Intelligent image preprocessing method for rock debris recognition

Abstract

The invention discloses an intelligent image preprocessing method for rock debris recognition. The intelligent image preprocessing method comprises the following steps: (1) a two-path decoder image denoising method based on an attention mechanism; (2) carrying out an LPG-PCA image denoising algorithm based on vector similarity; (3) a low-illumination image enhancement method based on illumination map guidance; (4) enhancing the rock-soil rock debris data set through a Mosaic algorithm; and (5) generating an adversarial network (GAN) to expand the'actual measurement 'data set. According to the method disclosed by the invention, the precision and generalization ability of the identification model are improved. Meanwhile, by adopting an image expansion technology and an enhancement technology, the training set quality and the training efficiency of the recognition model can be obviously improved.

Description

Technical Field

[0001] The present invention belongs to the field of underground lithology identification, and particularly relates to an intelligent image preprocessing method for underground rock image identification in this field based on computer image processing technology. Background Art

[0002] Underground space lithology identification refers to determining the type, composition, and properties of underground rocks during underground exploration using various equipment and technical means. Usually, drilling equipment, seismic exploration equipment, electrical method equipment, radar equipment, etc. are used for underground material exploration.

[0003] With the gradual application of artificial intelligence science in diverse scenarios, many geologists have introduced artificial intelligence theories and technologies to study geological engineering problems in interdisciplinary fields, such as lithology intelligent recognition and analysis algorithms, which overcome the defects of traditional lithology recognition methods, improve the accuracy of recognition results, not only greatly reduce the workload of researchers, but also reduce the dependence on recognition equipment, and have begun to be applied in the field of geological research. Artificial intelligence deep learning mainly refers to an algorithm that processes information through a neural network to achieve feature extraction. In the field of lithology recognition, traditional machine learning methods require manual extraction of features such as the color, shape, and texture of rocks. For example, Singh et al. extracted features such as local texture information from basalt rock thin section images and achieved good classification of basalt rock samples. Zhou Yongzhang et al. summarized the core application technologies of big data and mathematical geosciences and proposed that machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Ma Zedong et al. proposed a multi-scale lithology recognition method based on big data deep learning of rock images, which can better identify rock lithology. Wang Jie et al. established a neural network lithology recognition model using the BP neural network learning algorithm based on machine learning technology and can better use drilling parameters for rock prediction. The Inception model is a convolutional neural network model proposed by Google in the 2014 ILSVRC competition. It uses multiple different convolutional kernels to extract different scale information of the target image and finally fuses them to obtain high-dimensional features of the image.

[0004] When performing cuttings image recognition, the obtained samples have the characteristics of many noise points and little effective information and need to be preprocessed before recognition. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent image preprocessing method for cuttings recognition.

[0006] The present invention adopts the following technical solution:

[0007] An intelligent image preprocessing method for cuttings recognition, the improvement lies in including the following content:

[0008] (1) Dual-path decoder image denoising method based on attention mechanism;

[0009] (2) LPG-PCA image denoising algorithm based on vector similarity;

[0010] (3) Low-light image enhancement method guided by illumination map;

[0011] (4) Mosaic algorithm to enhance the geotechnical rock cuttings dataset;

[0012] (5) Generative adversarial network GAN to augment the "measured" dataset.

[0013] Furthermore, the specific content of method (1) is as follows:

[0014] An encoder-decoder structure is adopted to extract and summarize image features, including an encoder and two decoders, and skip connections are used;

[0015] In the encoder part, a residual dense block RDB structure is used to design the convolutional layer. At the same time, the batch normalization layer after the convolutional layer is removed, the activation function is replaced with PReLU, and the method of strided convolution is also selected;

[0016] The first decoder is symmetric with the encoder part. The same residual dense block structure is adopted in the convolutional layer, and skip connections are made with the feature maps of the corresponding encoder layers. The polarization self-attention mechanism PSA is also introduced, and the transposed convolution with a stride of 2 is used for the upsampling operation;

[0017] The second decoder cascades and upsamples the output of each layer of the first decoder through two consecutive 3×3 convolutions, batch normalization, and ReLU function operations, and the output of the corresponding encoder layer.

[0018] Furthermore, in method (1), a dual-decoder image denoising network EDDNet based on attention mechanism and a decoder with different depths are used.

[0019] Furthermore, the specific content of method (2) is as follows:

[0020] The target pixel points and their neighborhoods in the noisy image are denoised as a whole, and the denoising model of the image is expressed as:

[0021] x v = x + v

[0022] In the above formula, x v is the noisy image, x is the original noise-free image, and v is the noise;

[0023] Regarding each image block as a vector, the vector form of the target pixel block is expressed as The vector form of the training sample block is expressed as Calculate the vector similarity between the target pixel block and each training sample block:

[0024]

[0025] In the above formula: is the norm of the vector, m is the total number of sample categories, is the included angle of the vector;

[0026] Sort according to the vector similarity from large to small, and select the training sample blocks corresponding to 70% of the data as the preliminarily screened similar sample blocks. Then, calculate the mean square error MSE between these sample blocks and the target pixel block to represent the similarity between the image blocks:

[0027]

[0028] In the above formula, σ 2 is the noise variance;

[0029] Select a threshold T. When the MSE between the noiseless pixel blocks is less than this threshold, that is:

[0030]

[0031] At this time, if the MSE between the noisy image blocks satisfies: e i ≤T + 2σ 2 , then it is considered that this training sample block is similar to the target pixel block.

[0032] Furthermore, in method (2), when the number n of image blocks that satisfy the condition e i ≤T + 2σ 2 is less than cK 2 , take (n + cK 2 / 2) sample numbers. When the number n of image blocks that satisfy the condition e i ≤T + 2σ 2 is greater than cK 2 , take n sample numbers. c takes 8 - 10, and K represents the side length of the local pixel block centered on the target pixel point.

[0033] Furthermore, the specific content of method (3) is:

[0034] First, calculate the irradiance guidance map for the input low-illumination image. Then, after splicing the low-illumination image and the guidance map, input the multi-scale image into U-Net for feature learning. Use the guidance map as an attention map and multiply it with the network output feature map to adjust the local brightness of the image. Finally, use a transitional convolutional layer to transfer the shallow detail features of the input image to the deep layer.

[0035] Further, the specific content of method (4) is as follows:

[0036] During the training process, before each epoch of training, use the Mosaic algorithm. Randomly select 4 images from the training set samples, then arrange them in 4 fixed positions of the new image. Finally, draw a cross-cutting line in the middle part of the new image and cut off the overlapping part of the image.

[0037] Further, in method (4), in the enhancement task, use the cross-entropy loss function to calculate the error between the true value and the predicted value. In the calculation of the cross-entropy loss function, use the one-hot encoding form to distinguish positive and negative samples, that is, label 1 is the positive sample and label 0 is the negative sample;

[0038] The cross-entropy loss function is expressed as:

[0039]

[0040] In the above formula, m is the number of samples, n is the number of categories, y i is the one-hot encoded label information, and p i is the probability matrix.

[0041] Further, the specific content of method (5) is as follows:

[0042] The generative adversarial network GAN consists of two networks: the generative network Generator is responsible for generating simulated data; the discriminative network Discriminator is responsible for judging whether the input data is real or generated. The generator G learns the real data distribution Px and maps out fake data G(z) similar to the real data distribution. The discriminator D learns the real data distribution Px and the generated fake data distribution PG(z) to identify whether the source of the input data is real data or generated fake data, that is, perform binary classification on x and G(z). The two are alternately trained in the network and jointly learn and progress in the confrontation until the Nash equilibrium state is reached.

[0043] The beneficial effects of the present invention are:

[0044] The method disclosed by the present invention improves the accuracy and generalization ability of the recognition model. At the same time, by adopting the image augmentation technology and enhancement technology, the quality and training efficiency of the recognition model training set can be significantly improved. Description of the Drawings

[0045] Figure 1 It is a flowchart for denoising cuttings image data;

[0046] Figure 2 It is a network architecture diagram of low-light image enhancement technology;

[0047] Figure 3 It is a structure diagram of the GAN framework;

[0048] Figure 4 It is a structure diagram of the ACGAN framework. Specific implementation manners

[0049] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in 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 invention and are not used to limit the present invention.

[0050] Embodiment 1. The intelligent image preprocessing method for cuttings recognition disclosed in this embodiment includes the following 5 parts in total:

[0051] (1) Dual-path decoder image denoising method based on attention mechanism:

[0052] Use traditional image processing methods to process the obtained cuttings image samples, and at the same time use a variety of processing methods for image processing. For images from different image sensors, corresponding image denoising and image enhancement methods need to be carried out respectively. In order to ensure that the image recognition system can obtain high-quality and feature-obvious training data, an encoder-decoder structure is used to extract and summarize image features. After verification, this method can greatly reduce image noise while retaining the detailed information required for cuttings classification. The specific process is as Figure 1 shown.

[0053] This network is an improvement based on the U-Net network to address various problems that may occur during the image denoising process. Since the U-Net reduces the image size during the downsampling process, directly fusing shallow information with corresponding deep information may lead to limited utilization of shallow features and further result in the loss of image detail information. To solve this problem, a dual-decoder image denoising network EDDNet based on the attention mechanism is used. A decoder with different depths is designed to achieve multi-scale extraction of features at different layers, further strengthening the efficient fusion of shallow and deep features during the downsampling process. At the same time, a residual dense block structure is designed. Through this design, the retention time of shallow detail information is extended to achieve the purpose of improving the image denoising effect. The model structure includes an encoder and two decoders (DecoderA and DecoderB), and multiple skip connections are adopted. In the encoder part, the encoder of ResUnet is improved, and the residual dense block RDB structure is used to design the convolutional layer. This not only increases the depth of the network but also maximally ensures the propagation of information between feature layers. At the same time, by removing the batch normalization (BN) layer after the convolutional layer and replacing the activation function with PReLU, the encoder part is further optimized. The downsampling operation reduces the image size while expanding the receptive field and reduces information loss by increasing the number of channels. Different from the traditional U-Net that uses the max pooling layer for downsampling, the method of strided convolution is selected to reduce information loss during the downsampling process. DecoderA of the decoder has basic symmetry with the encoder part. The same residual dense block structure is adopted in the convolutional layer, and skip connections are made with the feature maps of the corresponding encoding layers. At the same time, to enhance the fine-grained processing of the image, the polarization self-attention mechanism PSA is introduced in DecoderA. The upsampling operation uses a transposed convolution with a stride of 2. DecoderB of the decoder cascades and upsamples the output of each layer of DecoderA after passing through two consecutive 3×3 convolutions, batch normalization, and ReLU function operations with the output of the corresponding encoder layer. Such a design enables the model to fuse feature information at multiple levels and achieve a more accurate image denoising effect.

[0054] (2) LPG-PCA image denoising algorithm based on vector similarity:

[0055] This method denoises the target pixel points and their neighborhoods in the noisy image as a whole, speeds up the operation while retaining the local properties of the image, and enhances the image denoising effect. The image denoising model can be expressed as:

[0056] x v = x + v

[0057] Among them, x v is the noisy image, x is the original noise-free image, and v is the noise. In order to use the PCA method for image denoising, a large number of training samples of the noisy image x v are needed to calculate the covariance matrix of x v and the PCA transformation matrix. Taking the local pixel block of size K×K centered on the target pixel as the target pixel block, and selecting a training window of size L×L (L>K) centered on this target pixel block, theoretically (L-K+1)2 training sample blocks of size K×K can be obtained.

[0058] However, since there may be large differences in structure between some training sample blocks and the target sample block, if all training sample blocks are regarded as the final similar sample blocks, it will cause errors in the covariance matrix of xv obtained, and have a certain impact on the final denoising effect. Therefore, it is necessary to select sample blocks similar to the target pixel block from these training sample blocks.

[0059] First, each image block is regarded as a vector. For example, the vector form of the target pixel block can be expressed as The vector form of the training sample block can be expressed as Calculate the vector similarity between the target pixel block and each training sample block Among them:

[0060]

[0061]

[0062] is the norm of the vector, m is the total number of sample categories, is the included angle of the vectors.

[0063] The vector similarity comprehensively considers the norm and included angle information of the vectors, directly reflecting the similarity degree between the vectors. The more similar two vectors are, the greater the vector similarity. Sort the training sample blocks according to the vector similarity from large to small, and select the training sample blocks corresponding to 70% of the data as the initially screened similar sample blocks. Then calculate the mean square error MSE between these sample blocks and the target pixel block to represent the similarity between the image blocks, that is:

[0064]

[0065] where σ 2 is the noise variance. Select a suitable threshold T. When the MSE between the noise-free pixel blocks is less than this threshold, that is:

[0066]

[0067] At this time, the mean squared error (MSE) between the noisy image patches satisfies e i ≤T + 2σ 2 , and this training sample patch is considered similar to the target pixel patch. At the same time, to ensure there are enough samples, the traditional LPG-PCA algorithm obtains an empirical formula based on a large number of simulation experiments, that is, at least cK 2 similar sample patches are selected. K represents the side length of the local pixel patch centered on the target pixel point, with the unit of pixel, and c is usually taken as 8 - 10. However, in fact, during the selection process of similar sample patches, the number of image patches that satisfy e i ≤T + 2σ 2 may be far less than cK 2 . At this time, if cK 2 sample patches are still selected, it will cause certain errors and make some dissimilar image patches participate in the subsequent calculation of the covariance matrix. Therefore, an improvement is made in the number of samples of similar image patches. When the number n of image patches that satisfy e i ≤T + 2σ 2 is less than cK 2 , take (n + cK 2 ) / 2 sample patches. When the number n of image patches that satisfy e i ≤T + 2σ 2 is greater than cK 2 , take n sample patches, that is:

[0068]

[0069] An important property of PCA is to completely remove the correlation of the original data. In the PCA transform domain, the energy of the signal is mainly concentrated on a small subset of the PCA transform data set, while the energy of the noise is evenly distributed over the entire data set. Therefore, signals and noise can be well distinguished in the PCA domain.

[0070] (3) Low-light image enhancement method based on illumination map guidance:

[0071] Affected by uneven illumination, the brightness of different regions in low-light images will be different. Some existing methods perform consistent enhancement on all regions during brightness enhancement, resulting in problems such as over-enhancement or under-enhancement of the image. To address this issue, a low-light image enhancement network based on illumination map guidance is proposed, as shown in Figure 2As shown below. First, calculate the irradiance guidance map for the input low-light image to reflect the brightness of different regions in the image. Then, after splicing the low-light image and the guidance map, input the multi-scale image into U-Net for feature learning. To perform different degrees of brightness enhancement operations on regions with different brightness and darkness in the image, use the guidance map as an attention map and multiply it with the network output feature map to adjust the local brightness of the image. Finally, use a transitional convolutional layer to transfer the shallow detail features of the input image to the deep layer, realize the reuse of shallow features to enhance the features of the image, and obtain the image with enhanced brightness.

[0072] To make the overall brightness of the enhanced image uniform, an enhancement network is constructed using the idea of different enhancement strengths for regions with different brightnesses. In an image, the gray value corresponding to a brighter region is larger, and the gray value corresponding to a darker region is smaller. Therefore, during image enhancement, a relatively smaller enhancement strength is used for regions with larger gray values, and a larger enhancement strength is used for regions with smaller gray values. By observing the image, it can be found that the inverted image of the image is exactly opposite to the intensity of the gray value of the image, that is, it is consistent with the selected enhancement strength trend. Therefore, in the network, the guidance map can be used to guide the network to learn the enhancement strength for different regions, give a larger enhancement strength to relatively darker regions to solve the problem of insufficient enhancement, and weaken the enhancement strength for relatively brighter regions to avoid over-enhancement, making the overall image brightness balanced and more in line with human visual observation. Based on this, construct the irradiance guidance map using the inverse transformation of the low-light image to guide the feature learning of the network.

[0073] (4) Enhance the rock and soil cuttings dataset using the Mosaic algorithm:

[0074] In the experiment, the number of underground rock and soil cutting samples is small, the abnormal target morphology is variable, and there are many interferences in the rock and soil cutting image data. The detection effect of abnormal targets is greatly affected by the background. The biggest difference between the Mosaic algorithm and the augmenters library enhancement algorithm is that the latter expands the number of samples before model training, while the Mosaic algorithm performs data augmentation in real time during training. During each epoch of training, use the Mosaic algorithm. Randomly select 4 images from the training set samples, then arrange them in 4 fixed positions of the new image in a certain order. Finally, draw a cross cutting line in the middle part of the new image and cut off the overlapping part of the image. The Mosaic algorithm performs random data augmentation during training, so that new samples will be added to the training set during each epoch iteration, effectively solving the problem that a small number of training samples causes the model to fall into overfitting. At the same time, the Mosaic data augmentation algorithm splices 4 images with different backgrounds, enriching the background of the abnormal targets in the GPR images in the training set and improving the detection performance of the model.

[0075] In enhancement tasks, the cross-entropy loss function is often used to calculate the error between the true value and the predicted value, thereby driving the model closer to more accurate predictions. In the calculation of the cross-entropy loss function, one-hot encoding is used to distinguish positive and negative samples, that is, label 1 is a positive sample and label 0 is a negative sample.

[0076] The cross-entropy loss function can be expressed as:

[0077]

[0078] where m is the number of samples, n is the number of classes, y i is the one-hot label information, and p i is the probability matrix.

[0079] For example, for a batch of samples with a total of 5 classes, if there is an unsmoothed one-hot label [0, 0, 0, 1, 0], and the probability matrix p after passing through the network model is [0.1, 0.1, 0.1, 0.36, 0.34], substituting it into the cross-entropy loss function gives the loss of a single sample as 1.47. After calculation, the smoothed one-hot label is [0.02, 0.02, 0.02, 0.92, 0.02], and substituting it into the cross-entropy loss function gives the loss of a single sample as 2.63.

[0080]

[0081] It can be seen that the loss without label smoothing only considers the loss at the correct label position, resulting in the model overfitting to the abnormal target morphology in the underground geotechnical cuttings image training set during the training process, and a large number of abnormal targets with complex features in the test set cannot be correctly detected, that is, the model overfits and has poor generalization ability; the loss after smoothing considers both the loss at the correct label position and the loss at the wrong label position, increasing the loss value, enabling the model to reduce the misclassification of abnormal targets while improving the correct classification of abnormal targets, enhancing the learning ability of the model, improving the generalization ability of the model, and enabling the trained model to accurately identify the semantic information of cuttings in geotechnical cuttings images with variable backgrounds and complex morphologies.

[0082] Image sample expansion technology for small-scale image datasets:

[0083] To address the problem of insufficient training samples of underground space geotechnical cuttings, image data augmentation technology is used to generate more data samples. Data augmentation is a general term for methods of expanding data. Data augmentation can increase the samples in the training set, effectively alleviate the situation of model overfitting, and also bring stronger generalization ability to the model.

[0084] (5) Generative Adversarial Network (GAN) is used to expand the "measured" dataset:

[0085] A Generative Adversarial Network (GAN) consists of two networks: the Generator is responsible for generating simulated data; the Discriminator is responsible for judging whether the input data is real or generated. The generator G learns the real data distribution Px and maps out fake data G(z) similar to the real data distribution. The discriminator D needs to learn the real data distribution Px and the generated fake data distribution PG(z) to identify whether the source of the input data is real data or generated fake data, that is, to perform binary classification on x and G(z). The two are alternately trained in the network and jointly learn and progress in the confrontation until the Nash equilibrium state is finally reached.

[0086] The formation process of a GAN: Real-world problem requirements → Establish a GAN framework (programming) to implement functions → Train GAN (generator network, adversarial network) → Mature GAN model → Application.

[0087] ACGAN is used to generate multi-class enhanced data. As Figure 3 shown in Figure 4, the difference between ACGAN and GAN is that for the generator, there are two inputs, one is the classification data c of the label, and the other is the random data z, and the generated data is X fake = G(c, z); for the discriminator, it is necessary to judge the probability distribution P(S|X) of whether the data source is real data and the probability distribution P(C|X) of the data source for the classification label respectively. The objective function of ACGAN includes the following two parts

[0088] L S = E[logP(s = real|X real )] + E[logP(s = fake|X fake )]

[0089] L C = E[logP(C = c|X real )] + E[logP(C = c|X fake )]

[0090] Among them, L S is the loss function for whether the data is real or not, and L C is the loss function for the accuracy of data classification. In the optimization process, the discriminator D tries to maximize L C + L S , and the generator G tries to maximize L C - L S .

Claims

1. An intelligent image preprocessing method for cuttings recognition, characterized in that, It includes the following contents: (1) Dual-path decoder image denoising method based on attention mechanism; (2) LPG-PCA image denoising algorithm based on vector similarity; (3) Low-light image enhancement method guided by illumination map; (4) Mosaic algorithm to enhance the geotechnical rock debris dataset; (5) Generative adversarial network GAN to expand the "measured" dataset.

2. The intelligent image preprocessing method for cuttings recognition according to claim 1, wherein The specific content of method (1) is as follows: An encoder-decoder structure is adopted to extract and summarize image features, including an encoder and two decoders, and skip connections are used; In the encoder part, a residual dense block (RDB) structure is used to design the convolutional layer. At the same time, the batch normalization layer after the convolutional layer is removed, the activation function is replaced with PReLU, and the method of strided convolution is also selected; The first decoder is symmetric with the encoder part. The same residual dense block structure is used in the convolutional layer, and skip connections are made with the feature maps of the corresponding encoder layers. In addition, a polarization self-attention mechanism (PSA) is introduced, and the upsampling operation uses a transposed convolution with a stride of 2; The second decoder cascades and upsamples the output of each layer of the first decoder through two consecutive 3×3 convolutions, batch normalization, and ReLU function operations, and the output of the corresponding encoder layer.

3. The intelligent image preprocessing method for cuttings recognition according to claim 2, wherein In method (1), a dual-decoder image denoising network (EDDNet) based on attention mechanism and decoders with different depths are used.

4. The intelligent image preprocessing method for cuttings recognition according to claim 1, wherein, The specific content of method (2) is as follows: The target pixel points and their neighborhoods in the noisy image are taken as a whole for denoising, and the denoising model of the image is expressed as: x v = x + v In the above formula, x v is the noisy image, x is the original noise-free image, and v is the noise; Regarding each image block as a vector, the vector form of the target pixel block is represented as The vector form of the training sample block is represented as Calculate the vector similarity between the target pixel block and each training sample block: In the above formula: is the norm of the vector, m is the total number of sample categories, is the included angle of the vectors; Sort according to the vector similarity from large to small, and select the training sample blocks corresponding to 70% of the data as the initially screened similar sample blocks. Then, calculate the mean square error (MSE) between these sample blocks and the target pixel block to represent the similarity between image blocks: In the above formula, σ 2 is the noise variance; Select a threshold T. When the MSE between the noise-free pixel blocks is less than this threshold, it is: At this time, if the MSE between the noisy image blocks satisfies: e i ≤T + 2σ 2 , then this training sample block is considered similar to the target pixel block.

5. The intelligent image preprocessing method for cuttings recognition according to claim 4, wherein In method (2), when e i ≤T + 2σ 2 and the number n of image blocks satisfying the condition is less than cK 2 , take (n + cK) 2 / 2 as the number of samples. When e i ≤T + 2σ 2 and the number n of image blocks satisfying the condition is greater than cK 2 , take n as the number of samples. c ranges from 8 to 10, and K represents the side length of the local pixel block centered on the target pixel point.

6. The intelligent image preprocessing method for cuttings identification according to claim 1, characterized in that The specific content of method (3) is as follows: First, calculate the irradiance guidance map for the input low-light image. Then, after splicing the low-light image and the guidance map, the multi-scale image is input into U-Net for feature learning. The guidance map is used as an attention map and multiplied by the network output feature map to adjust the local brightness of the image. Finally, a transitional convolutional layer is used to transfer the shallow detail features of the input image to the deep layer.

7. The intelligent image preprocessing method for cuttings recognition according to claim 1, characterized in that The specific content of method (4) is as follows: During the training process, before each epoch of training, the Mosaic algorithm is used. Randomly select 4 images from the training set samples, then arrange them in 4 fixed positions of the new image, and finally draw a cross cut line in the middle part of the new image to cut off the overlapping part of the image.

8. The intelligent image preprocessing method for cuttings recognition according to claim 7, wherein In method (4), in the enhancement task, the cross-entropy loss function is used to calculate the error between the true value and the predicted value. In the calculation of the cross-entropy loss function, one-hot encoding is used to distinguish positive and negative samples, that is, label 1 is the positive sample and label 0 is the negative sample; The cross-entropy loss function is expressed as: In the above formula, m is the number of samples, n is the number of categories, and y i is one-hot encoded label information, and p i is the probability matrix.

9. The intelligent image preprocessing method for cuttings recognition according to claim 1, wherein The specific content of method (5) is as follows: The generative adversarial network GAN consists of two networks: the generator network Generator is responsible for generating simulated data; the discriminator network Discriminator is responsible for judging whether the input data is real or generated. The generator G learns the real data distribution Px and maps out fake data G(z) similar to the real data distribution. The discriminator D learns the real data distribution Px and the generated fake data distribution PG(z) to identify whether the source of the input data is real data or generated fake data, that is, to perform binary classification on x and G(z). The two are alternately trained in the network and jointly learn and progress in the confrontation, and finally reach the Nash equilibrium state.

Images