Panchromatic sharpening method, storage medium and equipment

By building a generator and discriminator network, using adversarial learning and graph convolution to propagate high-frequency edge features, the problem of insufficient joint retention of hollow spectral information and high-frequency feature expression capabilities in the full-color sharpening method is solved, and higher quality is achieved Image fusion effect.

CN118941466APending Publication Date: 2024-11-12CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410943233.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing full-color sharpening methods have difficulties in the joint retention of null spectrum information and high-frequency feature expression capabilities, resulting in the loss of spatial spectrum information and poor quality of fused images.

Method used

Build generator networks and discriminator networks, optimize generator network outputs through adversarial learning, propagate high-frequency edge features using graph convolution, and introduce high-low frequency prior constraints, supplemented by spatial and spectral discriminators to preserve image texture details and spectral color.

Benefits of technology

Improves the spatial fusion effect, enhances the image color retention and detail clarity, and improves the quality of the fusion image.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941466A_ABST
    Figure CN118941466A_ABST
Patent Text Reader

Abstract

The invention discloses a panchromatic sharpening method, a storage medium and equipment, and relates to the technical field of image fusion, and the method comprises the steps: carrying out the size matching of an MS image and a PAN image, constructing a generator network, and generating an image G (I); the image G (I) is processed by an average pooling operator to obtain an image GPAN, the image G (I) is processed by a down-sampling operator to obtain an image GMS, the image GPAN and the image PAN are subjected to high-frequency extraction to obtain an image # imgabs0 # and an image PANhigh, and the image GMS and the image MS are subjected to low-frequency extraction to obtain an image # imgabs1 # and an image MSlow; inputting the GMS, the # imgabs2 # MS and the MSlow into a spectrum discriminator, and inputting the GPAN, the # imgabs3 # PAN and the PANhigh into a space discriminator; and the generator network and the discriminator network obtain images with high spatial resolution and high spectral resolution in the adversarial learning process. According to the method, the problems of difficulty in spatial-spectral information joint reservation and limited high-frequency feature expression capability can be solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image fusion, and in particular to a full color sharpening method, storage medium and device. Background Art

[0002] Multi-source remote sensing plays an increasingly important role in important fields such as precision agriculture, resource survey, environmental monitoring, military and national defense. Many remote sensing satellites have been launched by countries around the world, such as Quickbird, Worldview, Gaofen-1 and Gaofen-2. They are equipped with different optical remote sensing sensors to collect remote sensing images. However, due to the limitations of satellite payload, cost and bandwidth, existing satellite sensors cannot directly provide high spatial resolution multispectral (HRMS) images that contain high spatial resolution and high spectral resolution.

[0003] Panchromatic sharpening is a classic image fusion technique to solve this problem. It essentially utilizes the information complementarity between panchromatic (PAN) and multispectral (MS) images. This technology combines spatial information and spectral information to obtain an ideal high-resolution image. Traditional methods include Brovey, HPF, MTF-GLP, etc. They are widely used and highly efficient, but they have their inherent limitations: (1) they cannot fully express the underlying complex relationship between the two fields; (2) they are prone to spectral or spatial distortion.

[0004] Traditional deep learning-based methods can be divided into supervised and unsupervised methods. Supervised methods include PanNet, RED-cGAN, and PSGAN. Due to the lack of ideal ground truth images, these methods make a supervised setting based on the simulated training dataset of the Wald protocol. There is a gap between the downscaled fusion result and the fully scaled original image. In 2020, Ma proposed an unsupervised pan-sharpening framework based on generative adversarial networks. After that, PGMAN introduced the QNR loss and obtained a relatively ideal fused image. Unsupervised methods can learn directly from the fully scaled original image without any preprocessing, making them more practical. Existing unsupervised deep learning methods have achieved some goals, but there are still some limitations: implicit learning causes damage to spatial spectral information, and it is difficult to jointly preserve spatial spectral information. Most pan-sharpening methods in the unsupervised paradigm use the input image for latent feature learning, which inevitably leads to the loss of spatial spectral information and limited expression of high-frequency features. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of difficulty in jointly preserving spatial spectrum information and limited high-frequency feature expression capability in existing pan-sharpening methods, and to propose a pan-sharpening method, comprising the following steps:

[0006] S1, obtain PAN and MS images, match the size of the MS image with the PAN image, and obtain matched PAN and MS images;

[0007] S2. Construct a generator network, which includes an encoder, a feature separator, a graph structure, and a decoder. The matched PAN and MS images are input into the encoder for feature encoding to obtain high-dimensional features. The high-dimensional features are input into the feature separator to extract features of the two dimensions of space and channel, respectively, to obtain decoupled high-dimensional spatial features and high-dimensional channel features. The decoupled high-dimensional spatial features and high-dimensional channel features are input into the graph structure for feature relationship extraction and propagation to obtain feature representations of the two dimensions of space and channel. The decoder decodes and restores the feature representations of the two dimensions of space and channel to obtain the image G(I).

[0008] S3, image G(I) is compressed into image G after average pooling operator PAN , G(I) is obtained by downsampling operator to obtain image G MS , image G PAN The image is obtained by high frequency extraction operator Image G MS The image is obtained by low-frequency extraction operator The matched PAN image is obtained by high frequency extraction operator to obtain image PAN high , the matched MS image is extracted by low frequency operator to obtain image MS low ;

[0009] S4. Construct a discriminator network. The discriminator network consists of a spatial discriminator and a spectral discriminator. MS and images Matched MS image and image MS low Input spectral discriminator, image G PAN and images Matched PAN image and image PAN high Input spatial discriminator;

[0010] S5. The generator network and the discriminator network optimize the output of the generator network during the adversarial learning process to obtain images with high spatial resolution and high spectral resolution.

[0011] Furthermore, the size matching of the MS image and the PAN image is as follows:

[0012] The MS image and the PAN image are cropped, and the cropped MS image is upsampled to match the size with the PAN image.

[0013] Further,

[0014] The encoder-decoder is implemented based on the U-Net structure;

[0015] The high-dimensional features are decoupled into high-dimensional spatial features and high-dimensional channel features through point-by-point convolution and depth-wise convolution respectively;

[0016] Furthermore, the Sobel operator and graph convolution are used to construct the graph structure. x ,S y ) Extract edges of decoupled high-dimensional spatial features and high-dimensional channel features from the x and y directions, and construct an adjacency matrix of the graph structure:

[0017]

[0018] Among them, S x is the Sobel operator in the x direction, F s Represents high-dimensional space features, S y is the Sobel operator in the y direction, I s and I c are all unit matrices, representing the self-connection of nodes; ξ represents the downsampling operation in the channel dimension, F c represents high-dimensional channel features, represents the adjacency matrix of the spatial feature graph, Represents the adjacency matrix of the channel feature map;

[0019] The two-dimensional features of space and channel are expressed as follows:

[0020]

[0021]

[0022] in, represents the spatial dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation, η represents the nonlinear activation function, represents the degree matrix of the spatial dimension, Represents the characteristics of the spatial dimension of the lth layer of the graph structure, The weight matrix representing the change in spatial dimension during the training of the lth layer of the graph structure, It represents the channel dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation. represents the degree matrix of the channel dimension, (F c ) (l) Represents the characteristics of the channel dimension of the lth layer of the graph structure, The weight matrix representing the change in channel dimension during the training of the lth layer of the graph structure.

[0023] Furthermore, the discriminator network loss function is as follows:

[0024]

[0025] in, represents the loss of the spectral discriminator, represents the loss of the spatial discriminator, GP is the gradient penalty term, represents the channel connection operation, N represents the number of samples, and D l and D h are the outputs of the two discriminator networks, μ represents the weight factor of the gradient penalty term, (G MS ) (n) Indicates the nth G MS Image samples, Indicates the nth Image sample, (MS) (n) represents the nth MS image sample, (MS low ) (n) Indicates the nth MS low Image sample, (G PAN ) (n) Indicates the nth G PAN Image samples, Indicates the nth Image sample, (PAN) (n) Represents the nth PAN image sample, (PAN high ) (n) Indicates the nth PAN high Image samples.

[0026] Furthermore, the generator network loss function is as follows:

[0027]

[0028] Among them, L G represents the generator network loss, α and β represent weight factors, L Q Represents a loss designed based on the no-reference image quality metric QNR.

[0029] The present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned full-color sharpening method is implemented.

[0030] The present invention also proposes an electronic device, comprising a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned full-color sharpening method.

[0031] The beneficial effects brought by the technical solution provided by the present invention are:

[0032] The present invention proposes a full-color sharpening method, constructing a generator network and two discriminator networks, wherein the discriminator network includes a high-frequency enhanced spatial discriminator and a low-frequency enhanced spectral discriminator. The spatial discriminator assists the generator in generating images that retain the texture details of the PAN image; the spectral discriminator assists the generated images in retaining the spectral color of the MS image. In order to overcome the influence of implicit learning on the retention of spatial-spectral information, the decoupled frequency information is integrated into the adversarial learning process as an additional priori assumption and constraint. High frequency and low frequency can help the two discriminators distinguish the differences between the generated and full-resolution images in texture and color, thereby improving the network's discrimination ability. Graph convolution is used to propagate long-range relationships of high-frequency edge features, and ultimately a better spatial fusion effect is achieved. The full-color sharpening method proposed in the present invention solves the problems of difficulty in jointly retaining spatial-spectral information and limited expression ability of high-frequency features, and has more advantages in color retention, detail enhancement, and texture clarity, and the quality of the fused image is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of a full color sharpening method according to an embodiment of the present invention;

[0034] Figure 2 1 is an overall framework diagram of a generator network and a discriminator network according to an embodiment of the present invention;

[0035] Figure 3 is a block diagram of an electronic device in an exemplary embodiment of an embodiment of the present invention;

[0036] FIG4 is a comparison diagram of the effects of processing low-resolution multispectral images by the method of an embodiment of the present invention and other methods, wherein (a) is a low-resolution multispectral image acquired by the QuickBird satellite, (b) is a full-color image acquired by the QuickBird satellite, (c) is the image processed by multi-resolution analysis method on (a) in FIG4 , (d) is the image processed by principal component analysis method on FIG4 (a), and (e) is the image processed by the method described in this embodiment on FIG4 (a). DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0038] A flow chart of a full color sharpening method according to an embodiment of the present invention is as follows: Figure 1 , specifically including the following steps:

[0039] S1. Acquire PAN and MS images, match the sizes of the MS image and the PAN image, and obtain matched PAN and MS images.

[0040] PAN image (multispectral image) has low spatial resolution and high spectral resolution; MS image (panchromatic image) has high spatial resolution and low spectral resolution.

[0041] The images are physically cropped to images of corresponding scales. Specifically, the sizes of PAN and MS images of Gaofen-2 are approximately 30000×30000 and 7500×7500 pixels, respectively. The MS images are cropped to 64×64 and the PAN images are cropped to 256×256. The cropped MS images are then directly upsampled using bicubic interpolation to match the size of the corresponding PAN images.

[0042] S2. Construct a generator network, which includes an encoder, a feature separator, a graph structure, and a decoder. The matched PAN and MS images are input into the encoder for feature encoding to obtain high-dimensional features. The high-dimensional features are input into the feature separator to extract features of the two dimensions of space and channel respectively to obtain decoupled high-dimensional spatial features and high-dimensional channel features. The decoupled high-dimensional spatial features and high-dimensional channel features are input into the graph structure for feature relationship extraction and propagation to obtain feature representations of the two dimensions of space and channel. The decoder decodes and restores the feature representations of the two dimensions of space and channel to obtain the image G(I).

[0043] The encoder-decoder is implemented based on the U-Net structure. The four encoding blocks of the encoder network use dense connections, and through multiple jump connections from front to back, the information gradient flow within the network is finally improved. The convolution kernel of the first encoding block is slightly larger, which is used to roughly extract input features. The subsequent 3×3 filters are used to gradually receive and extract the deep feature map connected to the previous layer of channels from coarse to fine, and finally output the encoded high-dimensional features. The encoder maps information to high dimensions, which is conducive to better feature extraction. The purpose of multi-level step-by-step mapping is to extract deep features as smoothly as possible.

[0044] The high-dimensional features are decoupled by pointwise convolution (PW) and depthwise convolution (DW) to obtain the high-dimensional spatial features F. s and high-dimensional channel features F c The size of the DW convolution is k×k×1, with a total of C, where k represents the size of the convolution kernel and C represents the number of channels; the size of the PW convolution is 1×1×C, with a total of C.

[0045] Use Sobel operator and graph convolution to build graph structure. Sobel operator (S x ,S y ) Perform edge extraction on the decoupled high-dimensional spatial features and high-dimensional channel features from the x and y directions, and construct the adjacency matrix of the graph structure. The adjacency matrix is ​​the mathematical form of the graph topology.

[0046]

[0047] Among them, S x is the Sobel operator in the x direction, F s Represents high-dimensional space features, S y is the Sobel operator in the y direction, I s and I c are all unit matrices, representing the self-connection of nodes; ξ represents the downsampling operation of the channel dimension. In order to reduce the computational complexity of high-dimensional channels, the downsampling operation ξ of the channel dimension is used to give F c Matching feature size; F c represents high-dimensional channel features, represents the adjacency matrix of the spatial feature graph, Represents the adjacency matrix of the channel feature map;

[0048] The Sobel operator and graph convolution are used to construct a graph structure consisting of high-frequency edge information. The spatial and channel dimensional features are represented as follows:

[0049]

[0050]

[0051] in, represents the spatial dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation, η represents the nonlinear activation function, represents the degree matrix of the spatial dimension, Represents the characteristics of the spatial dimension of the lth layer of the graph structure, The weight matrix representing the change in spatial dimension during the training of the lth layer of the graph structure, It represents the channel dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation. represents the degree matrix of the channel dimension, (ξ(F c )) (1) Represents the characteristics of the channel dimension of the lth layer of the graph structure, The weight matrix representing the change in channel dimension during the training of the lth layer of the graph structure.

[0052] The output channel dimension feature representation and spatial dimension feature representation are added point by point and finally input into the decoder network to reconstruct the spatial detail map in the high-frequency domain to obtain an image with high spatial resolution and high spectral resolution.

[0053] S3, image G(I) is compressed into image G after average pooling operator PAN , G(I) is obtained by downsampling operator to generate the multispectral image G MS , image GPAN The high-resolution full-color image generated by the high-frequency extraction operator (HEO) Image G MS The low-resolution multispectral image generated by the low-frequency extraction operator The matched PAN image (real full-color (PAN) image) is obtained by high-frequency extraction operator to obtain the real high-resolution full-color image PAN high The matched MS image (real multispectral (MS) image) is obtained by low-frequency extraction operator (LEO) to obtain the real low-resolution multispectral image MS low .

[0054] Marked as a fake sample, {PAN after matching, PAN high}, {MS after matching, MS low} is marked as a real sample. By introducing the constraints of high and low frequency priors in the adversarial learning process, the following discriminator network is promoted to better distinguish between real samples and fake samples, and the joint optimization of spatial-spectral features is improved.

[0055] S4. Construct a discriminator network. The discriminator network consists of a high-frequency enhanced spatial discriminator and a low-frequency enhanced spectral discriminator. MS and images Matched MS image and image MS low Input low-frequency enhanced spectral discriminator, image G PAN and images Matched PAN image and image PAN high Enter the high-frequency enhanced spatial discriminator.

[0056] The overall framework diagram of the generator network and the discriminator network of the embodiment of the present invention is shown in FIG. Figure 2 .

[0057] S5. The generator network and the discriminator network optimize the output of the generator network during the adversarial learning process to obtain images with high spatial resolution and high spectral resolution.

[0058] During training, the discriminator network loss function is as follows:

[0059]

[0060] in, represents the spectral loss of low-frequency enhancement, i.e., the loss of the spectral discriminator, represents the spatial loss of high-frequency enhancement, that is, the loss of the spatial discriminator, GP is the gradient penalty term, represents the channel connection operation, N represents the number of samples, and D l and Dh are the outputs of the two discriminator networks, μ represents the weight factor of the gradient penalty term, (G MS ) (n) Indicates the nth G MS Image samples, Indicates the nth Image sample, (MS) (n) represents the nth MS image sample, (MS low ) (n) Indicates the nth MS low Image sample, (G PAN ) (n( Indicates the nth G PAN Image samples, Indicates the nth Image sample, (PAN) (n) Represents the nth PAN image sample, (PAN high ) (n) Indicates the nth PAN high Image samples.

[0061] The generator network loss function is as follows:

[0062]

[0063] Among them, L G represents the generator network loss, α represents a weight factor used to adjust the spectral discriminator D l The influence in the loss function; β represents another weight factor used to adjust the spatial discriminator D h The impact in the loss function; L Q Represents a loss designed based on the no-reference image quality metric QNR.

[0064] In an exemplary embodiment, a computer-readable storage medium is included, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned full-color sharpening method is implemented.

[0065] See also Figure 3 In an exemplary embodiment, an electronic device is also included, including at least one processor, at least one memory, and at least one communication bus.

[0066] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through a communication bus to execute the above-mentioned full-color sharpening method.

[0067] Specifically, multispectral remote sensing images and panchromatic remote sensing images from QuickBird satellite data are used to verify the method described in this embodiment. The resolution ratio of the panchromatic remote sensing image to the multispectral remote sensing image is 4:1, the size of the panchromatic remote sensing image is 256×256, and the size of the multispectral remote sensing image is 64×64.

[0068] FIG4(a) is a low-resolution multispectral image acquired by the QuickBird satellite, FIG4(b) is a panchromatic image acquired by the QuickBird satellite, FIG4(c) is an image processed by a multi-resolution analysis method, FIG4(d) is an image processed by a principal component analysis method, and FIG4(e) is an image processed by the method described in this embodiment.

[0069] The experimental results show that the multi-resolution analysis and principal component analysis methods have obvious deficiencies in detail preservation, the overall image is blurred, and there is a significant difference in spectral color from the MS image, and there are color casts and spectral mixing phenomena. However, the method described in this embodiment has the best spatial texture performance while maintaining acceptable color, especially in image details such as house corners, road edges, farmland textures, etc., which are closer to the original full-color image PAN.

[0070] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A panchromatic sharpening method, characterized in that: The following steps are involved: S1, obtain PAN and MS images, match the size of the MS image with the PAN image, and obtain matched PAN and MS images; S2. Construct a generator network, which includes an encoder, a feature separator, a graph structure, and a decoder. The matched PAN and MS images are input into the encoder for feature encoding to obtain high-dimensional features. The high-dimensional features are input into the feature separator to extract features of the two dimensions of space and channel, respectively, to obtain decoupled high-dimensional spatial features and high-dimensional channel features. The decoupled high-dimensional spatial features and high-dimensional channel features are input into the graph structure for feature relationship extraction and propagation to obtain feature representations of the two dimensions of space and channel. The decoder decodes and restores the feature representations of the two dimensions of space and channel to obtain the image G(I). S3, image G(I) is compressed into image G after average pooling operator PAN , G(I) is obtained by downsampling operator to obtain image G Ms , image G PAN The image is obtained by high frequency extraction operator Image G MS The image is obtained by low-frequency extraction operator The matched PAN image is obtained by high frequency extraction operator to obtain image PAN high , the matched MS image is extracted by low frequency operator to obtain image MS low ; S4. Construct a discriminator network. The discriminator network consists of a spatial discriminator and a spectral discriminator. MS and images Matched MS image and image MS low Input spectral discriminator, image G PAN and images Matched PAN image and image PAN high Input spatial discriminator; S5. The generator network and the discriminator network optimize the output of the generator network during the adversarial learning process to obtain images with high spatial resolution and high spectral resolution.

2. A full color sharpening method according to claim 1, characterized in that: Matching the size of MS image to PAN image is as follows: The MS image and the PAN image are cropped, and the cropped MS image is upsampled to match the size with the PAN image.

3. A full color sharpening method according to claim 1, characterized in that: The encoder-decoder is implemented based on the U-Net structure; The high-dimensional features are decoupled into high-dimensional spatial features and high-dimensional channel features through point-by-point convolution and depth-wise convolution respectively.

4. A full color sharpening method according to claim 1, characterized in that: Use Sobel operator and graph convolution to build graph structure. Sobel operator (S x ,S y ) Extract edges of decoupled high-dimensional spatial features and high-dimensional channel features from the x and y directions, and construct an adjacency matrix of the graph structure: Among them, S x is the Sobel operator in the x direction, F s Represents high-dimensional space features, S y is the Sobel operator in the y direction, I s and I c are all unit matrices, representing the self-connection of nodes; ξ represents the downsampling operation in the channel dimension, F c represents high-dimensional channel features, represents the adjacency matrix of the spatial feature graph, Represents the adjacency matrix of the channel feature map; The two-dimensional features of space and channel are expressed as follows: in, represents the spatial dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation, η represents the nonlinear activation function, represents the degree matrix of the spatial dimension, Represents the characteristics of the spatial dimension of the lth layer of the graph structure, The weight matrix representing the change in spatial dimension during the training of the lth layer of the graph structure, It represents the channel dimension feature representation of the output of the l+1th layer after the graph structure relationship aggregation. represents the degree matrix of the channel dimension, (ξ(F c )) (l) Represents the characteristics of the channel dimension of the lth layer of the graph structure, The weight matrix representing the change in channel dimension during the training of the lth layer of the graph structure.

5. A full color sharpening method according to claim 1, characterized in that: The discriminator network loss function is as follows: in, represents the loss of the spectral discriminator, represents the loss of the spatial discriminator, GP is the gradient penalty term, represents the channel connection operation, N represents the number of samples, and D l and D h are the outputs of the two discriminator networks, μ represents the weight factor of the gradient penalty term, (G MS ) (n) Indicates the G MS Image samples, Indicates the Image sample, (MS) (n) represents the th MS image sample, (MS low ) (n) Indicates the MS low Image sample, (G PAN ) (n) Indicates the G PAN Image samples, Indicates the Image sample, (PAN) (n) represents the PAN image sample, (PAN high ) (n) Indicates the PAN high Image samples.

6. A panchromatic sharpening method according to claim 5, characterized in that: The generator network loss function is as follows: Among them, L G represents the generator network loss, α and β represent weight factors, L Q Represents a loss designed based on the no-reference image quality metric QNR.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1 to 6.

Citation Information

Cited By

  • Multispectral remote sensing image panchromatic sharpening method based on frequency domain decomposition network

    CN119579847A

  • Blurred image sharpening method, system, equipment and medium

    CN120807354A