CN111242241A - A method of augmenting training samples for etched character recognition networks - Google Patents

Abstract

The invention discloses an etched character recognition network training sample augmentation method, which belongs to the field of image processing technology and deep learning. The method includes the following steps: collecting etched character images in a scene; generating content images and style images according to the etched character images; constructing a bidirectional generative adversarial network; training a bidirectional generative adversarial network; Bidirectional Generative Adversarial Network to generate etched character images. The invention generates a large number of etched character images by generating an adversarial network, and sufficient training samples can still be obtained in the case of a small sample size. Compared with manual sample collection, the invention is faster and more efficient, and the generated etched character images are more realistic and improve the performance. Accuracy of Recognizing Etched Characters Using Deep Learning Methods.

Description

A method of augmenting training samples for etched character recognition networks

technical field

The invention belongs to the field of image processing technology and deep learning, and particularly relates to a method for augmenting training samples of an etched character recognition network.

Background technique

Etched character recognition is commonly used in text recognition in industrial equipment signage, and is one of the difficulties in scene text recognition. Industrial equipment signs are usually made of metal and are partially placed in outdoor environments, so there are often degradations such as reflections, stains, blurs, and scratches in sign images, which bring many difficulties to the recognition of etched characters.

Using deep learning to recognize etched characters requires a large amount of data to train a character recognition model to meet the generalization ability of the model, and the model trained when the sample size is small is prone to overfitting. In the research of etched character recognition, the number of etched character images that can be collected in a specific scene is small, and the lack of sample data is serious, which cannot meet the needs of deep learning. In addition, the collection and sorting of samples consumes a lot of manpower and material resources, and the efficiency of collecting and sorting samples manually is very low. Therefore, the use of deep learning methods to identify etched characters needs to solve the problem of small sample size. Commonly used image augmentation methods include flipping, rotating, scaling, cropping, shifting, adding noise, etc., but these methods all make a series of random changes to the existing samples, and can only generate images similar to the original samples.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image sample augmentation method for the problem of small scale of training samples of etched character recognition network, which can quickly generate a large batch of etched character images to meet the training requirements of deep learning network.

The technical solution for realizing the purpose of the present invention is: an etched character recognition network training sample augmentation method, the method comprises the following steps:

Step 1, collect the etched character image in the scene;

Step 2, generating a content image and a style image according to the etched character image;

Step 3, build a bidirectional generative adversarial network;

Step 4, training the bidirectional generative adversarial network;

Step 5: Input the content image and the style image into the trained bidirectional generative adversarial network to generate the etched character image.

Further, generating the content image according to the etched character image in step 2 specifically includes:

Step 2-1, marking the text information of the etched character image;

Step 2-2, count character information according to the real label of the marked etched character image;

Step 2-3, generating content images of multiple fonts according to the character information.

Further, generating the style image according to the etched character image in step 2 is specifically: generating the style image according to the characteristics of the etched character image.

Further, generating a style image according to the etched character image described in step 2 specifically includes:

From the collected etched character images, an image whose resolution meets the first preset condition and or whose definition meets the second preset condition and or whose feature significance meets the third preset condition is selected as a style image.

Further, building a bidirectional generative adversarial network described in step 3 specifically includes:

Step 3-1, build a stylized generative adversarial network;

Step 3-2, build a de-stylized generative adversarial network;

Step 3-3, construct the loss function.

Further, the stylized generative adversarial network described in step 3-1 includes: a stylized generative network and a stylized discriminant network;

The stylized generation network, whose input is a content image and a style image, and an output is a stylized character image; the stylized generation network includes: a content encoder E _x1 , whose input is a content image, and the output is a content feature vector; a style encoder E _x2 , the input is a style image, and the output is a style feature vector; the generator G _x , whose input is the content feature vector and the style feature vector, and the output is a stylized character image;

The input of the stylized discriminating network is the stylized character image or the real etched character image, and the output is a number between 0 and 1, which is used to represent the probability that the input image is a real image.

Further, the de-stylized generative adversarial network described in step 3-2 includes: a de-stylized generation network and a de-stylized discriminant network;

The de-stylized generation network, whose input is the stylized character image, and the output is the de-stylized character image; the de-stylized generation network includes: a first encoder E _y , whose input is the stylized character image, and the output is is a feature vector; the second generator G _y , its input is the feature vector output by the first encoder E _y , and the output is a de-stylized character image;

The de-stylized discriminant network has the input of the de-stylized character image or the real content image, and the output is a number between 0 and 1, which is used to represent the probability that the input image is a real image.

Further, the loss function L described in step 3-3 includes: content image reconstruction loss L ₁ , stylized generative adversarial network loss L ₂ and de-stylized generative adversarial network loss L ₃ , the formula is:

L=L ₁ +L ₂ +L ₃

The content image reconstruction loss L ₁ is used to ensure that the content encoder E _x1 can extract the core information of the content image, and the formula is:

In the formula, x represents the input content image, λ _x represents the weight of the content image reconstruction loss, and the value of λ _x ranges from 0 to 1;

The stylized generative adversarial network loss L ₂ includes the first pixel loss L _spix and the first adversarial loss L _sadv , the formula is:

L ₂ =λ _x1 L _spix +λ _x2 L _sadv

where L _spix represents the first pixel loss, L _sadv represents the first adversarial loss, λ _x1 and λ _x2 represent the weights of L _{spix and L sadv ,} respectively, and the value range of λ _x1 and λ _x2 is 0 to 1;

Among them, the calculation formula of the first pixel loss L _spix is as follows:

In the formula, x and y represent the input content image and style image, respectively, and y' represents the image generated by the stylized generation network;

The calculation formula of the first adversarial loss L _sadv is as follows:

定义为在风格图像y和风格化生成网络生成的图像y'之间沿直线均匀采样的向量，λ_sadv为取值范围为0到1的权重参数，D_x代表风格化判别网络；In the formula,

Defined as a vector uniformly sampled along a straight line between the style image y and the image y' generated by the stylized generation network, λ _sadv is a weight parameter ranging from 0 to 1, and D _x represents the stylized discriminant network;

The de-stylized generative adversarial network loss L ₃ includes the second pixel loss L _dpix , the second adversarial loss L _dfeat and the content feature loss L _dadv , the formula is:

L ₃ =λ _y1 L _dpix +λ _y2 L _dadv +λ _y3 L _dfeat

In the formula, λ _y1 , λ _y2 , and λ _y3 are the weights of L _dpix , L _dadv , and L _dfeat respectively, and the values range from 0 to 1;

Among them, the calculation formula of the second pixel loss L _dpix is as follows:

The calculation formula of the second adversarial loss L _dfeat is as follows:

The formula for calculating the content loss L _dadv is as follows:

Further, in step 4, the two-way generative adversarial network is trained, and the specific process includes:

Step 4-1, initialize the parameters and the number of iterations of the bidirectional generative adversarial network;

Step 4-2, input the content image to the content encoder of the stylized generative adversarial network, input the features output by the content encoder to the stylized generative network, calculate the loss function, and use the gradient descent method to update the parameters of the content encoder ;

Step 4-3, input the style image to the de-stylization generation network to generate a fake content image;

Steps 4-4, respectively input the real content image and the fake content image to the de-stylized discriminating network, calculate the loss function and update the parameters of the de-stylizing discriminant network by using the gradient descent method;

Step 4-5, input the fake content image to the de-stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the de-stylizing generation network;

Steps 4-6, input the content image and the style image into the stylization generation network to generate a fake style image;

Steps 4-7, respectively input the real style image and the fake style image to the stylized discriminant network, calculate the loss function and update the parameters of the stylized discriminant network by using the gradient descent method;

Steps 4-8, input the fake style image to the stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the stylized generation network;

Step 4-9, determine whether the current number of iterations is less than the set threshold, if so, repeat steps 4-2 to 4-8; otherwise, end the training of the bidirectional generative adversarial network.

Further, the content image and the style image are input into the trained bidirectional generative adversarial network described in step 5, and the etched character image is generated, which specifically includes:

Step 5-1, input the content image and the style image into the trained stylized generation network to generate the etched character image;

In step 5-2, the generated etched character images are screened, and images that do not meet the preset requirements are deleted.

Compared with the prior art, the present invention has the following significant advantages: 1) generating a large number of etched character images through a generative adversarial network, sufficient training samples can still be obtained in the case of a small sample size; 2) generating a large number of etched characters through a generative network Compared with manual sample collection, etched character images are faster and more efficient; 3) The use of bidirectional generative adversarial network can generate realistic character images, which improves the accuracy of using deep learning methods to identify etched characters.

The present invention will be described in further detail below with reference to the accompanying drawings.

Description of drawings

FIG. 1 is a flowchart of a method for augmenting training samples of an etched character recognition network in one embodiment.

FIG. 2 is a schematic diagram of a content image in one embodiment.

FIG. 3 is a schematic diagram of an etching style image in one embodiment.

FIG. 4 is a schematic diagram of a bidirectional generative adversarial network in one embodiment.

FIG. 5 is a flow chart of bidirectional generative adversarial network training in one embodiment.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

In one embodiment, with reference to FIG. 1, a method for augmenting training samples of an etched character recognition network is provided, and the method includes the following steps:

Step 1, collect the etched character image in the scene;

Step 2, generating a content image and a style image according to the etched character image;

Step 3, build a bidirectional generative adversarial network;

Step 4, train a bidirectional generative adversarial network;

Step 5: Input the content image and the style image into the trained bidirectional generative adversarial network to generate the etched character image.

Further, in one of the embodiments, in the above step 2, the content image is generated according to the etched character image, which specifically includes:

Step 2-1, marking the text information of the etched character image;

Step 2-2, count character information according to the real label of the marked etched character image;

Step 2-3, generating content images of multiple fonts according to the character information, as shown in FIG. 2 .

Here, a variety of fonts include Song type, Kai type, fit type, imitation Song type and so on.

Here, as a specific example, the text color of the content image is black, and the background color is white.

Further, in one of the embodiments, in the above step 2, the style image is generated according to the etched character image, specifically: generating the style image according to the etched character image feature.

Further, in one of the embodiments, in the above step 2, the style image is generated according to the etched character image, which specifically includes:

From the collected etched character images, an image whose resolution meets the first preset condition and or whose clarity meets the second preset condition and or whose feature significance meets the third preset condition is selected as a style image as shown in FIG. 3 .

Here, the content image and style image generated above have the same size specification.

Further, in one of the embodiments, building a bidirectional generative adversarial network in step 3 is shown in FIG. 4 , and the specific process includes:

Step 3-1, build a stylized generative adversarial network;

Step 3-2, build a de-stylized generative adversarial network;

Step 3-3, construct the loss function.

Further, in one of the embodiments, the stylized generative adversarial network in step 3-1 includes: a stylized generative network and a stylized discriminant network.

Among them, the input of the stylized generation network is the content image and the style image, and the output is the stylized character image.

Here, the input and output images are three-channel images of the same size.

The stylization generation network includes: a content encoder E _x1 , a style encoder E _x2 and a generator G _x .

The input of the content encoder E _x1 is the content image, and the output is the content feature vector. The content encoder first uses convolutional layers to extract features of the content image. Then use deconvolution layer upsampling to fuse the output features with the features output by the previous network layers. The activation layer is set before the convolution layer and the deconvolution layer, and the batch normalization layer is set after the convolution layer and the deconvolution layer.

The input of the style encoder E _x2 is the style image, and the output is the style feature vector. The style encoder first uses convolutional layers to extract the features of style images. Then use deconvolution layer upsampling to fuse the output features with the features output by the previous network layers. The activation layer is set before the convolution layer and the deconvolution layer, and the batch normalization layer is set after the convolution layer and the deconvolution layer.

The input of the generator G _x is the above-mentioned content feature vector and style feature vector, and the output is a stylized character image. The content feature vector and the style feature vector are of the same size. The generator first concatenates content feature vectors and style feature vectors, and then uses multiple deconvolution layers to upsample to generate stylized character images. The activation layer is set before the deconvolution layer, and the batch normalization layer is set after the deconvolution layer.

Among them, the input of the stylized discriminant network is a stylized character image or a real etched character image, and the output is a number between 0 and 1, which is used to represent the probability that the input image is a real image. The stylized discriminant network includes a convolutional layer, an activation layer is set before the convolutional layer, and a batch normalization layer is set after the convolutional layer.

Further, in one of the embodiments, the de-stylized generative adversarial network in step 3-2 includes: a de-stylized generative network and a de-stylized discriminant network.

Among them, the input of the de-stylized generation network is the stylized character image, and the output is the de-stylized character image.

The de-stylized generative network includes: a first encoder E _y and a second generator G _y .

The input of the first encoder E _y is a stylized character image, and the output is a feature vector. The encoder first uses convolutional layers to extract features from the content image. Then use deconvolution layer upsampling to fuse the output features with the features output by the previous network layers. The activation layer is set before the convolution layer and the deconvolution layer, and the batch normalization layer is set after the convolution layer and the deconvolution layer.

The input of the second generator G _y is the feature vector output by the first encoder E _y , and the output is a de-stylized character image. The generator uses multiple deconvolution layers to upsample to generate stylized character images. The activation layer is set before the deconvolution layer, and the batch normalization layer is set after the deconvolution layer.

Among them, the input of the de-stylized discriminant network is the de-stylized character image or the real content image, and the output is a number between 0 and 1, which is used to indicate the probability that the input image is a real image. The de-stylized discriminant network includes a convolutional layer, an activation layer is set before the convolutional layer, and a batch normalization layer is set after the convolutional layer.

Further, in one of the embodiments, the loss function L in step 3-3 includes: a content image reconstruction loss L ₁ , a stylized generative adversarial network loss L ₂ and a de-stylized generative adversarial network loss L ₃ , the formula is:

L=L ₁ +L ₂ +L ₃

The content image reconstruction loss L ₁ is used to ensure that the content encoder E _x1 can extract the core information (including character structure, stroke information, etc.) of the content image, and the formula is:

In the formula, x represents the input content image, λ _x represents the weight of the content image reconstruction loss, and the value of λ _x ranges from 0 to 1;

The stylized generative adversarial network loss L ₂ , including the first pixel loss L _spix and the first adversarial loss L _sadv , is formulated as:

L ₂ =λ _x1 L _spix +λ _x2 L _sadv

where L _spix represents the first pixel loss, L _sadv represents the first adversarial loss, λ _x1 and λ _x2 represent the weights of L _{spix and L sadv ,} respectively, and the value range of λ _x1 and λ _x2 is 0 to 1;

Among them, the calculation formula of the first pixel loss L _spix is as follows:

In the formula, x and y represent the input content image and style image, respectively, and y' represents the image generated by the stylized generation network;

The calculation formula of the first adversarial loss L _sadv is as follows:

定义为在风格图像y和风格化生成网络生成的图像y'之间沿直线均匀采样的向量，λ_sadv为取值范围为0到1的权重参数，D_x代表风格化判别网络；In the formula,

Defined as a vector uniformly sampled along a straight line between the style image y and the image y' generated by the stylized generation network, λ _sadv is a weight parameter ranging from 0 to 1, and D _x represents the stylized discriminant network;

The de-stylized generative adversarial network loss L ₃ includes the second pixel loss L _dpix , the second adversarial loss L _dfeat and the content feature loss L _dadv , the formula is:

L ₃ =λ _y1 L _dpix +λ _y2 L _dadv +λ _y3 L _dfeat

In the formula, λ _y1 , λ _y2 , and λ _y3 are the weights of L _dpix , L _dadv , and L _dfeat respectively, and the values range from 0 to 1;

Among them, the calculation formula of the second pixel loss L _dpix is as follows:

The calculation formula of the second adversarial loss L _dfeat is as follows:

The formula for calculating the content loss L _dadv is as follows:

Further, in one of the embodiments, the goal of training the bidirectional generative adversarial network in the above step 4 is to use the gradient descent method to minimize the value of the above-mentioned loss function L. The gradient descent method utilizes the characteristics of convex functions, and moves the parameters of the convex function in the opposite direction of the gradient by one step to realize the decrease of the function value. By repeating iteratively, the local minimum of the convex function is finally found. The update formula of parameter θ in this process is:

In the formula, θ represents the parameter, η represents the iteration step size, and J(θ) represents the loss function.

Combined with Figure 5, the specific process of training a bidirectional generative adversarial network includes:

Step 4-1, initialize the parameters and the number of iterations of the bidirectional generative adversarial network;

Step 4-2, input the content image to the content encoder of the stylized generative adversarial network, input the features output by the content encoder to the stylized generative network, calculate the loss function, and use the gradient descent method to update the parameters of the content encoder ;

Step 4-3, input the style image to the de-stylization generation network to generate a fake content image;

Steps 4-4, respectively input the real content image and the fake content image to the de-stylized discriminating network, calculate the loss function and update the parameters of the de-stylizing discriminant network by using the gradient descent method;

Step 4-5, input the fake content image to the de-stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the de-stylizing generation network;

Steps 4-6, input the content image and the style image into the stylization generation network to generate a fake style image;

Steps 4-7, respectively input the real style image and the fake style image to the stylized discriminant network, calculate the loss function and update the parameters of the stylized discriminant network by using the gradient descent method;

Steps 4-8, input the fake style image to the stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the stylized generation network;

Step 4-9, determine whether the current number of iterations is less than the set threshold, if so, repeat steps 4-2 to 4-8; otherwise, end the training of the bidirectional generative adversarial network.

Further, in one of the embodiments, the above-mentioned step 5 inputs the content image and the style image into the trained bidirectional generative adversarial network to generate the etched character image, which specifically includes:

Step 5-1, input the content image and the style image into the trained stylized generation network to generate the etched character image;

In step 5-2, the generated etched character images are screened, and images that do not meet the preset requirements are deleted.

To sum up, the present invention generates a large number of etched character images through a generative adversarial network, and sufficient training samples can still be obtained in the case of a small sample size. Compared with manual sample collection, the present invention is faster and more efficient, and the generated etched character images are more realistic. , which improves the accuracy of recognizing etched characters using deep learning methods.

Claims (10)

1. an etching character recognition network training sample augmentation method, is characterized in that, described method comprises the following steps: Step 1, collect the etched character image in the scene; Step 2, generating a content image and a style image according to the etched character image; Step 3, build a bidirectional generative adversarial network; Step 4, training the bidirectional generative adversarial network; Step 5: Input the content image and the style image into the trained bidirectional generative adversarial network to generate the etched character image. 2. The etched character recognition network training sample augmentation method according to claim 1, wherein the step 2 generates a content image according to the etched character image, specifically comprising: Step 2-1, marking the text information of the etched character image; Step 2-2, count character information according to the real label of the marked etched character image; Step 2-3, generating content images of multiple fonts according to the character information. 3. The etched character recognition network training sample augmentation method according to claim 1 and 2, wherein the step 2 generates a style image according to the etched character image, specifically: generating a style according to the etched character image feature image. 4. The etched character recognition network training sample augmentation method according to claim 3, wherein the step 2 generates a style image according to the etched character image, specifically comprising: From the collected etched character images, an image whose resolution meets the first preset condition and or whose definition meets the second preset condition and or whose feature significance meets the third preset condition is selected as a style image. 5. etched character recognition network training sample augmentation method according to claim 4, is characterized in that, described in step 3 constructs bidirectional generative confrontation network, specifically comprises: Step 3-1, build a stylized generative adversarial network; Step 3-2, build a de-stylized generative adversarial network; Step 3-3, construct the loss function. 6. The etched character recognition network training sample augmentation method according to claim 5, wherein the stylized generative adversarial network described in step 3-1 comprises: a stylized generative network and a stylized discrimination network; The stylized generation network, whose input is a content image and a style image, and an output is a stylized character image; the stylized generation network includes: a content encoder E _x1 , whose input is a content image, and the output is a content feature vector; a style encoder E _x2 , the input is a style image, and the output is a style feature vector; the generator G _x , whose input is the content feature vector and the style feature vector, and the output is a stylized character image; The input of the stylized discriminating network is the stylized character image or the real etched character image, and the output is a number between 0 and 1, which is used to represent the probability that the input image is a real image. 7. The etched character recognition network training sample augmentation method according to claim 6, wherein the de-stylized generative adversarial network described in step 3-2 comprises: a de-stylized generation network and a de-stylized discriminant network ; The de-stylized generation network, whose input is the stylized character image, and the output is the de-stylized character image; the de-stylized generation network includes: a first encoder E _y , whose input is the stylized character image, and the output is is a feature vector; the second generator G _y , its input is the feature vector output by the first encoder E _y , and the output is a de-stylized character image; The de-stylized discriminant network has the input of the de-stylized character image or the real content image, and the output is a number between 0 and 1, which is used to represent the probability that the input image is a real image. 8. The method for augmenting etched character recognition network training samples according to claim 7, wherein the loss function L in step 3-3 comprises: content image reconstruction loss L ₁ , stylized generative adversarial network loss L ₂ and de-stylized GAN loss L ₃ , formulated as: L=L ₁ +L ₂ +L ₃ The content image reconstruction loss L ₁ is used to ensure that the content encoder E _x1 can extract the core information of the content image, and the formula is:

In the formula, x represents the input content image, λ _x represents the weight of the content image reconstruction loss, and the value of λ _x ranges from 0 to 1; The stylized generative adversarial network loss L ₂ includes the first pixel loss L _spix and the first adversarial loss L _sadv , the formula is: L ₂ =λ _x1 L _spix +λ _x2 L _sadv where L _spix represents the first pixel loss, L _sadv represents the first adversarial loss, λ _x1 and λ _x2 represent the weights of L _{spix and L sadv ,} respectively, and the value range of λ _x1 and λ _x2 is 0 to 1; Among them, the calculation formula of the first pixel loss L _spix is as follows:

In the formula, x and y represent the input content image and style image, respectively, and y' represents the image generated by the stylized generation network; The calculation formula of the first adversarial loss L _sadv is as follows:

In the formula,

Defined as a vector uniformly sampled along a straight line between the style image y and the image y' generated by the stylized generation network, λ _sadv is a weight parameter ranging from 0 to 1, and D _x represents the stylized discriminant network; The de-stylized generative adversarial network loss L ₃ includes the second pixel loss L _dpix , the second adversarial loss L _dfeat and the content feature loss L _dadv , the formula is: L ₃ =λ _y1 L _dpix +λ _y2 L _dadv +λ _y3 L _dfeat In the formula, λ _y1 , λ _y2 , and λ _y3 are the weights of L _dpix , L _dadv , and L _dfeat respectively, and the values range from 0 to 1; Among them, the calculation formula of the second pixel loss L _dpix is as follows:

The calculation formula of the second adversarial loss L _dfeat is as follows:

The formula for calculating the content loss L _dadv is as follows:

9. The etched character recognition network training sample augmentation method according to claim 8, wherein in step 4, the two-way generative adversarial network is trained, and the specific process comprises: Step 4-1, initialize the parameters and the number of iterations of the bidirectional generative adversarial network; Step 4-2, input the content image to the content encoder of the stylized generative adversarial network, input the features output by the content encoder to the stylized generative network, calculate the loss function, and use the gradient descent method to update the parameters of the content encoder ; Step 4-3, input the style image to the de-stylization generation network to generate a fake content image; Steps 4-4, respectively input the real content image and the fake content image to the de-stylized discriminating network, calculate the loss function and update the parameters of the de-stylizing discriminant network by using the gradient descent method; Step 4-5, input the fake content image to the de-stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the de-stylizing generation network; Steps 4-6, input the content image and the style image into the stylization generation network to generate a fake style image; Steps 4-7, respectively input the real style image and the fake style image to the stylized discriminant network, calculate the loss function and update the parameters of the stylized discriminant network by using the gradient descent method; Steps 4-8, input the fake style image to the stylized discriminating network, calculate the loss function, and use the gradient descent method to update the network parameters of the stylized generation network; Step 4-9, determine whether the current number of iterations is less than the set threshold, if so, repeat steps 4-2 to 4-8; otherwise, end the training of the bidirectional generative adversarial network. 10. etched character recognition network training sample augmentation method according to claim 9, it is characterized in that, described in step 5, content image and style image are input into trained bidirectional generative adversarial network, generate etched character image, Specifically include: Step 5-1, input the content image and the style image into the trained stylized generation network to generate the etched character image; In step 5-2, the generated etched character images are screened, and images that do not meet the preset requirements are deleted.

Images