Content-assured recurrent adversarial network finger OCT volume data denoising method

Abstract

A content-guaranteed recurrent generative adversarial network (RBG) method for denoising finger OCT volume data includes the following steps: 1) creating a high-quality, clean finger B-Scan image manually; 2) constructing a content-guaranteed recurrent generative adversarial network (RBG) model, setting parameters, and training it; 3) using the trained convolutional neural network model to denoise the OCT volume data. This invention can train a content-guaranteed recurrent generative adversarial network model using only a clean image and effectively denoise finger OCT volume data.

Description

Technical Field

[0001] This invention relates to the field of OCT image denoising, and in particular to a content-guaranteed recurrent adversarial network (ROA) method for denoising finger OCT volume data. Background Technology

[0002] With the increasing awareness of protecting personal identity information, biometric identification technology has also made great strides, and fingerprints have become the mainstream method of individual identification and authentication due to their ease of collection, high uniqueness and resistance to change.

[0003] Optical coherence tomography (OCT) is a novel bio-imaging technology that combines optics, computer science, and many other disciplines. It can acquire information up to 3 millimeters below the epidermis. Due to its non-invasive, high-resolution, and non-destructive advantages, OCT technology has begun to be used for fingerprint identification. Using dermal information acquired through OCT for fingerprint recognition can significantly reduce the impact of the external environment on fingerprint identification, improving security. However, due to the imaging principle of OCT, optical interference inevitably occurs when incident light encounters the internal tissue particles of the finger, forming irregular granular speckle noise in the generated scanned image. This noise can obscure important subcutaneous tissue structures needed for identification. Therefore, noise reduction is necessary to utilize finger OCT volume data for fingerprint recognition.

[0004] Cycle-Consistent Generative Adversarial Networks (CycleGANs) are adversarial neural networks that utilize unpaired datasets for image style transfer. Since noisy and noiseless images belong to two style domains, CycleGANs can be used to denoise images.

[0005] Because clean samples of OCT images are difficult to obtain and their numbers are far fewer than those of noisy samples, their content is difficult to guarantee. Summary of the Invention

[0006] This invention aims to overcome the aforementioned shortcomings of existing technologies by proposing a content-guaranteed recurrent adversarial network (ROV) method for denoising finger OCT volume data. This invention can train a content-guaranteed RRV model using only a clean B-Scan sample of a finger OCT, and then utilize the trained model to effectively denoise the finger OCT volume data.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A content-guaranteed recurrent adversarial network (ROA) finger OCT volume data denoising method includes the following steps:

[0009] 1) Create a high-quality, clean finger B-Scan image manually;

[0010] 2) Construct a content-guaranteed recurrent generative adversarial network model, determine the training process, set parameters, and conduct training;

[0011] 3) Use the trained convolutional neural network model to denoise the OCT volume data.

[0012] Furthermore, in step 1), the creation of a high-quality clean finger B-Scan image includes the following steps:

[0013] (11) First, perform simple filtering on the noisy finger B-Scan image to obtain a relatively clean finger B-Scan image;

[0014] (12) Among relatively clean finger B-Scan images, the cleanest image is selected by calculating its Peak Signal-to-Noise Ratio (PSNR) for artificial enhancement. The main contents include removing residual speckle particles and restoring the continuity between tissues; increasing the brightness of sweat gland and dermal contours; erasing all tissues between the epidermis and dermis of the finger except for sweat glands, highlighting sweat gland, dermal contours and epidermal contours, minimizing misjudgments by the reconstruction algorithm, and finally obtaining a high-quality clean finger B-Scan image for training.

[0015] Furthermore, step 2) includes the following steps:

[0016] (21) The content-guaranteed recurrent generative adversarial network model mainly consists of three parts: a generator, a discriminator, and a content loss network. The generator is responsible for image generation. It first downsamples the input image. The downsampling layer first fills the input image. The first time, it passes through 64 7×7 convolutional kernels with a stride of 1, an instance normalization layer, and a ReLU activation function to obtain a 64-channel feature map with the same dimensions as the original image. The second time, it passes through 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. The third time, it passes through 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. The feature map obtained from the downsampling layer is then passed through a transition layer containing six residual blocks, each with the same structure. The input image is first convolved with 256 3×3 kernels with a stride of 2, followed by a normalization layer and a ReLU activation function, then convolved again with 256 3×3 kernels with a stride of 2 and a normalization layer, before being connected to the original input residuals to complete feature fusion. Finally, the result from the transition layer is upsampled, with the upsampling process being the exact reverse of the downsampling process, restoring the image to its input state. The discriminator is responsible for identifying fake images generated by the generator. It first extracts features from the image. The first feature extraction layer consists of 64 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The second layer consists of 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The third layer consists of 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The fourth layer consists of 521 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function, resulting in a 512-channel feature map. This feature map is then passed through a fully connected layer with a stride of 2 and a 3×3 convolutional kernel to obtain the probability that the input image is a real image. The output value is then controlled between 0 and 1 using a sigmoid function. The content loss network is responsible for extracting the content information of the input noisy image and the content information of the image generated by the generator, respectively. The first layer consists of 64 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2, the second layer consists of 128 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2, and the third layer consists of 256 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2. Finally, a 256-channel feature map is output for calculating the content loss.

[0017] (22) Prepare two sets of training data. One set is a large number of noisy OCT finger B-Scan images collected by OCT without processing, and the other set is a high-quality clean finger B-Scan image obtained in step 1).

[0018] (23) Determine the training process of the content-guaranteed recurrent adversarial network. The content-guaranteed recurrent adversarial network component includes a dual generator and a dual discriminator. The generator G AB Generator G is responsible for generating a clean image from a noisy one. BA The discriminator D is responsible for generating a noisy image from a clean one. A The discriminator D is responsible for determining whether the generated clean image is real. B It is responsible for determining whether the generated noisy image is real. The content-guaranteed recurrent adversarial network is trained 11 times, with each training session containing a different number of rounds. The first 10 training sessions consist of only one round each, while the 11th training session consists of 20 rounds.

[0019] (24) The first training uses 1 clean image and n noisy images. The single training process is as follows: the n noisy images first pass through the generator G. AB A synthesized low-noise image is obtained. The synthesized low-noise image and the original noisy image are then passed through a content loss network to obtain their respective content feature maps. The average error between the two is calculated to obtain the content loss. The synthesized low-noise image is then passed through a G network... BA The synthesized noisy image is obtained, and then the average error between the original noisy image and the synthesized noisy image is calculated to obtain the cycle consistency loss. Discriminator D A The synthesized low-noise image is then evaluated to obtain the GAN loss. A clean image first passes through the generator G... BA The synthesized noisy image is obtained, and then the synthesized noisy image is processed by G. AB A synthesized clean image is obtained, and then the average error between the original clean image and the synthesized clean image is calculated to obtain the cycle consistency loss. Discriminator D B The synthesized low-noise image is then evaluated to obtain the adversarial loss.

[0020] (25) From the 2nd to the 10th training iterations, n noisy images are first processed by the generator G. AB n synthesized low-noise images are obtained. The differences between the synthesized low-noise images are compared by the structural similarity index (SSIM). The number of training times is represented by i. The n / (11-i) images with large structural differences are selected and added to the clean sample set for the i-th training. The training process of a single round is the same as step (24). While continuously enriching the clean image sample set, the quality of the clean image sample set is improved.

[0021] (26) During the 11th training, the dataset was enriched to include n noisy images and n′ clean images. The recurrent generative adversarial network model with guaranteed input content was trained for 20 rounds until the loss converged. The training process for a single round was the same as step (24).

[0022] (27) Determine the parameters of the lightweight 3D convolutional neural network, load the images in the training set into the model for training, and in step (24), the content loss function is the average error between the feature maps obtained by the content loss network from the synthesized low-noise image and the original noisy image:

[0023]

[0024] in: Represents a noisy image. Representative generator G AB A synthesized clean image, Y i Synthesized low-noise image feature map, X i This is the feature map of the original noisy image.

[0025] The cycle consistency loss function in step (24) includes the noisy image and G. BA The average error between the generated synthetic noisy images and the clean image and G AB Average error between the generated synthetic low-noise images:

[0026]

[0027] in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G AB The generated synthetic low-noise image.

[0028] The counter-loss in step (24) includes G AB The mapping adversarial loss and G between the generated synthetic low-noise image and the real clean image BA Adversarial loss between the generated synthetic noise image and the real noise image:

[0029]

[0030] in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G ABThe generated synthetic low-noise image. It is discriminator D A For G AB The discrimination results of the generated low-noise map. It is discriminator D B For G BA The discrimination results of the generated noisy image.

[0031] The process of step 3) is as follows: Since the original 1800×500 finger B-Scan image acquired by OCT contains a large number of black areas, a 256×256 window B-Scan image is cropped to obtain a series of sub-images, which are then used as input to the network, in order to improve the attention of the network training. When denoising the finger body data using the trained model, the body data is also first cropped and input into the network. Finally, the denoised sub-images are reassembled into 1800×500×1400 image data.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A network model that can effectively remove OCT fingertip data can be trained using only one clean image; 2. The denoising effect is better than traditional methods and the speed is faster. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the network architecture of the method of the present invention;

[0034] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0036] See Figures 1-2 A content-guaranteed recurrent adversarial network (ROA) finger OCT volume data denoising method includes the following steps:

[0037] 1) Creating a high-quality, clean finger B-Scan image involves the following steps:

[0038] (11) Perform preliminary filtering on noisy finger B-Scan images to obtain relatively clean finger B-Scan images;

[0039] (12) Among relatively clean finger B-Scan images, the cleanest image is selected by calculating its Peak Signal-to-Noise Ratio (PSNR) for artificial enhancement. The main contents include removing residual speckle particles and restoring the continuity between tissues; increasing the brightness of sweat gland and dermal contours; erasing all tissues between the epidermis and dermis of the finger except for sweat glands, highlighting sweat glands, dermal contours and epidermal contours, minimizing the misjudgment of the reconstruction algorithm, and finally obtaining a high-quality clean finger B-Scan image for training;

[0040] 2) See Figure 2 The construction of a content-guaranteed recurrent generative adversarial network model, the determination of the training process, the setting of parameters, and the training process include the following steps:

[0041] (21) The content-guaranteed recurrent generative adversarial network model mainly consists of three parts: a generator, a discriminator, and a content loss network. The generator is responsible for image generation. It first downsamples the input image. The downsampling layer first fills the input image. The first time, it passes through 64 7×7 convolutional kernels with a stride of 1, an instance normalization layer, and a ReLU activation function to obtain a 64-channel feature map with the same dimensions as the original image. The second time, it passes through 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. The third time, it passes through 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. The feature map obtained from the downsampling layer is then passed through a transition layer containing six residual blocks, each with the same structure. The input image is first convolved with 256 3×3 kernels with a stride of 2, followed by a normalization layer and a ReLU activation function, then convolved again with 256 3×3 kernels with a stride of 2 and a normalization layer, before being connected to the original input residuals to complete feature fusion. Finally, the result from the transition layer is upsampled, with the upsampling process being the exact reverse of the downsampling process, restoring the image to its input state. The discriminator is responsible for identifying fake images generated by the generator. It first extracts features from the image. The first feature extraction layer consists of 64 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The second layer consists of 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The third layer consists of 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The fourth layer consists of 521 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function, resulting in a 512-channel feature map. This feature map is then passed through a fully connected layer with a stride of 2 and a 3×3 convolutional kernel to obtain the probability that the input image is a real image. The output value is then controlled between 0 and 1 using a sigmoid function. The content loss network is responsible for extracting the content information of the input noisy image and the content information of the image generated by the generator, respectively. The first layer consists of 64 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2, the second layer consists of 128 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2, and the third layer consists of 256 3×3 convolutional kernels with a stride of 1, ReLU activation function and Max Pooling layer with a stride of 2. Finally, a 256-channel feature map is output for calculating the content loss.

[0042] (22) Prepare two sets of training data. One set is a large number of noisy OCT finger B-Scan images collected by OCT without processing, and the other set is a high-quality clean finger B-Scan image obtained in step 1).

[0043] (23) Determine the training process of the content-guaranteed recurrent adversarial network. The content-guaranteed recurrent adversarial network component includes a dual generator and a dual discriminator. The generator G AB Generator G is responsible for generating a clean image from a noisy one. BA The discriminator D is responsible for generating a noisy image from a clean one. A The discriminator D is responsible for determining whether the generated clean image is real. B It is responsible for determining whether the generated noisy image is real. The content-guaranteed recurrent adversarial network is trained 11 times, with each training session containing a different number of rounds. The first 10 training sessions consist of only one round each, while the 11th training session consists of 20 rounds.

[0044] (24) The first training uses 1 clean image and n noisy images. The single training process is as follows: the n noisy images first pass through the generator G. AB A synthesized low-noise image is obtained. The synthesized low-noise image and the original noisy image are then passed through a content loss network to obtain their respective content feature maps. The average error between the two is calculated to obtain the content loss. The synthesized low-noise image is then passed through a G network... BA The synthesized noisy image is obtained, and then the average error between the original noisy image and the synthesized noisy image is calculated to obtain the cycle consistency loss. Discriminator D A The synthesized low-noise image is then evaluated to obtain the GAN loss. A clean image first passes through the generator G... BA The synthesized noisy image is obtained, and then the synthesized noisy image is processed by G. AB A synthesized clean image is obtained, and then the average error between the original clean image and the synthesized clean image is calculated to obtain the cycle consistency loss. Discriminator D B The synthesized low-noise image is then evaluated to obtain the adversarial loss.

[0045] (25) From the 2nd to the 10th training iterations, n noisy images are first processed by the generator G. AB n synthesized low-noise images are obtained. The differences between the synthesized low-noise images are compared by the structural similarity index (SSIM). The number of training times is represented by i. The n / (11-i) images with large structural differences are selected and added to the clean sample set for the i-th training. The training process for a single round is the same as step (24). While continuously enriching the clean image sample set, the quality of the clean image sample set is improved.

[0046] (26) During the 11th training, the dataset is enriched to include n noisy images and n′ clean images. The input content guarantees the training of the recurrent generative adversarial network model for 20 rounds until the loss converges. The training process for a single round is the same as step (24).

[0047] (27) Determine the parameters of the lightweight 3D convolutional neural network, load the images in the training set into the model for training, and in step (23), the content loss function is the average error between the feature maps obtained by the content loss network from the synthesized low-noise image and the original noisy image:

[0048]

[0049] in: Represents a noisy image. Representative generator G AB A synthesized clean image, Y i Synthesized low-noise image feature map, X i This is the feature map of the original noisy image.

[0050] The cycle consistency loss function in step (23) includes the noisy image and G. BA The average error between the generated synthetic noisy images and the clean image and G AB Average error between the generated synthetic low-noise images:

[0051]

[0052] in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G AB The generated synthetic low-noise image.

[0053] The counter-loss in step (23) includes G AB The mapping adversarial loss and G between the generated synthetic low-noise image and the real clean image BA Adversarial loss between the generated synthetic noise image and the real noise image:

[0054]

[0055] in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G ABThe generated synthetic low-noise image. It is discriminator D A For G AB The discrimination results of the generated low-noise map. It is discriminator D B For G BA The discrimination results of the generated noisy image;

[0056] 3) The steps for denoising OCT volume data using a trained convolutional neural network model are as follows: A 256×256 window B-Scan image is cropped to obtain a series of sub-images, which are then used as input to the network. When denoising finger volume data using the trained model, the volume data is also first cropped and input into the network. Finally, the denoised sub-images are reassembled into a 1800×500×1400 image.

[0057] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A content-guaranteed recurrent adversarial network (ROA) finger OCT volume data denoising method, comprising the following steps: 1) Create a high-quality, clean finger B-Scan image manually; specifically including: (11) First, perform simple filtering on the noisy finger B-Scan image to obtain a relatively clean finger B-Scan image; (12) In the finger B-Scan images obtained by simple filtering, the cleanest image is selected by calculating its peak signal-to-noise ratio (PSNR) for artificial enhancement. This includes: removing residual speckle particles and repairing the continuity between tissues; increasing the brightness of the sweat gland and dermal contours; erasing all tissues between the finger epidermis and dermis except for the sweat glands, highlighting the sweat gland, dermal contour and epidermal contour, minimizing the misjudgment of the reconstruction algorithm, and finally obtaining a high-quality clean finger B-Scan image for training. 2) Construct a content-guaranteed recurrent generative adversarial network (RBAN) model, determine the training process, set parameters, and train it; input clean finger B-Scan images and noisy finger B-Scan images into the content-guaranteed RBAN model for training until the loss converges; the content-guaranteed RBAN model consists of three parts: a generator, a discriminator, and a content loss network. The generator is responsible for generating images, the discriminator is responsible for identifying fake images generated by the generator, and the content loss network is responsible for extracting the content information of the input noisy images and the content information of the images generated by the generator, and calculating the content loss; 3) Use the trained convolutional neural network model to denoise the OCT volume data.

2. The content-guaranteed recurrent generative adversarial network finger OCT volume data denoising method according to claim 1, characterized in that, Step 2) specifically includes: (21) The content-guaranteed recurrent generative adversarial network model consists of three parts: a generator, a discriminator, and a content loss network. The generator generates an image and first downsamples the input image. The downsampling layer first fills the input image. The first time, it passes through 64 7×7 convolutional kernels with a stride of 1, an instance normalization layer, and a ReLU activation function to obtain a 64-channel feature map with the same dimensions as the original image. The second time, it passes through 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. The third time, it passes through 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a ReLU activation function to obtain a 256-channel feature map. Then, the feature map obtained from the downsampling layer is... The map layer uses a transition layer containing six residual blocks, each with the same structure. The input image is first convolved with 256 3×3 convolutional kernels with a stride of 2, followed by a normalization layer and a ReLU activation function. Then, it is convolved again with 256 3×3 convolutional kernels with a stride of 2 and a normalization layer. Finally, it is concatenated with the original input residuals to complete feature fusion. The result from the transition layer is then upsampled, the upsampling process being the exact reverse of the downsampling process, restoring the image to its input state. The discriminator is responsible for identifying fake images generated by the generator. It first extracts features from the image, specifically the extraction layer: Layer 1 includes 64 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function; Layer 2 includes 128 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function; Layer 3 includes 256 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function. The ReLU activation function is used in the fourth layer, which includes 521 3×3 convolutional kernels with a stride of 2, an instance normalization layer, and a Leak ReLU activation function, resulting in a 512-channel feature map. This feature map is then passed through a 3×3 convolutional kernel with a stride of 2 and a fully connected layer to obtain the probability that the input image is a real image. A sigmoid function is used to control the output value between 0 and 1. The content loss network is responsible for extracting the content information from the input noisy image and the generator-generated image. The first layer consists of 64 3×3 convolutional kernels with a stride of 1, a ReLU activation function, and a Max Pooling layer with a stride of 2. The second layer consists of 128 3×3 convolutional kernels with a stride of 1, a ReLU activation function, and a Max Pooling layer with a stride of 2. The third layer consists of 256 3×3 convolutional kernels with a stride of 1, a ReLU activation function, and a Max Pooling layer with a stride of 2, ultimately outputting a 256-channel feature map for calculating the content loss. (22) Prepare two sets of training data. One set is a large number of noisy OCT finger B-Scan images collected by OCT without processing, and the other set is a high-quality clean finger B-Scan image obtained in step 1). (23) Determine the training process of the content-guaranteed recurrent adversarial network. The content-guaranteed recurrent adversarial network component includes a dual generator and a dual discriminator. The generator G AB Generator G is responsible for generating a clean image from a noisy one. BA The discriminator D is responsible for generating a noisy image from a clean one. A The discriminator D is responsible for determining whether the generated clean image is real. B The content-guaranteed recurrent adversarial network is responsible for determining whether the generated noisy image is real. It is trained 11 times, with each training session containing a different number of rounds. The first 10 training sessions consist of only 1 round each, while the 11th training session consists of 20 rounds. (24) The first training uses 1 clean image and n noisy images. The single training process is as follows: the n noisy images first pass through the generator G. AB A synthesized low-noise image is obtained. The synthesized low-noise image and the original noisy image are then passed through a content loss network to obtain their respective content feature maps. The average error between the two is calculated to obtain the content loss. The synthesized low-noise image is then passed through a G network... BA The synthesized noisy image is obtained, and then the average error between the original noisy image and the synthesized noisy image is calculated to obtain the cycle consistency loss; discriminator D A The synthesized low-noise image is then evaluated to obtain the GAN loss; a clean image first passes through the generator G... BA The synthesized noisy image is obtained, and then the synthesized noisy image is processed by G. AB A synthesized clean image is obtained, and then the average error between the original clean image and the synthesized clean image is calculated to obtain the cycle consistency loss; discriminator D B The synthesized low-noise image is then evaluated to obtain the adversarial loss. (25) From the 2nd to the 10th training iterations, n noisy images are first processed by the generator G. AB n synthesized low-noise images are obtained. The differences between the synthesized low-noise images are compared by the structural similarity index (SSIM). The number of training times is represented by i. The n / (11-i) images with large structural differences are selected and added to the clean sample set for the i-th training. The training process of a single round is the same as step (24). While continuously enriching the clean image sample set, the quality of the clean image sample set is improved. (26) During the 11th training, the dataset was enriched to include n noisy images and n′ clean images. The recurrent generative adversarial network model with guaranteed input content was trained for 20 rounds until the loss converged. The training process for a single round was the same as step (24). (27) Determine the parameters of the lightweight 3D convolutional neural network, load the images in the training set into the model for training, and in step (24), the content loss function is the average error between the feature maps obtained by the content loss network from the synthesized low-noise image and the original noisy image: in: Represents a noisy image. Representative generator G AB A synthesized clean image, Y i Synthesized low-noise image feature map, X i This is the feature map of the original noisy image; The cycle consistency loss function in step (24) includes the noisy image and G. BA The average error between the generated synthetic noisy images and the clean image and G AB Average error between the generated synthetic low-noise images: in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G AB The generated synthetic low-noise image; The counter-loss in step (24) includes G AB The mapping adversarial loss and G between the generated synthetic low-noise image and the real clean image BA Adversarial loss between the generated synthetic noise image and the real noise image: in: Represents a noisy image. Representing G BA The generated synthetic noise image, Represents a clean image. Representing G AB The generated synthetic low-noise image; It is discriminator D A For G AB The discrimination results of the generated low-noise map. It is discriminator D B For G BA The discrimination results of the generated noisy image.

3. The content-guaranteed recurrent generative adversarial network finger OCT volume data denoising method according to claim 2, characterized in that, Step 3) includes the following process: Because the original 1800×500 finger B-Scan image acquired by OCT contains a large number of black areas, a 256×256 window B-Scan image was cropped to obtain a series of sub-images, which were then used as input to the network in order to improve the attention of the network training. When denoising the finger body data using the trained model, the body data was also cropped and input into the network first. Finally, the denoised sub-images were re-stitched into 1800×500×1400 image data.

Images