A deep learning denoising method for single hyperspectral imaging

Abstract

The application provides a kind of deep learning denoising method for single hyperspectral imaging, comprising the following steps: undersampling for hyperspectral data containing noise, obtaining undersampling hyperspectral imaging containing noise;Model training, undersampling hyperspectral imaging is placed at the input end and output end of the model respectively, the model is mainly composed of two parts, namely encoder and decoder;The input end of the encoder and the output end of the decoder are both images containing noise, by using mean square error as loss function, the model is trained to convergence, and a trained model is obtained;Denoising using the model, input multiple undersampling hyperspectral imaging containing noise into the trained model, and output the undersampling image after denoising;The results of the undersampling image after denoising are accumulated and added to form the hyperspectral imaging after denoising;The method provided by the application can construct a deep learning denoising model only with a single hyperspectral image, and has good denoising performance.

Description

Technical Field

[0001] This invention relates to the field of image denoising, and in particular to a deep learning denoising method for single hyperspectral imaging. Background Technology

[0002] In non-in-situ modes, the characterization of reaction processes often lacks accuracy and precision, especially at crucial reaction intermediates where significant information loss occurs, leading to substantial discrepancies between the characterized results and the actual process. In recent years, in-situ process research has become a challenging and crucial area of ​​study in fundamental research fields such as surface science and biomedicine. Raman spectroscopy, a highly sensitive imaging technique capable of providing material fingerprint information, has been widely applied in in-situ process research due to its non-invasive advantage. However, in-situ process research typically requires high time resolution, while Raman imaging results acquired over short time periods have low signal-to-noise ratios, creating a clear contradiction. Generally, acquiring sufficiently high-quality Raman imaging results in a short time can be approached from two aspects: 1) From a hardware perspective, improving instrument design to enhance imaging quality, such as optimizing the optical path and increasing light throughput to improve the signal-to-noise ratio of the Raman signal; 2) From a software perspective, post-processing the imaging results using denoising algorithms to improve imaging quality. Clearly, denoising algorithms, with their advantages of low cost, high versatility, and high performance, play a key role in improving imaging quality.

[0003] It is important to note that Raman imaging, as a hyperspectral imaging technique, combines the characteristics of both spectra and images. Each Raman image contains not only image information between pixels, but each pixel also corresponds to all the detailed information of a Raman spectrum. Therefore, depending on the object to be denoised, Raman imaging denoising algorithms can be divided into two categories: 1) Spectral denoising, which involves denoising the Raman spectrum corresponding to each pixel in the Raman image to achieve final image denoising. This type of method is mostly based on classical denoising algorithms, such as SG filtering; 2) Image denoising, which involves directly denoising the image result. This type of method is mostly based on machine learning algorithms, such as the BM3D algorithm. While machine learning algorithms have good denoising performance, they usually require a reference image with a certain high signal-to-noise ratio to be effective. For in-situ Raman imaging, the imaging results have significant time sensitivity, and the algorithm should not interact with other imaging processes to ensure the accuracy of the denoising results. Therefore, this approach cannot meet the above requirements. Summary of the Invention

[0004] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and propose a deep learning denoising method for single hyperspectral imaging. This method requires only a single hyperspectral image to construct a deep learning denoising model and has excellent denoising performance.

[0005] The present invention adopts the following technical solution:

[0006] A deep learning-based denoising method for single hyperspectral images includes the following steps:

[0007] Undersampling is performed on noisy hyperspectral data, specifically as follows: the wavenumber range for imaging in the hyperspectral data is calibrated according to requirements; a fixed number of points are undersampled within the band range; the spectral intensities corresponding to the undersampled points are summed to generate the value of a single pixel in the Raman imaging; the operation is repeated to obtain noisy undersampled hyperspectral images.

[0008] Model training involves placing undersampled hyperspectral images at the input and output of the model, respectively. The model consists of two main parts: an encoder and a decoder. Both the input of the encoder and the output of the decoder contain noisy images. By using the mean square error as the loss function, the model is trained until convergence, resulting in a well-trained model.

[0009] The model-based denoising method involves inputting multiple noisy undersampled hyperspectral images into a trained model. The model denoises the input undersampled hyperspectral images and outputs the denoised undersampled images. The denoised undersampled image results are then summed to form the denoised hyperspectral image.

[0010] Specifically, the number of fixed points is between 1 / 4 and 1 / 6 of the total number of points.

[0011] Specifically, the model is the U-net model.

[0012] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0013] This invention provides a deep learning-based denoising method for single hyperspectral imaging, comprising the following steps: undersampling the noisy hyperspectral data, specifically: calibrating the wavenumber range for imaging in the hyperspectrum according to requirements; undersampling a fixed number of points within the waveband range; summing the spectral intensities corresponding to the undersampled points to generate the value of a single pixel in the hyperspectral imaging; repeating the operation to obtain the noisy undersampled hyperspectral image; and model training, specifically: placing the undersampled hyperspectral image at the input and output of the model, which mainly consists of two parts, namely, encoding... The method involves an encoder and a decoder. Both the input and output of the encoder and decoder are noisy images. The model is trained to convergence using the mean square error as the loss function, resulting in a well-trained model. Denoising is achieved by inputting multiple noisy undersampled hyperspectral images into the trained model. The model denoises the input undersampled Raman images and outputs a denoised undersampled image. The denoised undersampled image results are then summed to form a denoised hyperspectral image. This invention provides a method that requires only a single hyperspectral image to construct a deep learning denoising model and exhibits excellent denoising performance. Attached Figure Description

[0014] Figure 1 A flowchart of a deep learning denoising method for a single hyperspectral image provided in an embodiment of the present invention;

[0015] Figure 2 This is a diagram of the U-net network model architecture provided in an embodiment of the present invention;

[0016] Figure 3 A schematic diagram of the original image and the denoised image of the surface-enhanced Raman spectrum with different spectral peaks on a two-dimensional material provided in an embodiment of the present invention;

[0017] Figure 4 This is a schematic diagram of the original image and the denoised image of the tip-enhanced Raman spectrum of different spectral peaks on a two-dimensional material provided in an embodiment of the present invention.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0019] This invention provides a deep learning denoising method for single hyperspectral imaging, which can construct a deep learning denoising model with only a single hyperspectral image and has good denoising performance.

[0020] like Figure 1 The following is a flowchart of a deep learning denoising method for a single hyperspectral image provided by an embodiment of the present invention, including the following steps:

[0021] S101: Undersampling is performed on noisy hyperspectral data. Specifically, the wavenumber range for imaging in the hyperspectral data is calibrated according to the requirements; a fixed number of points are undersampled within the band range; the spectral intensities corresponding to the undersampled points are summed to generate the value of a single pixel in the hyperspectral imaging; the operation is repeated to obtain noisy undersampled hyperspectral images.

[0022] The wavenumber range is selected based on requirements. Different wavenumber ranges reflect different chemical information. Specifically, the fixed number of points is one-fifth of the total number of points. The purpose of undersampling is to obtain images with different noise components for subsequent model training. Using this method, even images with different noise distributions and consistent subject information, or even two images with noise, can be used to train a denoising network.

[0023] S102: Model training, specifically: undersampled hyperspectral images are placed at the input and output of the model, respectively. This model mainly consists of two parts: an encoder and a decoder. Both the input of the encoder and the output of the decoder contain noisy images. The model is trained until convergence using the mean square error as the loss function, resulting in a trained model. Specifically, the model is a U-net model, such as... Figure 2 This is a diagram of the U-net network model architecture provided in an embodiment of the present invention.

[0024] S103: Denoising using a model, specifically: inputting multiple noisy undersampled hyperspectral images into a trained model, the model denoising the input undersampled hyperspectral images and outputting the denoised undersampled images; accumulating and summing the denoised undersampled image results to form the denoised hyperspectral image.

[0025] Example 1

[0026] like Figure 3 As shown, this invention can effectively improve the signal-to-noise ratio of surface-enhanced Raman spectroscopy imaging. Different scanning times, 0.1s and 0.5s, were used at a depth of 137cm. -1 and 255cm -1 Four Raman images with different signal-to-noise ratios were obtained by imaging the spectral peaks and then input into the model constructed in this invention. The results are shown in the figure. This invention exhibits good denoising performance in imaging at different signal-to-noise ratios and demonstrates good fidelity when matched with high signal-to-noise ratio imaging and reference images.

[0027] Example 2

[0028] like Figure 4 As shown, the method of this invention can be used for tip-enhanced Raman spectroscopy imaging, effectively improving the signal-to-noise ratio of the image. (The image is located at 1580 cm⁻¹.) -1and 1990cm -1 Two spectral peaks with different signal-to-noise ratios (SNRs) were imaged to obtain two Raman images with different SNRs, and these results were input into the model constructed in this invention. The results are shown in the figure. This invention exhibits good denoising performance in imaging at different SNRs and demonstrates good fidelity when matched with high SNR imaging and reference images.

[0029] This invention provides a deep learning-based denoising method for single hyperspectral imaging, comprising the following steps: undersampling the noisy hyperspectral data, specifically: calibrating the wavenumber range for imaging in the hyperspectrum according to requirements; undersampling a fixed number of points within the waveband range; summing the spectral intensities corresponding to the undersampled points to generate the value of a single pixel in the hyperspectral imaging; repeating the operation to obtain the noisy undersampled hyperspectral image; and model training, specifically: placing the undersampled hyperspectral image at the input and output of the model, which mainly consists of two parts, namely encoding... The encoder and decoder are both input images containing noise. By using the mean squared error as the loss function, the model is trained until convergence, resulting in a trained model. Denoising is achieved using the model by inputting multiple noisy undersampled hyperspectral images into the trained model. The model denoises the input undersampled hyperspectral images and outputs a denoised undersampled image. The denoised undersampled image results are then summed to form a denoised hyperspectral image. The method provided by this invention requires only a single hyperspectral image to construct a deep learning denoising model and exhibits excellent denoising performance.

[0030] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A deep learning-based denoising method for single hyperspectral imaging, characterized in that, Includes the following steps: Undersampling is performed on noisy hyperspectral data, specifically as follows: the wavenumber range for imaging in the hyperspectral data is calibrated according to requirements; a fixed number of points are undersampled within the waveband range; the spectral intensities corresponding to the undersampled points are summed to generate the value of a single pixel in the hyperspectral imaging; the operation is repeated to obtain noisy undersampled hyperspectral images. Model training involves placing undersampled hyperspectral images at the input and output of the model, respectively. The model consists of two main parts: an encoder and a decoder. Both the input of the encoder and the output of the decoder contain noisy images. By using the mean square error as the loss function, the model is trained until convergence, resulting in a well-trained model. The model-based denoising method involves inputting multiple noisy undersampled hyperspectral images into a trained model. The model denoises the input undersampled hyperspectral images and outputs the denoised undersampled images. The denoised undersampled image results are then summed to form the denoised hyperspectral image.

2. The deep learning denoising method for a single hyperspectral image according to claim 1, characterized in that, The fixed number of points is one-fifth of the total number of points.

3. The deep learning denoising method for a single hyperspectral image according to claim 1, characterized in that, The model in question is the U-net model.

Images