random data erasure and the like, constructs a simplified residual block network structure, increases the number of characteristic image channels, reduces the performance reduction by using the residual network while deepening the network, and improves the quality of the dam crack data enhanced image.

Drawings

FIG. 1 is a flow chart of image generation in an exemplary embodiment;

FIG. 2 is a flow chart of image discrimination in an embodiment.

Detailed Description

The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.

On the basis of carrying out basic data enhancement operation on dam crack image data, taking the dam real crack image data as input, generating a dam crack image by adopting an image generator (DGM) for generating a countermeasure network based on depth convolution, wherein the structure of the generator is shown as figure 1, an image discriminator (DDM) for generating the countermeasure network based on depth convolution is used for judging the quality of the image generated by the DGM, the structure of the discriminator is shown as figure 2, and the specific flow is described as follows:

(1) image generation:

(1.1) inputting a picture of a dam real crack subjected to basic data enhancement operation;

(1.2) performing convolution operation, setting parameters of a convolution layer to be 3 multiplied by 64, completing missing values by using 0, and using a Rectified Linear Unit (ReLu) as an activation function;

(1.3) training and learning by sequentially using 12 same residual blocks, removing a BN layer of each residual block to simplify a network structure, and training by using a Skip Connection (Skip Connection) mode to solve the problems of gradient explosion and gradient disappearance in the training process;

and (1.4) after entering the convolutional layer, using Skip Connection, and finally outputting a generated picture.

(2) And (3) image discrimination:

(2.1) inputting a picture to be distinguished;

(2.2) extracting picture features by using 5 layers of convolution layers, wherein the convolution layers adopt convolution kernels with the size of 4 multiplied by 4, and the number of channels in each layer is increased by 2 times, so that the local receptive field is increased;

(2.3) performing dimensionality reduction using a convolution layer with a convolution kernel of 1 × 1;

(2.4) training in a Skip Connection manner using 3 convolutional layers;

(2.5) performing a flattening (Flatten) operation on the image data;

and (2.6) outputting a judgment result by using a full Connected Layer (FC) and a Sigmoid function, wherein the judgment result is the probability value that the picture is a real picture.

Claims (1)

1. A dam crack image data enhancement method based on a depth convolution generation countermeasure network is characterized by comprising the following steps:

(1) the method comprises the following steps of taking dam real crack image data as input, generating a dam crack image, and specifically comprising the following steps:

(1.1) inputting a picture of a dam real crack subjected to basic data enhancement operation;

(1.2) performing convolution operation, setting parameters of a convolution layer to be 3 multiplied by 64, completing missing values by using 0, and using a rectification Linear Unit, namely a Rectified Linear Unit, ReLu as an activation function;

(1.3) sequentially using 12 same residual blocks to train and learn, removing a BN layer of each residual block to simplify a network structure, and using a Skip Connection mode to train, so that the problems of gradient explosion and gradient disappearance in the training process are solved;

(1.4) using Skip Connection after entering the convolutional layer, and finally outputting a generated picture;

(2) and (2) judging the dam crack image generated in the step (1), and specifically comprising the following steps:

(2.1) inputting a picture to be distinguished;

(2.2) extracting picture features by using 5 layers of convolution layers, wherein the convolution layers adopt convolution kernels with the size of 4 multiplied by 4, and the number of channels in each layer is increased by 2 times, so that the local receptive field is increased;

(2.3) performing dimensionality reduction using a convolution layer with a convolution kernel of 1 × 1;

(2.4) training in a Skip Connection manner using 3 convolutional layers;

(2.5) performing a flattening Flatten operation on the image data;

and (2.6) outputting a judgment result by using full Connected Layer functions (namely Fully Connected Layer, FC and Sigmoid function), wherein the judgment result is the probability value that the picture is a real picture.

Images