WO2021088101A1 - Insulator segmentation method based on improved conditional generative adversarial network - Google Patents

Abstract

An insulator segmentation method and system based on an improved conditional generative adversarial network, and an apparatus, aiming at solving the problems of poor segmentation precision and low efficiency of the insulator segmentation method based on the conditional generative adversarial network. The system method comprises: obtaining an image comprising an insulator as an input image (S100); and obtaining an insulator segmentation image by means of an insulator segmentation model on the basis of the input image, the insulator segmentation model being constructed on the basis of a generator of the conditional generative adversarial network, the generator being constructed on the basis of an auto-encoder and comprising an encoder and a decoder, the encoder comprising an asymmetric convolution layer and a maximum pooling layer, the decoder comprising an asymmetric convolution layer and an up-sampling layer, a training sample of the insulator segmentation model comprising an input image sample and a real segmentation image of an insulator comprised in the training sample (S200). By means of the improved conditional generative adversarial network, the insulator segmentation precision and efficiency are improved.

Description

Insulator segmentation method based on improved conditional generation confrontation network Technical field

The invention belongs to the field of image segmentation and high-voltage transmission line inspection, and in particular relates to an insulator segmentation method, system and device based on an improved condition generation counter network.

Background technique

Insulators are widely used in power systems. Once damaged, they will cause the power network to collapse and cause serious power economic losses. Therefore, detecting insulators has become an indispensable task in the power inspection process. With the rapid development of robots and unmanned aerial vehicles and the improvement of image detection technology, dangerous and complicated manual power inspections are gradually replaced by machines. In recent years, with the continuous development of artificial intelligence neural networks, power inspection based on deep learning has become the focus of researchers in recent years. How to use deep learning to accurately segment and identify insulators has become an important research direction in current line inspection.

Image segmentation is an important research topic in computer vision. It mainly studies the task of assigning a label to each pixel in the image, and recognizes the target at the pixel level. Image segmentation methods can be divided into three main categories. The first is based on traditional methods, such as "Normalized cut" and "Grab cut". This type of method mainly uses pixel-level bottom-level information for segmentation. The overall algorithm has low complexity, does not require training, and the segmentation efficiency is relatively high. However, in the face of an image with a slightly complicated background, auxiliary information needs to be added to help its segmentation, otherwise the effect is not ideal.

Another very important method is the deep learning method. Currently, deep learning image segmentation methods mainly include full convolutional networks, autoencoder networks, and generative adversarial networks (GAN). The full convolutional network FCN uses the deconvolution layer to upsample the feature map output by the last convolution layer to obtain an image with the same size as the input, and complete pixel-level segmentation on this image. FCN uses a fully convolutional network and has become a classic algorithm in image segmentation algorithms. The self-encoder method mainly uses an encoder to extract image features, obtain a feature map, and use a decoder to divide the feature map at the pixel level. This method is more complicated and time-consuming than a full convolutional network. The image segmentation completed by the generative confrontation network is generally used by the generator to map the extracted low-dimensional features into the segmentation model of the object. The discriminator network is generally symmetrical with the generator network to distinguish whether the predicted segmentation model is true. Through continuous training, a high-quality generator model is obtained, and the trained generator model is used to complete image segmentation. This method has higher segmentation accuracy and can be better improved in efficiency. However, the accuracy and efficiency of segmentation are still poor for insulator images with complex image backgrounds and various postures.

Therefore, this patent proposes an insulator segmentation method based on an improved conditional generation adversarial network, which has greater advantages over existing methods in terms of image segmentation accuracy and efficiency.

Summary of the invention

In order to solve the above-mentioned problems in the prior art, that is, to solve the problems of poor segmentation accuracy and low efficiency of the existing insulator segmentation method based on the conditional generation confrontation network, the first aspect of the present invention proposes an improved conditional generation confrontation network The method for dividing the insulator includes:

Step S100: Obtain an image containing an insulator as an input image;

Step S200, based on the input image, obtain an insulator segmentation image through an insulator segmentation model;

The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.

In some preferred embodiments, the asymmetric convolution layer of the encoder is composed of a convolution function, a batch normalization function, and a linear rectification function; the asymmetric convolution layer of the decoder is composed of a deconvolution function, Batch normalization function and linear rectification function are formed.

In some preferred embodiments, the training method of the insulator segmentation model is:

Step A100: Obtain an image containing insulators, and construct a sample set by a preset image enhancement method, the sample set including input image samples and real segmentation images of the insulators contained therein; split the sample set into training samples Set and test sample set;

Step A200: Obtain an insulator segmentation image through an insulator segmentation model based on the input image samples in the training sample set; use it as an insulator to generate a segmentation image;

Step A300: Generate a segmented image according to the insulator and the real segmented image of the insulator corresponding to the training sample, obtain the segmentation results of each region in the insulator segmented image through the conditional generation against the network cGAN discriminator, and obtain the loss value of the insulator segmentation model ；

Step A400: Obtain the current number of iterations. If the loss value is less than the preset training loss threshold or the number of iterations is greater than the preset number of training iterations, output the trained insulator segmentation model and use it as the first model. Go to step A500; otherwise, based on the loss value, use the backpropagation algorithm to update the parameters of the insulator segmentation model, increase the number of iterations by 1, and skip to step A200;

Step A500: Obtain insulator segmentation images of all input image samples in the test sample set through the first model, and compare the insulator segmentation images with the real segmentation images of the insulators contained in the test sample set to obtain mIoU The assessed value;

Step A600, if the mIoU evaluation value is greater than the preset evaluation value, use the first model as the finally trained insulator segmentation model; otherwise, skip to step A200.

In some preferred embodiments, the method of "constructing a sample set through a preset image enhancement method" in step A100 is as follows:

Obtain an image containing insulators as a pre-processed image sample;

Based on a preset set of brightness multiples, randomly select a brightness multiple to perform brightness processing on the preprocessed image sample to obtain a brightness processed image sample;

Rotate the preprocessed image samples to obtain multiple rotated processed image samples;

The brightness processed image sample and the rotation processed image sample are scaled to a preset size; and a sample set is constructed based on the scaled image.

In some preferred embodiments, the discriminator of the conditional generation against network cGAN is composed of five layers of convolutional layers; the first layer of convolutional layer is composed of convolution function and Leaky ReLU function, and the last layer is composed of convolution function , The remaining three convolutional layers are composed of convolution function, Leaky ReLU function, and batch normalization function.

In some preferred embodiments, the output of the discriminator of the conditional generation against network cGAN is a 16×16 matrix.

In the second aspect of the present invention, an insulator segmentation system based on an improved conditional generation confrontation network is proposed. The system includes an acquisition module and an output module;

The acquisition module is configured to acquire an image containing an insulator as an input image;

The output module is configured to obtain an insulator segmentation image through an insulator segmentation model based on the input image;

The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.

In a third aspect of the present invention, a storage device is provided, in which a plurality of programs are stored, and the program applications are loaded and executed by a processor to realize the above-mentioned insulator segmentation method based on the improved conditional generation confrontation network.

In the fourth aspect of the present invention, a processing device is proposed, including a processor and a storage device; the processor is suitable for executing each program; the storage device is suitable for storing multiple programs; the program is suitable for being loaded by the processor And execute to realize the above-mentioned insulator segmentation method based on the improved condition to generate the confrontation network.

The beneficial effects of the present invention:

The invention generates a confrontation network through an improved condition, which improves the accuracy and efficiency of the insulator segmentation. The present invention forms a self-encoder network by constructing encoders and decoders containing asymmetric convolutional layers, and uses them as an improved conditional generation counter-network cGAN generator, which reduces the calculation amount of insulator segmentation and improves segmentation efficiency.

At the same time, the present invention changes the output of the discriminator of the conditional generation against network cGAN to a 16×16 matrix, which can discriminate the segmentation results of each region of the insulator segmented image output by the generator, and update the generator (insulator segmentation model) based on the result output by the discriminator ), compared with the 0 and 1 values output by the existing discriminator, the trained insulator model has higher accuracy, and achieves high-quality segmentation in insulator images with complex backgrounds and diverse types and poses.

Description of the drawings

By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, purposes, and advantages of the present application will become more apparent.

FIG. 1 is a schematic flowchart of an insulator segmentation method based on an improved conditional generation confrontation network according to an embodiment of the present invention;

2 is a schematic diagram of an insulator segmentation system based on an improved conditional generation confrontation network according to an embodiment of the present invention;

FIG. 3 is a schematic flowchart of a training method of an insulator model according to an embodiment of the present invention;

4 is an example diagram of an improved conditional generation confrontation network according to an embodiment of the present invention;

FIG. 5 is an example diagram of comparison of detection results of different network models according to an embodiment of the present invention; FIG.

Fig. 6 is an exemplary diagram of the detection result of the present invention according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not All examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

The application will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described here are only used to explain the related invention, but not to limit the invention. In addition, it should be noted that, for ease of description, only the parts related to the relevant invention are shown in the drawings.

It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other if there is no conflict.

The insulator segmentation method based on the improved condition generation confrontation network of the present invention, as shown in Fig. 1, includes the following steps:

Step S100: Obtain an image containing an insulator as an input image;

Step S200, based on the input image, obtain an insulator segmentation image through an insulator segmentation model;

The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.

In order to more clearly describe the insulator segmentation method of the present invention based on the improved condition generation countermeasure network, the steps in an embodiment of the method of the present invention will be described in detail below with reference to the accompanying drawings.

In the following preferred embodiments, the training method of the insulator segmentation model will be described in detail first, and the insulator segmentation method based on the improved condition generation adversarial network will be described in detail.

1. Training method of insulator segmentation model

The insulator segmentation model, as shown in Figure 3, includes the following steps:

Step A100: Obtain an image containing insulators, and construct a sample set by a preset image enhancement method, the sample set including input image samples and real segmentation images of the insulators contained therein; split the sample set into training samples Set and test sample set;

Step A200: Obtain an insulator segmentation image through an insulator segmentation model based on the input image samples in the training sample set; use it as an insulator to generate a segmentation image;

Step A300: Generate a segmented image according to the insulator and the real segmented image of the insulator corresponding to the training sample, obtain the segmentation results of each region in the insulator segmented image through the conditional generation against the network cGAN discriminator, and obtain the loss value of the insulator segmentation model ；

Step A400: Obtain the current number of iterations. If the loss value is less than the preset training loss threshold or the number of iterations is greater than the preset number of training iterations, output the trained insulator segmentation model and use it as the first model. Go to step A500; otherwise, based on the loss value, use the backpropagation algorithm to update the parameters of the insulator segmentation model, increase the number of iterations by 1, and skip to step A200;

Step A500: Obtain insulator segmentation images of all input image samples in the test sample set through the first model, and compare the insulator segmentation images with the real segmentation images of the insulators contained in the test sample set to obtain mIoU The assessed value;

Step A600, if the mIoU evaluation value is greater than the preset evaluation value, use the first model as the finally trained insulator segmentation model; otherwise, skip to step A200.

The following is a detailed expansion of the above training steps, as follows:

Step A100: Obtain an image containing insulators, and construct a sample set by a preset image enhancement method, the sample set including input image samples and real segmentation images of the insulators contained therein; split the sample set into training samples Set and test sample set.

Step A101: Obtain an image containing the insulator as the original image.

Step A102, in order to enrich the samples, the present invention performs brightness processing and rotation processing on the original image.

The brightness processing method is: randomly select two values in the range of 0.5 to 1.5 times the brightness of the original image to process the brightness of the original image.

Rotation processing: Rotate the original image once every 24°, and each image can get 15 enhanced images. The present invention preferably rotates once every 24°. In other preferred embodiments, it can be set according to actual conditions.

In the present invention, the above image enhancement method is used to process 350 insulator images to obtain 6000 enhanced images.

In step A103, the enhanced image is scaled to a size of 256×256, and named according to 1.jpg to 6000.jpg, and the sample set image is annotated to obtain the insulator segmentation map, and the production of the image sample set is completed.

The sample set includes input image samples and real segmented images of the insulators contained therein.

The sample set is divided into a training sample set and a test sample set, where 5000 sheets are used as the training sample set and 1000 sheets are used as the test sample set.

Step A200: Obtain an insulator segmented image through an insulator segmentation model based on the input image samples in the training sample set; use it as an insulator to generate a segmented image.

In this embodiment, the conditional generation confrontation network cGAN is composed of a generator and a discriminator, as shown in FIG. 4. The insulator segmentation model is constructed based on the generator of the conditional generation confrontation network cGAN. The generator is a self-encoder, including an encoder and a decoder, that is, the encoding part and the decoding part. Save the sampling index of the coding part and pass it to the up-sampling layer, thereby reducing information loss. The real segmented image of the insulator and its corresponding image sample, the segmented image of the insulator output by the generator and its corresponding input image sample are regarded as a true and false image pair. Based on the insulator segmentation image output by the generator in the true and false image pair and the real segmentation image of the insulator, as input, a 16×16 matrix is output through the discriminator, and each element of the matrix represents whether the corresponding patch (region) is true or not false.

Each encoder in the generator includes an asymmetric convolutional layer and a maximum pooling layer. The asymmetric convolution layer is composed of a convolution function, a batch normalization function, and a linear rectification function, that is, the structure of Conv+BN+ReLU, which can deepen the network complexity while reducing network parameters. The decoder includes an asymmetric convolution layer and an up-sampling layer; the decoder and the encoder are symmetrical, using a deconvolution function, a batch normalization function, and a linear rectification function, that is, the structure of Conv+BN+ReLU. As shown in Table 1:

Table 1

Set	Layer name	Type of layers	Output size
Input	RGB image	To	256×256×3
Encoder1	Conv1	Conv+BN+ReLU,fs=(3,1)	256×256×64
To	Conv2	Conv+BN+ReLU,fs=(1,3)	256×256×64
To	MP1	Max-pooling(window 2×2)	128×128×64
Encoder2	Conv3	Conv+BN+ReLU,fs=(3,1)	128×128×128
To	Conv4	Conv+BN+ReLU,fs=(1,3)	128×128×128
To	MP2	Max-pooling(window 2×2)	64×64×128
Encoder3	Conv5	Conv+BN+ReLU,fs=(3,1)	64×64×256
To	Conv6	Conv+BN+ReLU,fs=(1,3)	64×64×256
To	Conv7	Conv+BN+ReLU,fs=(3,3)	64×64×256
To	MP3	Max-pooling(window 2×2)	32×32×256
Encoder4	Conv8	Conv+BN+ReLU,fs=(3,1)	32×32×512

To	Conv9	Conv+BN+ReLU,fs=(1,3)	32×32×512
To	Conv10	Conv+BN+ReLU,fs=(3,3)	32×32×512
To	MP4	Max-pooling(window 2×2)	16×16×512
Encoder5	Conv11	Conv+BN+ReLU,fs=(3,1)	16×16×512
To	Conv12	Conv+BN+ReLU,fs=(1,3)	16×16×512
To	Conv13	Conv+BN+ReLU,fs=(3,3)	16×16×512
To	MP5	Max-pooling(window 2×2)	8×8×512
Decoder1	UP1	UpSampling	16×16×512
To	Deconv1	Deconv+BN+ReLU,fs=(3,1)	16×16×512
To	Deconv2	Deconv+BN+ReLU,fs=(1,3)	16×16×512
To	Deconv3	Deconv+BN+ReLU,fs=(3,3)	16×16×512
Decoder2	UP2	UpSampling	32×32×512
To	Deconv4	Deconv+BN+ReLU,fs=(3,1)	32×32×512
To	Deconv5	Deconv+BN+ReLU,fs=(1,3)	32×32×512
To	Deconv6	Deconv+BN+ReLU,fs=(3,3)	32×32×256
Decoder3	UP3	UpSampling	64×64×256
To	Deconv7	Deconv+BN+ReLU,fs=(3,1)	64×64×256
To	Deconv8	Deconv+BN+ReLU,fs=(1,3)	64×64×256
To	Deconv9	Deconv+BN+ReLU,fs=(3,3)	64×64×128
Decoder4	UP4	UpSampling	128×128×128
To	Deconv10	Deconv+BN+ReLU,fs=(3,1)	128×128×128
To	Deconv11	Deconv+BN+ReLU,fs=(1,3)	128×128×64
Decoder5	UP5	UpSampling	256×256×64
To	Deconv12	Deconv+BN+ReLU,fs=(3,3)	256×256×64
To	Deconv13	Deconv+tanh,fs=(4,4)	256×256×3

In Table 1, Set is the frame, including Input, Encoder1-Encoder5 (encoder), Decoder1-Decoder5 (decoder), Layer name represents the name of each layer in the generator, RGB image is the input image, Conv represents the volume Multiplying layer, MP means maximum pooling layer, Deconv means deconvolution layer, UP means upsampling layer, Type of layers means layer type, fs means convolution kernel, window means window size, Max-pooling means maximum pooling, UpSampling For upsampling, Output size represents the output size.

The input image samples in the training sample set are input to the generator, and the resolution of the input image samples is 256×256×3, and the insulator segmentation image is obtained; and the segmentation image is generated as the insulator.

When training the insulator segmentation model, initialize the various parameters of the conditional generation against network cGAN. The present invention sets the batch sent to the training model at one time to 8, the initial learning rate is set to 0.0001, and the optimizer parameter is β ₁ = 0.9, β ₂ =0.99, and the maximum number of training iterations is set to 50000.

Step A300: Generate a segmented image according to the insulator and the real segmented image of the insulator corresponding to the training sample, obtain the segmentation results of each region in the insulator segmented image through the conditional generation against the network cGAN discriminator, and obtain the loss value of the insulator segmentation model .

In this embodiment, the discriminator of the conditional generation adversarial network cGAN is mainly composed of five-layer encoders, that is, five-layer convolutional layers. The first convolutional layer is composed of convolutional functions and Leaky ReLU functions; and the last layer is composed of convolutional functions and Leaky ReLU functions. The other three convolutional layers are composed of convolution function, Leaky ReLU function, and batch normalization function, namely Conv+Leaky ReLU+BN. The convolution kernel adopts 4×4, and the step size is 2. As shown in table 2:

Table 2

In Table 2, RGB image and generated image are the generated image of the generator and the real segmentation image of the insulator corresponding to the training sample.

The input of the discriminator is the generated image of the generator and the real segmentation image of the insulator corresponding to the training sample, and the output is a 16×16 matrix. Each element of the matrix represents whether the corresponding patch is true or false.

According to the insulator, the segmented image and the real segmented image of the insulator corresponding to the training sample are generated, and the segmentation results of each region in the insulator segmented image are obtained through the conditional generation of the discriminator against the network cGAN. Let the input samples conduct feedforward conduction in the network, and obtain the training error after comparing the generated segmentation map with the real segmentation map. According to the current network parameter values, the generator and discriminator continue to operate on the read training samples until the network output Training loss value of generator and discriminator.

Step A400: Obtain the current number of iterations. If the loss value is less than the preset training loss threshold or the number of iterations is greater than the preset number of training iterations, output the trained insulator segmentation model and use it as the first model. Go to step A500; otherwise, based on the loss value, use the back propagation algorithm to update the parameters of the insulator segmentation model, increase the number of iterations by 1, and skip to step A200.

In this embodiment, according to the current number of iterations or the loss value threshold, it is judged whether the training is continued or terminated. If it is terminated, the trained insulator segmentation model is obtained; otherwise, based on the loss value, the conditional generation is obtained through backpropagation. Fight against the change of the network cGAN, update the parameters, increase the number of iterations by 1, and jump to step A200.

Step A500: Obtain insulator segmentation images of all input image samples in the test sample set through the first model, and compare the insulator segmentation images with the real segmentation images of the insulators contained in the test sample set to obtain mIoU The assessed value.

In this embodiment, based on the insulator segmentation model trained in step S400, the model is tested through a test sample set. That is to say, all the input image samples in the test sample set generate insulator segmentation images, which are compared with the real segmentation images of the insulators contained in the test sample set to obtain the mIoU evaluation value and calculate the average test time.

Step A600, if the mIoU evaluation value is greater than the preset evaluation value, use the first model as the finally trained insulator segmentation model; otherwise, skip to step A200.

Generally speaking, the higher the mIoU evaluation value, the better the average segmentation effect. Therefore, if the obtained mIoU value is greater than the preset evaluation value, it means that the current training model has a better effect, otherwise it will be retrained.

In order to evaluate the insulator segmentation effect of the improved conditional generation adversarial network, the present invention compares this network with other classic network models. The experimental results are shown in Table 3:

table 3

Among them, Models in Table 3 represents the network model used in the experiment, Ours is the improved network extracted by the present invention, that is, the network in Figures 5 and 6, Pix2pix, SegNet, Unet, and FCN are pixel-to-pixel models, semantic segmentation networks, Deep learning segmentation network, full convolutional neural network, Trainable Para(M) represents the amount of training parameters, and Time represents the average test time. It can be seen from Table 3 that the mIoU evaluation value of the segmentation in the present invention is the highest, indicating the average segmentation effect the best. At the same time, the network requires the least parameters, which greatly reduces network complexity and improves segmentation efficiency. Figure 5 shows the final segmentation result. It can be seen that because the discriminator is added to the network of the present invention, the segmentation accuracy of the insulator can be improved, and pixel-level segmentation can also be completed in minute details.

In order to verify the ability of this network to segment insulators of different scales, we selected images with more complex backgrounds and background objects much larger than the size of the insulators as the test objects, and conducted an insulator segmentation experiment. As shown in Figure 6, even if the background is very complex and the object is larger than the size of the insulator, the network can still accurately identify the position of the insulator and segment it with high precision. It can be seen that this network solves the difficult problem of insulator detection in a complex environment.

2. Insulator segmentation method based on improved conditional generation confrontation network

In step S100, an image containing an insulator is obtained as an input image.

In this embodiment, an image containing an insulator actually acquired is used as the input image. The image containing the insulator can be taken manually, or obtained by aerial photography or other means.

Step S200, based on the input image, obtain an insulator segmentation image through an insulator segmentation model.

In this embodiment, based on the acquired image containing the insulator, the insulator segmentation image is acquired through the trained insulator segmentation model.

According to the second embodiment of the present invention, an insulator segmentation system based on an improved condition generation confrontation network, as shown in FIG. 2, includes: an acquisition module 100 and an output module 200;

The acquiring module 100 is configured to acquire an image containing an insulator as an input image;

The output module 200 is configured to obtain an insulator segmentation image through an insulator segmentation model based on the input image;

The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.

Those skilled in the technical field can clearly understand that, for the convenience and conciseness of description, the specific working process and related description of the system described above can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

It should be noted that the insulator segmentation system based on the improved condition to generate the confrontation network provided by the above embodiment only uses the division of the above functional modules as an example. In practical applications, the above function can be assigned to different functions according to needs. Functional modules are implemented, that is, the modules or steps in the embodiments of the present invention are further decomposed or combined. For example, the modules of the above-mentioned embodiments can be combined into one module, or further divided into multiple sub-modules to complete all or the steps described above. Part of the function. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing each module or step, and are not regarded as improper limitations on the present invention.

In a storage device according to a third embodiment of the present invention, a plurality of programs are stored therein, and the programs are suitable for being loaded by a processor and implementing the above-mentioned insulator segmentation method based on the improved condition generation confrontation network.

A processing device according to a fourth embodiment of the present invention includes a processor and a storage device; the processor is suitable for executing each program; the storage device is suitable for storing multiple programs; the program is suitable for being loaded and executed by the processor In order to realize the above-mentioned insulator segmentation method based on the improved condition to generate the confrontation network.

Those skilled in the technical field can clearly understand that what is not described is convenient and concise. For the specific working process and related description of the storage device and processing device described above, you can refer to the corresponding process in the foregoing method example, and will not be repeated here. Go into details.

Those skilled in the art should be able to realize that the modules and method steps of the examples described in the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two, and the software modules and method steps correspond to the program Can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or known in the technical field Any other form of storage medium. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in accordance with the function in the above description. Whether these functions are performed by electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered as going beyond the scope of the present invention.

The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or indicate a specific order or sequence.

The term "including" or any other similar term is intended to cover non-exclusive inclusion, so that a process, method, article or device/device including a series of elements includes not only those elements, but also other elements not explicitly listed, or It also includes the inherent elements of these processes, methods, articles, or equipment/devices.

So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims (9)

1. An insulator segmentation method based on an improved conditional generation confrontation network is characterized in that the method includes the following steps:

   Step S100: Obtain an image containing an insulator as an input image;

   Step S200, based on the input image, obtain an insulator segmentation image through an insulator segmentation model;

   The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.
2. The insulator segmentation method based on the improved conditional generation confrontation network according to claim 1, wherein the asymmetric convolution layer of the encoder is composed of a convolution function, a batch normalization function, and a linear rectification function; The asymmetric convolutional layer of the decoder is composed of a deconvolution function, a batch normalization function, and a linear rectification function.
3. The insulator segmentation method based on the improved conditional generation adversarial network according to claim 1, wherein the training method of the insulator segmentation model is:

   Step A100: Obtain an image containing insulators, and construct a sample set by a preset image enhancement method, the sample set including input image samples and real segmentation images of the insulators contained therein; split the sample set into training samples Set and test sample set;

   Step A200: Obtain an insulator segmentation image through an insulator segmentation model based on the input image samples in the training sample set; use it as an insulator to generate a segmentation image;

   Step A300: Generate a segmented image according to the insulator and the real segmented image of the insulator corresponding to the training sample, obtain the segmentation results of each region in the insulator segmented image through the conditional generation against the network cGAN discriminator, and obtain the loss value of the insulator segmentation model ；

   Step A400: Obtain the current number of iterations. If the loss value is less than the preset training loss threshold or the number of iterations is greater than the preset number of training iterations, output the trained insulator segmentation model and use it as the first model. Go to step A500; otherwise, based on the loss value, use the backpropagation algorithm to update the parameters of the insulator segmentation model, increase the number of iterations by 1, and skip to step A200;

   Step A500: Obtain insulator segmentation images of all input image samples in the test sample set through the first model, and compare the insulator segmentation images with the real segmentation images of the insulators contained in the test sample set to obtain mIoU The assessed value;

   Step A600, if the mIoU evaluation value is greater than the preset evaluation value, use the first model as the finally trained insulator segmentation model; otherwise, skip to step A200.
4. The insulator segmentation method based on the improved conditional generation confrontation network according to claim 3, characterized in that, in step A100, the method of "constructing a sample set through a preset image enhancement method" is:

   Obtain an image containing insulators as a pre-processed image sample;

   Based on a preset set of brightness multiples, randomly select a brightness multiple to perform brightness processing on the preprocessed image sample to obtain a brightness processed image sample;

   Rotate the preprocessed image samples to obtain multiple rotated processed image samples;

   The brightness processed image sample and the rotation processed image sample are scaled to a preset size; and a sample set is constructed based on the scaled image.
5. The insulator segmentation method based on the improved conditional generation confrontation network according to claim 3, wherein the discriminator of the conditional generation confrontation network cGAN is composed of five layers of convolutional layers; the first layer of convolutional layer is composed of convolutional layers. Function, Leaky ReLU function, the last layer is composed of convolution function, and the remaining three convolution layers are composed of Convolution function, Leaky ReLU function, and batch normalization function.
6. The insulator segmentation method based on an improved conditional generation confrontation network according to claim 5, wherein the discriminator of the conditional generation confrontation network cGAN has an output of a 16×16 matrix.
7. An insulator segmentation system based on an improved conditional generation confrontation network, which is characterized in that the system includes an acquisition module and an output module;

   The acquisition module is configured to acquire an image containing an insulator as an input image;

   The output module is configured to obtain an insulator segmentation image through an insulator segmentation model based on the input image;

   The insulator segmentation model is constructed based on a generator of the conditional generation confrontation network cGAN; the generator is constructed based on a self-encoder, which includes an encoder and a decoder; the encoder includes an asymmetric convolutional layer and a maximum pooling layer; The decoder includes an asymmetric convolutional layer and an up-sampling layer; the training samples of the insulator segmentation model include input image samples and real segmented images of the insulators contained therein.
8. A storage device, wherein a plurality of programs are stored, wherein the program application is loaded and executed by a processor to realize the insulator segmentation method based on the improved condition generation confrontation network according to any one of claims 1-6 .
9. A processing device, including a processor and a storage device; a processor, suitable for executing each program; a storage device, suitable for storing multiple programs; characterized in that the program is suitable for being loaded and executed by the processor to realize rights The insulator segmentation method based on the improved condition generation counter network described in any one of claims 1-6.

Images