CN113793263B - Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution - Google Patents
Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution Download PDFInfo
- Publication number
- CN113793263B CN113793263B CN202110967124.7A CN202110967124A CN113793263B CN 113793263 B CN113793263 B CN 113793263B CN 202110967124 A CN202110967124 A CN 202110967124A CN 113793263 B CN113793263 B CN 113793263B
- Authority
- CN
- China
- Prior art keywords
- convolution
- resolution image
- layer
- residual error
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 15
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 12
- 230000004927 fusion Effects 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 9
- 238000003384 imaging method Methods 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 238000005096 rolling process Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 2
- 239000011800 void material Substances 0.000 claims description 2
- 238000013507 mapping Methods 0.000 abstract description 2
- 230000008707 rearrangement Effects 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 9
- 230000006870 function Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 2
- 239000003129 oil well Substances 0.000 description 2
- 239000000523 sample Substances 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002592 echocardiography Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000002604 ultrasonography Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10132—Ultrasound image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution, which comprises the steps of firstly carrying out shallow feature extraction on convolution layers with the size of 9 × 9 and the number of channels of 64, then utilizing the characteristic that the cavity convolution improves the receptive field under the condition that the parameter quantity is not changed to construct a multi-scale cavity convolution block, then combining the multi-scale cavity convolution block with a common 3 × 3 convolution layer and a BN layer to form a residual error block, connecting 16 residual error blocks in series to form a residual error network, carrying out nonlinear mapping on features by adopting a multi-path parallel structure to obtain high-level features, and finally carrying out rearrangement on feature maps by sub-pixel convolution layers to finally obtain a high-resolution image SR.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution.
Background
In the field of oil logging, the well-around imaging logging is an important branch. The imaging logging around the well can reflect the condition of the oil well by an intuitive well wall image, the development condition of cracks and holes on the well wall can be clearly seen, the imaging logging around the well is an important means for evaluating the oil well, and the resolution evaluation of logging personnel on the well is directly influenced by the definition of the obtained well wall image. High-resolution reconstruction is one of main research directions for image enhancement, the enhancement effect of traditional methods such as a bilinear interpolation method and a bicubic interpolation method is not obvious, and super-resolution reconstruction based on deep learning is researched by more and more people due to the obvious enhancement effect. For example: the deep learning SRCNN algorithm for image super-resolution reconstruction uses bicubic interpolation for upsampling, and then a three-layer neural network is constructed for feature learning and reconstruction.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution so as to realize the rapid enhancement of a logging image and facilitate logging personnel to better observe the image and analyze the underground condition in real time.
In order to achieve the purpose, the invention discloses a parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution, which is characterized by comprising the following steps of:
(1) Acquiring and preprocessing an image;
acquiring a plurality of original high-resolution well logging images by using a well periphery ultrasonic imaging instrument, cutting each high-resolution image to obtain a high-resolution image HR with the same size, and performing n-time down-sampling on each high-resolution well logging image to obtain a low-resolution image LR with the size of H x W, wherein n is a sampling multiple;
(2) Constructing an image reconstruction network based on multi-scale cavity convolution and training;
(2.1) extracting a feature map containing shallow features;
inputting the low-resolution image LR into a convolution layer v1 with the size of 9 × 9 and the number of channels of 64, and performing shallow feature extraction by using a PRelu activation function to obtain 64 feature maps with the size of H × W;
(2.2) constructing two parallel residual error networks;
(2.2.1) constructing a multi-scale cavity rolling block;
simultaneously extracting the features of 64 feature maps by adopting 64 convolution layers v2 with convolution kernels of 3 × 3 and 64 convolution layers v3 with convolution kernels of 3 × 3 and 2 expansion rates, adding the output results of v2 and v3, inputting the added result into v2 and v3 again, and finally performing feature fusion on the output results of v2 and v3 by using convolution kernel 1, and directly adding the fused result with the input 64 feature maps to construct a multi-scale hole convolution block;
(2.2.2) constructing a single-path residual error network: connecting a multi-scale cavity convolution block, a convolution layer v4 of a convolution kernel with the size of 3 x 3 and a normalization layer together, adding the result to 64 input feature maps to form a residual block, and connecting 16 residual blocks in series to form a single-path residual network;
(2.2.3) constructing two parallel residual error networks: simultaneously extracting features of a convolutional layer v5 with 64 convolutional kernels with 5 × 5 sizes and a convolutional layer v6 with 64 convolutional kernels with 5 × 5 sizes and 2 expansion rates, adding output results of v5 and v6, inputting the added output results into v5 and v6 again, and finally performing feature fusion on the output results of v5 and v6 by using 1 × 1 convolutional kernel and directly adding the input output results with 64 feature maps to construct another multi-scale hole convolutional block; then, the multi-scale hole convolution block is added with a convolution layer v4 and a normalization layer of a convolution kernel of 3 x 3 size and input 64 feature maps to form another residual block, and then 16 residual blocks are connected in series to form another single-path residual network;
connecting two single-path residual error networks to the convolution layer in the step (2.1) in a parallel mode, performing feature fusion on the results output by the two parallel residual error networks in the step (2.2) by using a 1-to-1 convolution kernel, inputting the results into a convolution layer v4 and a normalization layer of a 3-to-3 convolution kernel, and directly adding the results and the output of the convolution layer in the step (2.1) to obtain a feature map containing 64 high-level features;
(2.3) reconstructing a high-resolution image;
inputting 64 feature graphs containing high-level features into channels with the number of 64 x n 2 The channel number is widened by the convolution layer v7, and then the widened channel number is input into the sub-pixel convolution layer, so that the single pixels on the multiple channel feature maps are combined and arranged into a group of pixels on one channel feature map, namely: h W r n 2 → n H W r, r is the number of channels after the last stage output; finally, outputting a reconstructed high-resolution image SR with the size of (n × H) × (n × W) and the number of channels of 3 through a convolution layer v8 with the size of 9 × 9 and the number of channels of 3;
(2.4) calculating a loss function value;
calculating pixel mean square error MSE of the reconstructed high-resolution image SR and the original high-resolution image HR, and taking the MSE as a loss function value;
wherein SR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image SR, and HR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image HR;
(2.5) repeating the steps (2.1) - (2.4), continuing to train the image reconstruction network, and performing parameter optimization by using an Adam optimization algorithm to minimize MSE (mean Square error), so as to finally obtain a trained image reconstruction network model;
(3) And acquiring a logging image in real time, and inputting the logging image into the trained image reconstruction network so as to output a reconstructed high-resolution image.
The invention aims to realize the following steps:
the invention relates to a parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution, which comprises the steps of firstly carrying out shallow feature extraction on convolution layers with the size of 9 × 9 and the number of channels of 64, then utilizing the property of improving the receptive field of the cavity convolution under the condition of constant parameter quantity to construct a multi-scale cavity convolution block, then combining the multi-scale cavity convolution block with a common 3 × 3 convolution layer and a BN layer to form a residual error block, connecting 16 residual error blocks in series to form a residual error network, carrying out nonlinear mapping on features by adopting a multi-path parallel structure to obtain high-level features, finally carrying out rearrangement on feature maps by using sub-pixel convolution layers to finally obtain a high-resolution image SR.
Meanwhile, the parallel residual error network high-resolution image reconstruction method based on the multi-scale cavity convolution also has the following beneficial effects:
(1) The receptive field is improved by using the cavity convolution under the condition that the parameter quantity is not changed, and more global characteristics are obtained.
(2) And feature information of different scales is complemented by using a parallel network structure.
(3) Compared with the traditional bicubic interpolation and deep learning super-resolution reconstruction classical algorithms SRCNN, VDSR, SRResNet and the like, the multi-scale parallel network based on the cavity convolution obviously improves the PSNR and SSIM of image reconstruction objective indexes.
Drawings
FIG. 1 is a flow chart of a parallel residual error network high-resolution image reconstruction method based on multi-scale hole convolution according to the invention;
FIG. 2 is a block diagram of a peri-borehole ultrasound imager;
FIG. 3 is a multi-scale hole volume block;
FIG. 4 is a parallel network structure based on multi-scale hole rolling blocks;
FIG. 5 is a 4-fold reconstructed mean index analysis of the test set for this and other classical algorithms;
FIG. 6 is a test set 2-fold reconstructed mean index analysis of the present algorithm and other classical algorithms;
FIG. 7 is a comparison of the effect of multiple single log image participation of the present algorithm and other classical algorithms;
Detailed Description
The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the main content of the present invention.
Examples
In this embodiment, as shown in fig. 1, the parallel residual error network high resolution image reconstruction method based on multi-scale hole convolution of the present invention includes the following steps:
s1, image acquisition and pretreatment;
as shown in fig. 2, the borehole ultrasonic imaging instrument includes a ground control system and an underground logging circuit system, during logging, the ultrasonic transducer probe is driven by the motor transmission device to rotate for 360 degrees, each time the ultrasonic transducer probe rotates for one circle, the tooth sensor generates 250 pulses and the body maark sensor generates 1 pulse, then the main control module of the underground logging circuit shapes the signals, the shaped tooth signals serve as transmitting signals to enable the FPGA to drive the transmitting circuit to generate high-voltage pulses and excite the ultrasonic transducer, and the shaped body mark signals serve as acquisition cycle synchronization signals to mark a new circle acquisition starting point.
The ultrasonic transducer collects full-wave column data of echoes, the full-wave column data are sent to the ground control system through an EDIB bus, the upper computer receives the data converted through the USB port and then analyzes the data, echo amplitude and arrival time data in a data packet are extracted, and a final borehole wall image is synthesized by using software.
After actual logging, 127 images with the resolution of 140 × 140, 152 images with the resolution of 180 × 180, 860 images with the resolution of 352 × 352 are finally obtained as a training set and 366 images with the resolution of 352 × 352, 265 images with the resolution of 180 × 180, and 30 images with the resolution of 140 × 140 are finally obtained as a testing set;
and taking the original images in the training set as original high-resolution images, cutting each original high-resolution image to obtain an original high-resolution image HR with the size of 96 × 96, and performing four-time down-sampling on the HR to obtain a low-resolution image LR with the size of H × W =24 × 24 so as to perform network training.
S2, constructing an image reconstruction network based on multi-scale hole convolution and training;
s2.1, extracting a characteristic diagram containing shallow layer characteristics;
in this embodiment, 20 LR images are randomly extracted into the feature extraction layer each time, the extracted low-resolution image LR is input to the convolution layer v1 with the size of 9 × 9 and the number of channels of 64, and the shallow feature extraction is performed by using the prilu activation function, so as to obtain 64 feature maps with the size of H × W;
s2.2, constructing two parallel residual error networks;
s2.2.1, constructing a multi-scale cavity rolling block;
simultaneously extracting the features of 64 feature maps by adopting a convolutional layer v2 with 64 convolution kernels with 3 × 3 sizes and a convolutional layer v3 with 64 convolution kernels with 3 × 3 convolution kernels with 2 expansion rates, then adding the output results of v2 and v3 and inputting the added result into v2 and v3 again, and finally performing feature fusion on the output results of v2 and v3 by using a 1 × 1 convolution kernel and directly adding the input result and the input 64 feature maps to construct a multi-scale cavity convolutional block, as shown in fig. 3;
s2.2.2, constructing a single-path residual error network: connecting a multi-scale cavity convolution block, a convolution layer v4 of a convolution kernel with the size of 3 x 3 and a normalization layer together, adding the result to 64 input feature maps to form a residual block, and connecting 16 residual blocks in series to form a single-path residual network;
s2.2.3, as shown in FIG. 4, two parallel residual error networks are constructed: simultaneously extracting features of a convolutional layer v5 with 64 convolutional kernels with 5 × 5 sizes and 2 expansion rates, adding output results of v5 and v6, inputting the added output results into v5 and v6 again, and finally performing feature fusion on the output results of v5 and v6 by using 1 × 1 convolutional kernel and directly adding the input output results with 64 feature maps to construct another multi-scale hole convolutional block; then, the multi-scale hole convolution block is added with a convolution layer v4 and a normalization layer of a convolution kernel with the size of 3 x 3 and input 64 feature maps to form another residual block, and then 16 residual blocks are connected in series to form another single-path residual network;
connecting two single-path residual error networks to the convolution layer in the step S2.1 in a parallel mode, performing feature fusion on the results output by the two parallel residual error networks in the step S2.2 by using a 1-to-1 convolution kernel, inputting the results into a convolution layer v4 and a normalization layer of a 3-to-3 convolution kernel, and directly adding the results and the output of the convolution layer in the step S2.1 to obtain a feature map containing 64 high-level features;
s2.3, reconstructing a high-resolution image;
the 64 feature maps containing high-level features are input to the convolution layer v7 with the channel number 64 × 4 to widen the channel number, and then input to the sub-pixel convolution layer, so that the single pixels on the multiple channel feature maps are combined and arranged into a group of pixels on one channel feature map, namely: h W R4 2 → 4 × h (4 × w) × r, in this embodiment, r is the number of channels 64 after the last stage output; finally, outputting a reconstructed high-resolution image SR with the size of (4 × H) × (4 × W) and the number of channels of 3 through a convolution layer v8 with the size of 9 × 9 and the number of channels of 3;
s2.4, calculating a loss function value;
calculating pixel mean square error MSE of the reconstructed high-resolution image SR and the original high-resolution image HR, and taking the MSE as a loss function value;
wherein SR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image SR, and HR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image HR;
s2.5, repeating the steps S2.1-S2.4, continuing to train the image reconstruction network, and performing parameter optimization by using an Adam optimization algorithm to minimize MSE (mean Square error), so as to finally obtain a trained image reconstruction network model;
and S3, collecting a logging image in real time, and inputting the logging image into the trained image reconstruction network so as to output a reconstructed high-resolution image.
Authentication
In this embodiment, the reconstructed high resolution image (SR) and the original high resolution image (HR) are reconstructed at high resolution, and a peak signal-to-noise ratio (PSNR) and a Structural Similarity (SSIM) index, which are commonly used in an algorithm, are compared. PSNR is compared from pixels, and the higher PSNR, the less distortion representing the reconstructed image, and the more similar the pixels are to those of HR. The peak signal-to-noise ratio is calculated as follows:
where n is the number of bits per pixel, and is typically 8.
SSIM is to compare two images from contrast, structural features and brightness, and the closer the SSIM is to 1, the more similar the representative images are, the better the reconstruction effect is. The structural similarity is calculated as follows:
SSIM(X,Y)=l(X,Y)·c(X,Y)·s(X,Y)
wherein X and Y represent the image SR and the image HR, μ, respectively X 、μ Y Representing the mean, σ, of the images SR and HR, respectively X 、σ Y Denotes the standard deviation, σ, of the images SR and HR, respectively XY Representing the covariance of the images SR and HR, c 1 、c 2 、c 3 Is a constant.
In order to verify the reconstruction effect of the algorithm, the algorithm and the classical algorithms SRCNN, VDSR and SRResNet for deep learning super-resolution reconstruction, the void convolutional layer with the expansion rate of 2 which is changed into the convolutional layer in SRResNet, and the average PSNR and average SSIM which are used for performing quadruple reconstruction effect by using the test set are shown in fig. 5. Then, the above steps are repeated to change the network proposed by the algorithm and the classical algorithms SRCNN, VDSR and SRResNet of deep learning super-resolution reconstruction, change the convolution layer in SRResNet into a cavity convolution layer with the expansion rate of 2, and change the single-path residual network in SRResNet into a parallel two-path residual network, and simultaneously use the average PSNR and the average SSIM pair with the double reconstruction effect of the test set as shown in FIG. 6. The image shows that the multi-scale cavity convolution blocks and the parallel network structure are effective to the reconstruction effect, and the parallel residual error network high-resolution image reconstruction method based on the multi-scale cavity convolution, which is provided by the algorithm, is improved to the reconstruction effect.
As shown in fig. 7, the algorithm and the conventional bicubic interpolation are shown, and based on the SRCNN, VDSR, ESPCN, SRResNet of the deep learning, the objective indicators PSNR and SSIM comparison and subjective visual comparison are performed on four well-logging images randomly selected from a test set.
Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.
Claims (1)
1. A parallel residual error network high-resolution image reconstruction method based on multi-scale cavity convolution is characterized by comprising the following steps:
(1) Acquiring and preprocessing an image;
acquiring a plurality of original high-resolution well logging images by using a well periphery ultrasonic imaging instrument, cutting each high-resolution image to obtain a high-resolution image HR with the same size, and performing n-time down-sampling on each high-resolution well logging image to obtain a low-resolution image LR with the size of H x W, wherein n is a sampling multiple;
(2) Constructing an image reconstruction network based on multi-scale void convolution and training;
(2.1) extracting a feature map containing shallow features;
inputting the low-resolution image LR into a convolution layer v1 with the size of 9 × 9 and the number of channels of 64, and performing shallow feature extraction by using a PRelu activation function to obtain 64 feature maps with the size of H × W;
(2.2) constructing two parallel residual error networks;
(2.2.1) constructing a multi-scale cavity rolling block;
simultaneously extracting the features of 64 feature maps by adopting a convolutional layer v2 with 64 convolutional kernels with 3 × 3 sizes and 64 convolutional layers v3 with 3 × 3 convolutional kernels with 2 expansion rates, then adding the output results of v2 and v3 and inputting the added results into v2 and v3 again, and finally performing feature fusion on the output results of v2 and v3 by using a 1 × 1 convolutional kernel and directly adding the input results with 64 feature maps to construct a multi-scale cavity convolutional block;
(2.2.2) constructing a single-path residual error network: connecting a multi-scale cavity convolution block, a convolution layer v4 of a convolution kernel with the size of 3 x 3 and a normalization layer together, adding the result to 64 input feature maps to form a residual block, and connecting 16 residual blocks in series to form a single-path residual network;
(2.2.3) constructing two parallel residual error networks: simultaneously extracting features of a convolutional layer v5 with 64 convolutional kernels with 5 × 5 sizes and a convolutional layer v6 with 64 convolutional kernels with 5 × 5 sizes and 2 expansion rates, adding output results of v5 and v6, inputting the added output results into v5 and v6 again, and finally performing feature fusion on the output results of v5 and v6 by using 1 × 1 convolutional kernels and directly adding the fused output results with 64 input feature graphs to construct another multi-scale hole convolutional block; then, the multi-scale hole convolution block is added with a convolution layer v4 and a normalization layer of a convolution kernel of 3 x 3 size and input 64 feature maps to form another residual block, and then 16 residual blocks are connected in series to form another single-path residual network;
connecting two single-path residual error networks to the convolution layer in the step (2.1) in a parallel mode, performing feature fusion on the result output by the two parallel residual error networks in the step (2.2) by using a 1-to-1 convolution kernel, inputting the result into a convolution layer v4 of a 3-to-3 convolution kernel and a normalization layer, and directly adding the result and the output of the convolution layer in the step (2.1) to obtain a feature map containing 64 high-level features;
(2.3) reconstructing a high-resolution image;
inputting 64 feature graphs containing high-level features into channels with the number of 64 x n 2 The channel number is widened by the convolution layer v7, and then the widened channel number is input into the sub-pixel convolution layer, so that the single pixels on the multiple channel feature maps are combined and arranged into a group of pixels on one channel feature map, namely: h W r n 2 → n H W r, r is the number of channels after the last stage output; finally, outputting a reconstructed high-resolution image SR with the size of (n × H) × (n × W) and the number of channels of 3 through a convolution layer v8 with the size of 9 × 9 and the number of channels of 3;
(2.4) calculating a loss function value;
calculating pixel mean square error MSE of the reconstructed high-resolution image SR and the original high-resolution image HR, and taking the MSE as a loss function value;
wherein SR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image SR, and HR (i, j) represents the pixel value of the pixel point with the coordinate (i, j) in the high-resolution image HR;
(2.5) repeating the steps (2.1) - (2.4), continuing to train the image reconstruction network, and performing parameter optimization by using an Adam optimization algorithm to minimize MSE (mean Square error), so as to finally obtain a trained image reconstruction network model;
(3) And acquiring a logging image in real time, and inputting the logging image into the trained image reconstruction network so as to output a reconstructed high-resolution image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110967124.7A CN113793263B (en) | 2021-08-23 | 2021-08-23 | Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110967124.7A CN113793263B (en) | 2021-08-23 | 2021-08-23 | Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113793263A CN113793263A (en) | 2021-12-14 |
CN113793263B true CN113793263B (en) | 2023-04-07 |
Family
ID=78876216
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110967124.7A Active CN113793263B (en) | 2021-08-23 | 2021-08-23 | Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113793263B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114025118A (en) * | 2022-01-06 | 2022-02-08 | 广东电网有限责任公司中山供电局 | Low-bit-rate video reconstruction method and system, electronic equipment and storage medium |
CN114529519B (en) * | 2022-01-25 | 2024-07-12 | 河南大学 | Image compressed sensing reconstruction method and system based on multi-scale depth cavity residual error network |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097512A (en) * | 2019-04-16 | 2019-08-06 | 四川大学 | Construction method and the application of the three-dimensional MRI image denoising model of confrontation network are generated based on Wasserstein |
CN110211038A (en) * | 2019-04-29 | 2019-09-06 | 南京航空航天大学 | Super resolution ratio reconstruction method based on dirac residual error deep neural network |
CN110232653A (en) * | 2018-12-12 | 2019-09-13 | 天津大学青岛海洋技术研究院 | The quick light-duty intensive residual error network of super-resolution rebuilding |
CN110930306A (en) * | 2019-10-28 | 2020-03-27 | 杭州电子科技大学 | Depth map super-resolution reconstruction network construction method based on non-local perception |
CN111047515A (en) * | 2019-12-29 | 2020-04-21 | 兰州理工大学 | Cavity convolution neural network image super-resolution reconstruction method based on attention mechanism |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11237111B2 (en) * | 2020-01-30 | 2022-02-01 | Trustees Of Boston University | High-speed delay scanning and deep learning techniques for spectroscopic SRS imaging |
-
2021
- 2021-08-23 CN CN202110967124.7A patent/CN113793263B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110232653A (en) * | 2018-12-12 | 2019-09-13 | 天津大学青岛海洋技术研究院 | The quick light-duty intensive residual error network of super-resolution rebuilding |
CN110097512A (en) * | 2019-04-16 | 2019-08-06 | 四川大学 | Construction method and the application of the three-dimensional MRI image denoising model of confrontation network are generated based on Wasserstein |
CN110211038A (en) * | 2019-04-29 | 2019-09-06 | 南京航空航天大学 | Super resolution ratio reconstruction method based on dirac residual error deep neural network |
CN110930306A (en) * | 2019-10-28 | 2020-03-27 | 杭州电子科技大学 | Depth map super-resolution reconstruction network construction method based on non-local perception |
CN111047515A (en) * | 2019-12-29 | 2020-04-21 | 兰州理工大学 | Cavity convolution neural network image super-resolution reconstruction method based on attention mechanism |
Non-Patent Citations (3)
Title |
---|
PRED: A PARALLEL NETWORK FOR HANDLING MULTIPLE DEGRADATIONS VIA SINGLE MODEL IN SINGLE IMAGE SUPER-RESOLUTION;Guangyang Wu等;《网页在线公开:https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8804409》;第1-5页 * |
基于并行残差卷积网络的图像超分辨重建;杨伟铭等;《空军工程大学学报(自然科学版)》;第20卷(第4期);第84-89页 * |
基于并行通道-空间注意力机制的腹部MRI影像多尺度超分辨率重建;樊帆等;《计算机应用》;第40卷(第12期);第3624-3630页 * |
Also Published As
Publication number | Publication date |
---|---|
CN113793263A (en) | 2021-12-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113793263B (en) | Parallel residual error network high-resolution image reconstruction method for multi-scale cavity convolution | |
CN111127374B (en) | Pan-sharing method based on multi-scale dense network | |
CN112734646B (en) | Image super-resolution reconstruction method based on feature channel division | |
CN110310227A (en) | A kind of image super-resolution rebuilding method decomposed based on high and low frequency information | |
CN109214989B (en) | Single image super resolution ratio reconstruction method based on Orientation Features prediction priori | |
CN108629816A (en) | The method for carrying out thin layer MR image reconstruction based on deep learning | |
CN109712077B (en) | Depth dictionary learning-based HARDI (hybrid automatic repeat-based) compressed sensing super-resolution reconstruction method | |
CN113379601A (en) | Real world image super-resolution method and system based on degradation variational self-encoder | |
CN110533591B (en) | Super-resolution image reconstruction method based on codec structure | |
CN109584162A (en) | A method of based on the image super-resolution reconstruct for generating network | |
CN111178499B (en) | Medical image super-resolution method based on generation countermeasure network improvement | |
CN111833261A (en) | Image super-resolution restoration method for generating countermeasure network based on attention | |
CN111353935A (en) | Magnetic resonance imaging optimization method and device based on deep learning | |
CN113284046A (en) | Remote sensing image enhancement and restoration method and network based on no high-resolution reference image | |
CN115880158A (en) | Blind image super-resolution reconstruction method and system based on variational self-coding | |
CN115760814A (en) | Remote sensing image fusion method and system based on double-coupling deep neural network | |
CN112365505A (en) | Lightweight tongue body segmentation method based on coding and decoding structure | |
CN117455770A (en) | Lightweight image super-resolution method based on layer-by-layer context information aggregation network | |
CN102298768B (en) | High-resolution image reconstruction method based on sparse samples | |
CN108267311A (en) | A kind of mechanical multidimensional big data processing method based on tensor resolution | |
CN112818777A (en) | Remote sensing image target detection method based on dense connection and feature enhancement | |
CN117173022A (en) | Remote sensing image super-resolution reconstruction method based on multipath fusion and attention | |
CN116664587A (en) | Pseudo-color enhancement-based mixed attention UNet ultrasonic image segmentation method and device | |
CN113554070B (en) | Bearing fault diagnosis method based on transition probability matrix and lightweight network | |
CN115375968A (en) | Fault diagnosis method for planetary gearbox |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |