A core CT image denoising method based on deep learning

Abstract

The present invention relates to the technical field of image denoising processing, and relates to a core CT image denoising method based on deep learning, which comprises the following steps: 1. Data set preparation; 2. Model forward reasoning, passing the noise image into the model, and obtaining the output of the forward reasoning graph; 3. Loss calculation, the feature graph output by the model and the real data label; 4. Repeating steps 2 and 3, after multiple iterations and image resolution category judgment, and optimizing the image data according to the loss ratio weight; finally using the trained model to output the noise result, thereby achieving noise removal. The denoised image obtained by the method of the present invention is quite close to the real image data in terms of low-level pixel values, high-level abstract features, or overall concept and style.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising processing, and in particular to a core CT image denoising method based on deep learning. Background Art

[0002] The study of the rock physical structure of the core of the formation reservoir is of great significance for improving the recovery rate. Due to the complex core structure inside the formation reservoir, the early two-dimensional image detection is far from meeting the needs of current industrial production. The oil and gas parameters of the rock are often calculated through traditional rock physics experiments, which is too expensive and has a long cycle. Nowadays, the use of digital core technology has become one of the important means to study the internal structure of the core. With the rapid development of numerical simulation technology and computer hardware, digital core technology has become one of the best methods for rock physics research. Due to the heterogeneity of the actual core, the core established only by simple two-dimensional CT images and simple algorithms is difficult to reflect the actual situation of the core and cannot accurately describe its complex internal pore structure. Based on the two-dimensional information of the core, the core sequence image is obtained based on the CT scanning method. In the process of establishing the core, the skeleton and pore structure are often divided by the threshold algorithm. Such a simple calculation often produces some invalid points that have no reference value in the actual experiment, commonly known as noise points. How to use more effective methods to remove noise points so that the established digital core can have a similar pore structure to the real core is of great value. Summary of the invention

[0003] The present invention provides a core CT image denoising method based on deep learning, which can overcome certain or some defects of the prior art.

[0004] A core CT image denoising method based on deep learning according to the present invention comprises the following steps:

[0005] 1. Dataset preparation;

[0006] Second, model forward reasoning, passing the noise image into the model to obtain the output of the forward reasoning graph;

[0007] 3. Loss calculation, the feature map output by the model and the real data label;

[0008] 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgment, and optimize the image data according to the loss ratio weight; finally, use the trained model to output the noise result to achieve noise removal.

[0009] Preferably, in step one, the core CT is visualized by three-dimensional digitization of the core; random noise is added to the core, and a single image of the core is obtained by slicing, and a standard 128*128 noise image and a corresponding standard original image are obtained by adding salt and pepper and Gaussian white noise, and the noise points are marked as 1 and the original images are marked as 0, and the data are grouped by naming to obtain a complete data set; the data set is then sorted according to the ratio of training set: test set: validation set = 7:2:1, and the data preparation stage is completed.

[0010] Preferably, in step 2, the model is performed as follows:

[0011] a) The input image Imagegt format is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the IRA module;

[0012] b) Acquisition of stage 1 layer features: After IRA, the output channels are set to 3, 16, 16 respectively, and the stage 1 layer features are acquired;

[0013] c) Perform maxpooling on the stage1 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 32, 32, 32; finally, obtain the stage2 features;

[0014] d) Perform maxpooling on the stage2 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 64, 64, 64; finally, obtain the stage3 features;

[0015] e) Perform maxpooling on the stage3 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 128, 128, 128; finally, obtain the stage4 features;

[0016] f) Perform maxpooling on the stage4 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 256, 256, 256; finally, obtain the stage5 features;

[0017] g) Upsample the stage5 feature through deconvolution, set the output channel to 256, shorten the width and height to twice the previous value, and concatenate it with the copied stage4 feature; after three IRA operations, the output channels are 128, 128, 128; finally, obtain the stage4_1 feature;

[0018] h) Upsample the stage4_1 feature through deconvolution, set the output channel to 128, shorten the width and height to twice the previous ones, and concatenate it with the copied stage3 feature; after three IRA operations, the output channels are 64, 64, 64; finally, the stage3_1 feature is obtained;

[0019] i) Upsample the stage3_1 feature through deconvolution, set the output channel to 32, shorten the width and height to twice the previous value, and concatenate it with the copied stage2 feature; after three IRA operations, the output channels are 32, 32, 32; finally, obtain the stage2_1 feature;

[0020] j) Upsample the stage2_1 feature through deconvolution, set the output channel to 16, shorten the width and height to twice the previous value, and concatenate it with the copied stage1 feature; after three IRA operations, the output channels are 16, 16, 16; finally, obtain the stage1_1 feature;

[0021] k) The stage1_1 feature is subjected to three IRA operations, and the output channels are 16, 1, 1; the final feature map Image is obtained pred .

[0022] Preferably, in step a), the IRA module calculation steps are as follows:

[0023] 1) The IRA module consists of conv+BN+ReLU+conv+BN+ReLU+conv+BN+CBAM modules, where conv is a regular convolution, the kernels of the three convs are 1*1, 3*3, and 1*1, and the step size is 1 by default; the expansion coefficient is set in the module to control the amplification channel multiple of the 3*3 module in IRA; IRA is an inverted residual attention structure, the output is the same as the input, the number of intermediate channels is transformed to C*expansion, and the size is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively;

[0024] 2) The input image first passes through the Conv module, the input channel is C, the output channel is C, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0025] 3) The feature map passes through the BN module with C channels;

[0026] 4) The feature map passes through the ReLU module for activation;

[0027] 5) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the parameter kernel is 3*3, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0028] 6) The feature map passes through the BN module, and the number of channels is C*expansion;

[0029] 7) The feature map passes through the ReLU module for activation;

[0030] 8) The input image first passes through the Conv module, the input channel is C*expansion, the output channel is C*expansion, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0031] 9) The feature map passes through the BN module, and the number of channels is C*expansion;

[0032] 10) The feature map passes through the CBAM module to calculate the attention features;

[0033] 11) The original features and the attention features are added together to obtain the final result; the shape is (N, C, H, W), thus completing an IRA operation.

[0034] Preferably, the CBAM module includes two independent submodules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively; the calculation steps of the CBAM module are as follows:

[0035] A) The input feature map F (H×W×C) is subjected to global maxpooling based on width and global average pooling based on height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer;

[0036] B) Send them to a two-layer Conv2d, where the first layer has C input channels, C / r ​​output channels, 1 kernel, 1 stride, bias=False, and Relu activation function; the second layer has C / r input channels, C output channels, 1 kernel, 1 stride, and bias=False. The two layers of the neural network are shared.

[0037] C) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated;

[0038] D) Perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the Spatial attention module;

[0039] E) Use the feature map F output by the Channel attention module as the input feature map of the module; perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps;

[0040] F) Perform concat operation on these two feature maps based on channel;

[0041] G) After a 7×7 convolution operation;

[0042] H) is reduced to 1 channel, i.e. H×W×1; after sigmoid, a spatial attention feature is generated, i.e. M_s;

[0043] I) Multiply the feature by the input feature of the module to obtain the final generated attention feature.

[0044] Preferably, in step 3, the training loss is calculated as follows:

[0045] Calculate the loss cost between the real original image and the high-scoring image calculated by the model, where the loss is calculated as follows:

[0046] ①MSE loss calculation:

[0047]

[0048] represents the total MSE loss, is the classification value corresponding to the real picture; The classification value corresponding to the width and height image calculated by the model; W, H represent width and height respectively; x, y represent the target position index;

[0049] ②Perceptual loss calculation:

[0050]

[0051] represents the perceptual loss, C j is the channel, W, H represent the width and height of the image respectively; φ j (x) h,w,cis the perceptual feature map calculated by inference of the 11th, 24th, 42nd, and 50th layers of the real image in the trained Resnet50 residual network; j (x) h,w,c' The perceptual feature map calculated by inference of the trained Resnet50 model for the corresponding denoised image.

[0052] The present invention provides a denoising method suitable for core CT images. The method uses a coding model to train the noisy image to obtain the required denoised original image data, adopts different feature calculation modules of the residual 50 network, and uses the edge module information to establish feature map information as a priori loss to obtain a more accurate denoised image. At the same time, in order to avoid the problems that the reconstruction results in the previous methods only have a high signal-to-noise ratio PSNR, lack of high-frequency information, and the appearance of over-smooth textures, under the support of the CBAM attention module and multi-scale perceptual loss, the denoised image obtained by the method of the present invention is quite close to the real image data in terms of low-level pixel values, high-level abstract features, or overall concepts and styles. It has the following advantages in style reconstruction: high-level features, global structure, and texture are obvious. When denoising, there is more detailed information such as the underlying features, edges, colors, etc. of the image, and the effect is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a core CT image denoising method based on deep learning in Example 1;

[0054] Figure 2 This is a schematic diagram of the model in Example 1;

[0055] Figure 3 This is a schematic diagram of the IRA module in Example 1;

[0056] Figure 4 This is a schematic diagram of the CBAM module in Example 1. DETAILED DESCRIPTION

[0057] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0058] Example 1

[0059] like Figure 1 As shown, this embodiment provides a core CT image denoising method based on deep learning, which includes the following steps:

[0060] 1. Dataset preparation: visualize the core CT through three-dimensional digitization of the core; add random noise to the core, use the slicing method to get a single picture of the core, and obtain the standard 128*128 noise image and the corresponding standard original image by adding salt and pepper and Gaussian white noise. Mark the noise point as 1 and the original image as 0, group them by naming, and get a complete data set; then divide the data set into training set: test set: validation set = 7:2:1. At this point, the data preparation stage is completed.

[0061] Second, model forward reasoning, passing the noise image into the model to obtain the output of the forward reasoning graph;

[0062] The model proceeds as follows:

[0063] a) The input image Imagegt format is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the IRA module;

[0064] The IRA module calculation steps are as follows:

[0065] 1) The IRA module consists of conv+BN+ReLU+conv+BN+ReLU+conv+BN+CBAM modules, where conv is a conventional convolution, the kernels of the three convs are 1*1, 3*3, and 1*1, and the step size is 1 by default; the expansion coefficient is set in the module to control the amplification channel multiple of the 3*3 module in IRA; IRA is an inverted residual attention structure, the output is the same as the input, and the number of intermediate channels is transformed into C*expansion, which can extract more semantic information and summarize the semantic information. More directly, the inverted residual structure design can improve memory efficiency more, and can also improve the ability of gradients in back propagation, which is convenient for training; the sizes are (N, C, H, W), which represent the batch, channel, width, and height of the image respectively;

[0066] 2) The input image first passes through the Conv module, the input channel is C, the output channel is C, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0067] 3) The feature map passes through the BN module with C channels;

[0068] 4) The feature map passes through the ReLU module for activation;

[0069] 5) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the parameter kernel is 3*3, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0070] 6) The feature map passes through the BN module, and the number of channels is C*expansion;

[0071] 7) The feature map passes through the ReLU module for activation;

[0072] 8) The input image first passes through the Conv module, the input channel is C*expansion, the output channel is C*expansion, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0073] 9) The feature map passes through the BN module, and the number of channels is C*expansion;

[0074] 10) The feature map passes through the CBAM module to calculate the attention features;

[0075] 11) The original features and the attention features are added together to obtain the final result; the shape is (N, C, H, W), thus completing an IRA operation.

[0076] The CBAM module contains two independent submodules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively; this not only saves parameters and computing power, but also ensures that it can be integrated into the existing network architecture as a plug-and-play module. The calculation steps of the CBAM module are as follows:

[0077] A) The input feature map F (H×W×C) is subjected to global maxpooling and global average pooling based on width and height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer;

[0078] B) Then send them to a two-layer Conv2d, where the first layer has C input channels, C / r ​​output channels (r is the reduction rate), 1 kernel, 1 stride, bias = False, and Relu activation function; the second layer has C / r input channels, C output channels, 1 kernel, 1 stride, and bias = False. The two layers of the neural network are shared.

[0079] C) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated;

[0080] D) Finally, perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the SpatialAttention module;

[0081] E) Use the feature map F output by the Channel attention module as the input feature map of this module; first, perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps;

[0082] F) Then perform concat operation (channel splicing) on ​​these two feature maps based on channel;

[0083] G) Then it goes through a 7×7 convolution (7×7 has a larger receptive field than 3×3, and the effect is better) operation;

[0084] H) is reduced to 1 channel, i.e. H×W×1; then sigmoid is used to generate the spatial attention feature, i.e. M_s;

[0085] I) Finally, the feature is multiplied by the input feature of the module to obtain the final generated attention feature.

[0086] b) Acquisition of stage 1 layer features: After IRA, since IRA does not change the feature shape, it will not be repeated here. The output channels are set to 3, 16, 16 respectively, and the stage 1 layer features are obtained;

[0087] c) Perform maxpooling on the stage1 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 32, 32, 32; finally, obtain the stage2 features;

[0088] d) Perform maxpooling on the stage2 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 64, 64, 64; finally, obtain the stage3 features;

[0089] e) Perform maxpooling on the stage3 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 128, 128, 128; finally, obtain the stage4 features;

[0090] f) Perform maxpooling on the stage4 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 256, 256, 256; finally, obtain the stage5 features;

[0091] g) Upsample the stage5 feature through deconvolution, set the output channel to 256, shorten the width and height to twice the previous value, and concatenate it with the copied stage4 feature. After three IRA operations, the output channels are 128, 128, 128, and finally obtain the stage4_1 feature;

[0092] h) Upsample the stage4_1 feature through deconvolution, set the output channel to 128, shorten the width and height to twice the previous value, and concatenate it with the copied stage3 feature. After three IRA operations, the output channels are 64, 64, 64, and finally the stage3_1 feature is obtained;

[0093] i) Upsample the stage3_1 feature through deconvolution, set the output channel to 32, shorten the width and height to twice the previous value, and concatenate it with the copied stage2 feature. After three IRA operations, the output channels are 32, 32, 32, and finally the stage2_1 feature is obtained;

[0094] j) Upsample the stage2_1 feature through deconvolution, set the output channel to 16, shorten the width and height to twice the previous value, and concatenate it with the copied stage1 feature. After three IRA operations, the output channels are 16, 16, 16, and finally the stage1_1 feature is obtained;

[0095] k) The stage1_1 feature is subjected to three IRA operations, and the output channels are 16, 1, 1; the final feature map Image is obtained pred .

[0096] 3. Loss calculation: the feature map output by the model and the real data label; the training loss calculation of the generator image data is used as the optimization target, so that the converted high-resolution image is more realistic;

[0097] The training loss is calculated as follows:

[0098] Calculate the loss cost between the real original image and the high-scoring image calculated by the model, where the loss is calculated as follows:

[0099] ①MSE loss calculation:

[0100]

[0101] represents the total MSE loss, is the classification value corresponding to the real picture; The classification value corresponding to the width and height image calculated by the model; W, H represent width and height respectively; x, y represent the target position index;

[0102] ②Perceptual loss calculation:

[0103] The original cost function MSE makes the denoising result have a higher signal-to-noise ratio PSNR, but lacks high-frequency information and overly smooth textures. After the addition of perceptual loss prior, the denoised image is quite close to the real image in terms of low-level pixel values, high-level abstract features, and overall concept and style.

[0104]

[0105] represents the perceptual loss, C j is the channel, W, H represent the width and height of the image respectively; φ j (x) h,w,c is the perceptual feature map calculated by inference of the 11th, 24th, 42nd, and 50th layers of the real image in the trained Resnet50 residual network; j (x) h,w,c' The perceptual feature map calculated by inference of the trained Resnet50 model for the corresponding denoised image.

[0106] 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgment, and optimize the image data according to the loss ratio weight (the experimental ratio is 8:2, that is, the weights are 0.8 and 0.2); finally, use the trained model to output the noise result, thereby achieving noise removal.

[0107] This embodiment only uses two modules, the algorithm is simpler and more intuitive, and the codec model can be used to train the noisy image to obtain the required denoised original image data. Different feature calculation modules of the residual 50 network are used, and the edge module information is used to establish the feature map information as the prior loss to obtain a more accurate denoised image.

[0108] This embodiment avoids the problems of the reconstruction results having a high signal-to-noise ratio (PSNR), lack of high-frequency information, and the appearance of overly smooth textures that occur in previous methods. The denoised image obtained in this embodiment is quite close to the real image data in terms of low-level pixel values, high-level abstract features, and overall concepts and styles. When reconstructing the style, it has the following advantages: high-level features, global structure, and texture are obvious. When denoising, there is more detailed information on the underlying features, edges, colors, etc. of the image, and the effect is better.

[0109] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A core CT image denoising method based on deep learning, characterized by: The following steps are involved:

1. Dataset preparation; Second, model forward reasoning, passing the noise image into the model to obtain the output of the forward reasoning graph; In step 2, the model proceeds as follows: a) Input image Image gt The format is (N,C,H,W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the IRA module; The IRA module calculation steps are as follows: 1) The IRA module consists of conv+BN+ReLU+conv+BN+ReLU+conv+BN+CBAM modules, where conv is a conventional convolution, the kernels of the three convs are 1*1, 3*3, and 1*1, and the step size is 1 by default; the expansion coefficient is set in the module to control the amplification channel multiple of the 3*3 module in IRA; the IRA module is an inverted residual attention structure, the output is the same as the input, the number of intermediate channels is transformed to C*expansion, and the size is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively; 2) The input image first passes through the Conv module, the input channel is C, the output channel is C, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False; 3) The feature map passes through the BN module with C channels; 4) The feature map passes through the ReLU module for activation; 5) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the parameter kernel is 3*3, the step size is 1*1, the padding defaults to 1, and the bias is set to False; 6) The feature map passes through the BN module, and the number of channels is C*expansion; 7) The feature map passes through the ReLU module for activation; 8) The input image first passes through the Conv module, the input channel is C*expansion, the output channel is C*expansion, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False; 9) The feature map passes through the BN module, and the number of channels is C*expansion; 10) The feature map passes through the CBAM module to calculate the attention features; 11) The original features and the attention features are added together to obtain the final result; the shape is (N, C, H, W), and thus an IRA operation is completed; b) Acquisition of stage 1 layer features: After IRA, the output channels are set to 3, 16, 16 respectively, and the stage 1 layer features are acquired; c) Perform maxpooling on the stage1 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 32, 32, 32; finally, obtain the stage2 features; d) Perform maxpooling on the stage2 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 64, 64, 64; finally, obtain the stage3 features; e) Perform maxpooling on the stage3 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 128, 128, 128; finally, obtain the stage4 features; f) Perform maxpooling on the stage4 features with a pooling kernel of 2*2, and then perform three IRA operations with output channels of 256, 256, 256; finally, obtain the stage5 features; g) Upsample the stage5 feature through deconvolution, set the output channel to 256, shorten the width and height to twice the previous value, and concatenate it with the copied stage4 feature; after three IRA operations, the output channels are 128, 128, 128; finally, obtain the stage4_1 feature; h) Upsample the stage4_1 feature through deconvolution, set the output channel to 128, shorten the width and height to twice the previous ones, and concatenate it with the copied stage3 feature; after three IRA operations, the output channels are 64, 64, 64; finally, the stage3_1 feature is obtained; i) Upsample the stage3_1 feature through deconvolution, set the output channel to 32, shorten the width and height to twice the previous value, and concatenate it with the copied stage2 feature; after three IRA operations, the output channels are 32, 32, 32; finally, obtain the stage2_1 feature; j) Upsample the stage2_1 feature through deconvolution, set the output channel to 16, shorten the width and height to twice the previous value, and concatenate it with the copied stage1 feature; after three IRA operations, the output channels are 16, 16, 16; finally, obtain the stage1_1 feature; k) The stage1_1 feature is subjected to three IRA operations, and the output channels are 16, 1, 1; the final feature map Image is obtained pred ; 3. Loss calculation, the feature map output by the model and the real data label; 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgment, and optimize the image data according to the loss ratio weight; finally, use the trained model to output the noise result to achieve noise removal.

2. The core CT image denoising method based on deep learning according to claim 1, characterized in that: In step one, the core CT is visualized by three-dimensional digitization of the core; random noise is added to the core, and a single image of the core is obtained by slicing. The standard 128*128 noise image and the corresponding standard original image are obtained by adding salt and pepper and Gaussian white noise. The noise points are marked as 1 and the original images are marked as 0. The complete data set is obtained by grouping through naming; the data set is then divided into training set: test set: validation set = 7:2:

1. At this point, the data preparation stage is completed.

3. The core CT image denoising method based on deep learning according to claim 2, characterized in that: The CBAM module contains two independent sub-modules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively; the calculation steps of the CBAM module are as follows: A) The input feature map F (H×W×C) is subjected to global max pooling based on width and global average pooling based on height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer; B) Send them to a two-layer Conv2d, where the first layer has C channels as input, C / r ​​as output, 1 kernel, 1 stride, bias=False, and ReLU as activation function; The second layer has C / r input channels, C output channels, 1 kernel, 1 stride, and bias = False. The two layers of the neural network are shared. C) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated; D) Perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the Spatial attention module; E) Use the feature map F output by the Channel attention module as the input feature map of the module; perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps; F) Perform concat operation on these two feature maps based on channel; G) After a 7×7 convolution operation; H) is reduced to 1 channel, i.e. H×W×1; after sigmoid, spatial attention feature is generated, i.e. M_s; I) Multiply the feature by the input feature of the module to obtain the final generated attention feature.

4. The core CT image denoising method based on deep learning according to claim 3, characterized in that: In step 3, the training loss is calculated as follows: Calculate the loss cost between the real original image and the high-scoring image calculated by the model, where the loss is calculated as follows: ①MSEloss calculation: represents the total MSE loss, is the classification value corresponding to the real picture; The classification value corresponding to the width and height image calculated by the model; W, H represent the width and height respectively; x, y represent the target position index; ②Perceptual loss calculation: represents the perceptual loss, C j is the channel, W, H represent the width and height of the image respectively; φ j (x) h,w,c is the perceptual feature map calculated by inference of the 11th, 24th, 42nd, and 50th layers of the real image in the trained Resnet50 residual network; j (x) h,w,c′ The perceptual feature map calculated by inference of the trained Resnet50 model for the corresponding denoised image.

Images