Three-dimensional fluorescence spectrum image super-resolution reconstruction method for water quality detection instrument

Abstract

The invention relates to a three-dimensional fluorescence spectrum image super-resolution reconstruction method for a water quality detection instrument, which is suitable for reconstructing a low-resolution and sparse three-dimensional fluorescence spectrum image collected by a portable water quality detection instrument into a high-resolution full spectrum. According to the method, an efficient super-resolution reconstruction network architecture composed of a feature extraction module, an up-sampling module and a trainable multi-Gaussian mixture point spread function is constructed. A joint loss function combining a reconstruction error and a degradation consistency error is designed, and a joint training strategy is adopted to synchronously optimize a network and PSF parameters, so that accurate reduction of a fluorescence peak shape structure and enhancement of spectrum authenticity are realized. The applicability and robustness of the portable water quality detection instrument based on the three-dimensional fluorescence spectrum in pollutant identification and analysis tasks are improved, and the portable water quality detection instrument is suitable for actual scenes such as intelligent water quality monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of water quality detection technology and image processing, and specifically relates to a three-dimensional fluorescence spectral image super-resolution reconstruction method for water quality detection instruments, which reconstructs sparse spectral data collected by a portable device into a high-resolution complete three-dimensional fluorescence spectral image. Background Art

[0002] To achieve precise management and effective treatment of water quality, it is urgent to establish an efficient and reliable means of real-time monitoring of water pollutants. Currently, a variety of mature technologies and solutions have been developed in the field of water quality monitoring. Common methods include electrochemical detection, biosensors, ultrasonic detection, and spectroscopy. Among them, three-dimensional fluorescence spectroscopy technology is particularly suitable for water quality monitoring because it can obtain a complete excitation-emission matrix and realize the "chemical fingerprint" identification of pollutants in complex water samples. It has the advantages of high sensitivity, rapid detection, and non-destructive analysis.

[0003] However, in practical applications, especially in portable water quality detectors used in monitoring terminals, due to the limitations of their hardware design, such as the excitation light source is often LED, and the emission light is selected using bandpass filters or fixed wavelength photodetectors, which are limited in number and have problems such as wavelength bandwidth broadening, the collected three-dimensional fluorescence data often presents low resolution and sparse characteristics. This type of low-resolution data often manifests as blurred fluorescence peak morphology and loss of spectral details, making it difficult to fully reflect the fine structural characteristics of pollutants in water bodies. As a result, the data collected is significantly different from the high-resolution full-spectrum data obtained by laboratory-level high-end three-dimensional fluorescence spectrometers. The two are difficult to compare in terms of resolution and data integrity. This difference directly affects the effectiveness of subsequent data processing. In addition, the existing pollutant identification model relies on high-resolution full-spectrum data training, which is difficult to directly migrate and apply to the sparse spectral images collected by portable devices.

[0004] Therefore, an efficient super-resolution image reconstruction method is needed that can reconstruct the low-resolution three-dimensional fluorescence images obtained by portable terminals into high-resolution full-spectral data consistent with laboratory equipment, and accurately restore the morphological details of the fluorescence peaks. This will not only significantly improve the adaptability of portable devices in pollutant detection, identification and classification, but also break the dependence of high-precision detection technology on expensive hardware, and provide low-cost, high-efficiency technical support for intelligent water quality monitoring and management. In traditional super-resolution reconstruction methods, interpolation-based methods obtain unrealistic images, while learning-based methods such as CNN and GNN, while relying on data volume and network depth, have risks such as overfitting and poor robustness. Summary of the Invention

[0005] In order to overcome the above problems, the present invention provides a three-dimensional fluorescence spectral image super-resolution reconstruction method for water quality testing instruments with high authenticity and good robustness, which reconstructs the low-resolution and sparse three-dimensional fluorescence spectrum collected by a portable three-dimensional fluorescence spectrometer into a high-resolution full-spectrum image.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A super-resolution reconstruction method for three-dimensional fluorescence spectral images of water quality testing instruments is developed. An efficient feature extraction and upsampling network is designed, and a trainable multi-Gaussian mixture point spread function is introduced for degradation simulation to accurately capture changes in fluorescence peak shape and improve the authenticity of the reconstruction results. A joint training strategy is used to simultaneously optimize the network parameters and the point spread function parameters to improve the accuracy of the reconstructed image and the robustness of the algorithm, thereby enhancing the practicality and reliability of spectral data collected by portable instruments in pollutant identification and analysis.

[0008] The method comprises the following steps:

[0009] Step 1, Dataset Construction: For each water sample, use a water quality testing instrument and a large-scale three-dimensional fluorescence spectrum scanner F4600 to obtain the corresponding sparse spectral data and high-resolution full spectral data, and perform preprocessing to obtain the dataset:

[0010] Step 2, preliminary prediction of high-resolution images;

[0011] Step 3, low-resolution image degradation simulation;

[0012] Step 4: Joint training loss function construction;

[0013] Step 5: Joint training of deep learning network and PSF parameters. The process is as follows:

[0014] During the training process, the deep learning network parameters and PSF parameters are simultaneously used as trainable variables. The Adam optimizer is used to back-propagate and update them using the objective function in step 4, so that the PSF parameters can adaptively learn and compensate for the complex changes in the spectral degradation process.

[0015] In step 6, the model parameters trained in step 5 are saved, and the low-resolution spectral data of the test sample is input into the trained super-resolution model for reconstruction to obtain the corresponding high-resolution full-spectrum image.

[0016] Preferably, in step 2, the process of preliminary prediction of the high-resolution image is:

[0017] First, the preprocessed data is input into the feature extraction module of the convolutional neural network according to the batch size, and the data features in the image are extracted through the convolution operation; then, the nonlinear activation function ReLU is connected to enhance the network's ability to fit nonlinear relationships, thereby capturing the detailed features of the fuzzy peak edges in the fluorescence image; the output is a set of low-level feature maps to prepare for subsequent upsampling; let the input image be X LR , then the low-level feature map is Y1, and its specific calculation method is as follows:

[0018] Y1=ReLU(W1*X LR +b1) (1)

[0019] Among them, W1 represents the convolution kernel operation, b1 is the bias term;

[0020] Then, upsampling based on flax pixel convolution is performed to achieve image size expansion to obtain a high-resolution image. The convolution layer is used to expand the dimension of Y1 to the square of the upsampling ratio r to obtain the dimension-expanded image Y2. The PixelShuffle algorithm (flax pixel rearrangement) is then used to further reconstruct it into spatial dimensions (H×r)×(W×r), realizing pixel interpolation and rearrangement in the spatial domain to obtain a high-resolution image Y pred , which effectively avoids the common checkerboard effect of deconvolution operation and has a stronger ability to preserve image structure. The formula is as follows:

[0021]

[0022] Y pred =PixelShuffle(Y2),Y pred ∈R B×(H×r)×(W×r) (3)

[0023] Among them, W2 represents the convolution kernel operation, b2 is the bias term, R is the dataset, B is the batch size, and H and W are the height and width of the original image.

[0024] More preferably, in step 3, the process of simulating low-resolution image degradation is:

[0025] The point spread function (PSF) is used to simulate the degradation process from a high-resolution spectral image to a sparse spectral image, thereby capturing subtle changes in spectral degradation and enhancing the authenticity of the reconstructed spectrum. To match the characteristics of the fluorescence peak shape, the point spread function h is defined as a weighted mixture of N two-dimensional Gaussian functions to constrain the spectral peak shape, which is defined as follows:

[0026]

[0027] Among them, x and y represent two variables in the two-dimensional space of the point spread function, Ai , σ i , Δx i and Δy i is the amplitude, standard deviation and spatial displacement of the i-th Gaussian component;

[0028] Then the high-resolution image Y is predicted in step 2 pred Convolution with PSF and then downsampling are performed to achieve the degradation process from high to low resolution. The calculation formula is as follows:

[0029] X pred =D(Y pred *h)+η (5)

[0030] Among them, * is the convolution operation, D is the downsampling operator, and η is the noise term. The module finally outputs a degraded simulated low-resolution image X pred , used for comparison with real images.

[0031] More preferably, in step 4, the process of constructing the joint training loss function is:

[0032] In order to enable the network to take into account both the feature expression and spectral degradation pattern at the data level, enhance robustness and reduce the risk of overfitting, the present invention adopts an end-to-end joint training strategy, whose objective function consists of two parts: reconstruction loss and degradation consistency loss;

[0033] The reconstruction loss is used to compare the predicted high-resolution spectrum Y in step 2 pred With the real high-resolution spectrum Y HR The specific expression is as follows.

[0034]

[0035] in, and Represents the predicted high-resolution image and the true high-resolution image of the i-th sample;

[0036] The degradation consistency loss is used to compare the degradation simulated low-resolution image X obtained in step 3 pred , and the real low-resolution image X LR The specific expression is

[0037]

[0038] in, and represents the degraded simulated low-resolution image and the real low-resolution image of the jth sample;

[0039] Therefore, the overall loss function is defined as follows:

[0040] L total =L recon +λ deg L deg (8)

[0041] Among them, λ deg is a weight hyperparameter used to balance the contribution of the two losses during training.

[0042] The technical concept of the present invention is to propose a super-resolution reconstruction method for three-dimensional fluorescence spectra that combines a deep learning network with degradation simulation. This method constructs a neural network structure composed of a convolutional feature extraction module and an upsampling module based on flax pixel convolution, which can effectively extract spatial and spectral features from low-resolution images and achieve image amplification. At the same time, a trainable multi-Gaussian mixture point spread function is designed to simulate the actual spectral degradation process, and the degradation law of the spectrum from high to low resolution is embedded in the network learning framework. By constructing a joint training strategy including reconstruction loss and degradation consistency loss, the simultaneous optimization of network parameters and PSF parameters is achieved. It is suitable for super-resolution reconstruction tasks of sparse spectral data.

[0043] The beneficial effects of the present invention are as follows:

[0044] 1. The designed network structure combines a convolution-based feature extraction module with a sub-pixel convolution-based upsampling module, which can effectively restore the overall structure and local details of the fluorescence peak in the fluorescence spectrum image, thereby improving the image reconstruction quality.

[0045] 2. A multi-Gaussian mixture point spread function is introduced to simulate the spectral degradation process of spectral data from high resolution to low resolution, which enhances the ability to capture the details of spectral changes and improves the authenticity of the reconstructed spectrum.

[0046] 3. A joint training strategy is introduced to make the reconstruction process not only data-driven but also integrate the spectral degradation law. This improves the ability of the reconstruction results to fit the real spectral features while reducing the dependence on large-scale training data, effectively reducing the risk of overfitting and improving robustness and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a process framework diagram of the present invention;

[0048] Figure 2 are real low-resolution grayscale pixel images, where (A) is a real low-resolution grayscale pixel image of aniline; (B) is a real low-resolution grayscale pixel image of salicylic acid; (C) is a real low-resolution grayscale pixel image of fluorescent whitening agent; (D) is a real low-resolution grayscale pixel image of ciprofloxacin; (E) is a real low-resolution grayscale pixel image of norfloxacin;

[0049] Figure 3 are real high-resolution grayscale pixel images; (A) is a real high-resolution grayscale pixel image of aniline; (B) is a real high-resolution grayscale pixel image of salicylic acid; (C) is a real high-resolution grayscale pixel image of fluorescent whitening agent; (D) is a real high-resolution grayscale pixel image of ciprofloxacin; (E) is a real high-resolution grayscale pixel image of norfloxacin;

[0050] Figure 4 are reconstructed high-resolution grayscale pixel images, where (A) is the reconstructed high-resolution grayscale pixel image of aniline; (B) is the reconstructed high-resolution grayscale pixel image of salicylic acid; (C) is the reconstructed high-resolution grayscale pixel image of fluorescent whitening agent; (D) is the reconstructed high-resolution grayscale pixel image of ciprofloxacin; and (E) is the reconstructed high-resolution grayscale pixel image of norfloxacin. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0052] Reference Figures 1 to 4 A three-dimensional fluorescence spectrum image super-resolution reconstruction method for water quality detection instruments is used to reconstruct the sparse three-dimensional fluorescence spectrum collected by the portable terminal three-dimensional fluorescence spectrum instrument into a high-resolution full spectrum. The method first needs to construct the corresponding high and low resolution data sets under the same sample and send them to Figure 1 The prediction is performed in the network framework shown in the figure, the loss values ​​of the predicted high and low resolution images and the true high and low resolution spectra are calculated, and then the various parameters are updated through back propagation.

[0053] In this embodiment, aqueous solution samples of five substances, namely, aniline, salicylic acid, fluorescent whitening agent, ciprofloxacin, and norfloxacin, with various concentrations are used for training and testing, including the following steps:

[0054] Step 1, dataset construction;

[0055] In this method, a portable instrument and a large-scale three-dimensional fluorescence spectrum scanner F4600 are used for each water sample to obtain the corresponding sparse spectrum data and high-resolution full spectrum data, and then preprocess them to obtain a data set. Figure 2 and Figure 3 As shown;

[0056] Step 2, preliminary prediction of high-resolution images;

[0057] First, the preprocessed data is fed into the feature extraction module of the convolutional neural network according to the batch size, and the data features in the image are extracted through the convolution operation. Subsequently, the nonlinear activation function ReLU is connected to enhance the network's ability to fit nonlinear relationships, thereby capturing the detailed features of the blurred peak edges in the fluorescence image. The output is a set of low-level feature maps, which prepare for subsequent upsampling. Let the input image be X HR , then the low-level feature map is Y1, and its specific calculation method is as follows:

[0058] Y1=ReLU(W1*X HR +b1) (1)

[0059] Among them, W1 represents the convolution kernel operation, b1 is the bias term;

[0060] Then, upsampling based on flax pixel convolution is performed to expand the size of the image to obtain a high-resolution image. The convolution layer is used to expand the dimension of Y1 to the square of the upsampling factor r, and the dimension-expanded image Y2 is obtained. The PixelShuffle algorithm (flax pixel rearrangement) is then used to further reconstruct it into spatial dimensions (H×r)×(W×r), realizing pixel interpolation and rearrangement in the spatial domain, and obtaining a high-resolution image Y pred , which effectively avoids the common checkerboard effect of deconvolution operation and has a stronger ability to preserve image structure. The formula is as follows:

[0061]

[0062] Y pred =PixelShuffle(Y2),Y pred ∈R B×(H×r)×(W×r) (3)

[0063] Where W2 represents the convolution kernel operation, b2 is the bias term, R is the dataset, B is the batch size, H and W are the height and width of the original image;

[0064] The detailed parameters of the two convolutional layers are as follows:

[0065] Table 1 shows the detailed parameters of the convolutional layer.

[0066]

[0067] Step 3, low-resolution image degradation simulation;

[0068] The point spread function (PSF) is used to simulate the degradation process from a high-resolution spectral image to a sparse spectral image, thereby capturing subtle changes in spectral degradation and enhancing the authenticity of the reconstructed spectrum. In order to match the characteristics of the fluorescence peak shape, the point spread function h is defined as a weighted mixture of N two-dimensional Gaussian functions to constrain the spectral peak shape. The specific definition is as follows:

[0069]

[0070] Among them, x and y represent two variables in the two-dimensional space of the point spread function, A i , σ i , Δx i and Δy i are the amplitude, standard deviation, and spatial displacement of the i-th Gaussian component.

[0071] Then the high-resolution image Y is predicted in step 2 pred Convolution with PSF and then downsampling are performed to achieve the degradation process from high to low resolution. The specific calculation steps are as follows.

[0072] X pred =D(Y pred *h)+η (5)

[0073] Among them, * is the convolution operation, D is the downsampling operator, η is the noise term, and the module finally outputs a simulated low-resolution image X pred , used for comparison with real images.

[0074] Step 4: Joint training loss function construction;

[0075] In order to enable the network to simultaneously take into account the feature expression and spectral degradation pattern at the data level, enhance robustness and reduce the risk of overfitting, the present invention adopts an end-to-end joint training strategy, whose objective function consists of two parts: reconstruction loss and degradation consistency loss.

[0076] The reconstruction loss is used to compare the predicted high-resolution spectrum Y in step 2 pred With the real high-resolution spectrum Y HR The specific expression is as follows.

[0077]

[0078] in, and Represents the predicted high-resolution image and the true high-resolution image of the i-th sample.

[0079] The degradation consistency loss is used to compare the degradation simulated low-resolution image X obtained in step 3 pred , and the real low-resolution image XLR The specific expression is

[0080]

[0081] in, and Represents the j-th corresponding predicted low-resolution image and the true low-resolution image.

[0082] Therefore, the overall loss function is defined as follows:

[0083] L total =L recon +λ deg L deg (8)

[0084] Among them, λ deg is a weight hyperparameter used to balance the contribution of the two losses during training

[0085] Step 5: Joint training of deep learning network and PSF parameters;

[0086] During training, the parameters of the deep learning network and the PSF parameters are treated as jointly trainable variables and optimized simultaneously using the Adam optimizer using the objective function from step 4. The learning rate is set to 0.001, and the number of training rounds is 1000. Through iterative training, the super-resolution model's parameters are fitted and converged, ultimately obtaining a complete model for image reconstruction.

[0087] Step 6: Input the low-resolution spectral image data of the test samples of the five substances (aniline, salicylic acid, fluorescent whitening agent, ciprofloxacin and norfloxacin) into the trained super-resolution model for reconstruction to obtain the corresponding high-resolution full-spectral images. Figure 4 As shown in Figure 2, the reconstruction effect of the trained model on the above-mentioned pollutant samples is demonstrated.

[0088] This method can reconstruct the low-resolution, sparse 3D fluorescence spectral data collected by the terminal device into full spectral data consistent with the 3D fluorescence spectral scanner used in the laboratory, which is conducive to further qualitative and quantitative analysis of contaminated water samples.

[0089] Reference Figure 2 The following are pixel grayscale images of the real low-resolution three-dimensional fluorescence of five substances, including aniline, salicylic acid, fluorescent whitening agent, ciprofloxacin and norfloxacin, collected by a portable water quality detection instrument. Figure 3 The following are the real high-resolution three-dimensional fluorescence pixel grayscale images of five substances, including aniline, salicylic acid, fluorescent whitening agent, ciprofloxacin and norfloxacin, collected by the three-dimensional fluorescence spectrum scanner F4600, as shown in the reference Figure 4The figure shows the reconstruction of the sparse spectra of five substances, including aniline, salicylic acid, fluorescent whitening agent, ciprofloxacin and norfloxacin, into high-resolution full-spectrum pixel grayscale images using this method. Figure 3 and Figure 4 Clearly, the full spectral data from the 3D fluorescence spectrum reconstructed using this super-resolution method is highly similar to the true spectral data. This approach could play a greater role in the subsequent qualitative identification and quantitative analysis of pollutants. For example, after reconstructing the sparse spectrum into full spectral data, existing analytical tools or general models can be used directly to identify or analyze pollutants.

[0090] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. A three-dimensional fluorescence spectral image super-resolution reconstruction method for water quality detection instruments, characterized in that: An efficient feature extraction and upsampling network was designed, and a trainable multi-Gaussian mixture point spread function was introduced for degradation simulation to accurately capture the changes in fluorescence peak shape and improve the authenticity of the reconstruction results. Through a joint training strategy, the network parameters and the parameters of the point spread function were simultaneously optimized to improve the accuracy of the reconstructed image and the robustness of the algorithm, thereby enhancing the practicality and reliability of spectral data collected by portable instruments in pollutant identification and analysis.

2. The three-dimensional fluorescence spectrum image super-resolution reconstruction method for water quality detection instrument according to claim 1, characterized in that: The method comprises the following steps: Step 1, Dataset Construction: For each water sample, use a water quality testing instrument and a large-scale three-dimensional fluorescence spectrum scanner to obtain the corresponding sparse spectral data and high-resolution full spectral data, and perform preprocessing to obtain the dataset: Step 2, preliminary prediction of high-resolution images; Step 3, low-resolution image degradation simulation; Step 4: Joint training loss function construction; Step 5: Joint training of deep learning network and PSF parameters. The process is as follows: During the training process, the deep learning network parameters and PSF parameters are simultaneously used as trainable variables. The Adam optimizer is used to back-propagate and update them using the objective function in step 4, so that the PSF parameters can adaptively learn and compensate for the complex changes in the spectral degradation process. In step 6, the model parameters trained in step 5 are saved, and the low-resolution spectral data of the test sample is input into the trained super-resolution model for reconstruction to obtain the corresponding high-resolution full-spectrum image.

3. The method for super-resolution reconstruction of three-dimensional fluorescence spectrum images for water quality testing instruments according to claim 2, characterized in that: In step 2, the process of preliminary prediction of high-resolution images is as follows: First, the preprocessed data is fed into the feature extraction module of a convolutional neural network according to the batch size, where the data features in the image are extracted through convolution operations. Subsequently, a nonlinear activation function (ReLU) is introduced to enhance the network's ability to fit nonlinear relationships, thereby capturing the detailed features of the blurred peak edges in the fluorescence image. The output is a set of low-level feature maps, which are prepared for subsequent upsampling. Let the input image be X LR , then the low-level feature map is Y1, and its specific calculation method is as follows: Y1=ReLU(W1*X LR +b1) (1) Among them, W1 represents the convolution kernel operation, b1 is the bias term; Then, upsampling based on flax pixel convolution is performed to expand the size of the image to obtain a high-resolution image. The convolution layer is used to expand the dimension of Y1 to the square of the upsampling ratio r to obtain the dimension-expanded image Y2. The PixelShuffle algorithm is then used to further reconstruct the spatial dimension (H×r)×(W×r) to achieve pixel interpolation and rearrangement in the spatial domain to obtain a high-resolution image Y. pred , which effectively avoids the common checkerboard effect of deconvolution operation and has a stronger ability to preserve image structure. The formula is as follows: AND pred =PixelShuffle(Y2),Y pred ∈R B×(H×r)×(W×r) (3) Among them, W2 represents the convolution kernel operation, b2 is the bias term, R is the dataset, B is the batch size, and H and W are the height and width of the original image.

4. The three-dimensional fluorescence spectrum image super-resolution reconstruction method for water quality detection instrument according to claim 3, characterized in that: The process of step 3, low-resolution image degradation simulation is as follows: The point spread function (PSF) is used to simulate the degradation process from a high-resolution spectral image to a sparse spectral image, thereby capturing subtle changes in spectral degradation and enhancing the authenticity of the reconstructed spectrum. In order to match the characteristics of the fluorescence peak shape, the point spread function h is defined as a weighted mixture of N two-dimensional Gaussian functions to constrain the spectral peak shape, which is defined as follows: Among them, x and y represent two variables in the two-dimensional space of the point spread function, A i , σ i , Δx i and Δy i is the amplitude, standard deviation and spatial displacement of the i-th Gaussian component; Then the high-resolution image Y is predicted in step 2 pred Convolution with PSF and then downsampling are performed to achieve the degradation process from high to low resolution. The calculation formula is as follows: X pred =D(Y pred *h)+η (5) Among them, * is the convolution operation, D is the downsampling operator, η is the noise term, and the final output is a degraded simulated low-resolution image X pred , used for comparison with real images.

5. The three-dimensional fluorescence spectrum image super-resolution reconstruction method for water quality detection instrument according to claim 4, characterized in that: In step 4, the process of constructing the joint training loss function is as follows: An end-to-end joint training strategy is adopted, and its objective function consists of two parts: reconstruction loss and degradation consistency loss; The reconstruction loss is used to compare the predicted high-resolution spectrum Y in step 2 pred With the real high-resolution spectrum Y HR The specific expression is as follows: in, and Represents the predicted high-resolution image and the true high-resolution image of the i-th sample; The degradation consistency loss is used to compare the degradation simulated low-resolution image X obtained in step 3 pred , and the real low-resolution image X LR The specific expression is in, and represents the degraded simulated low-resolution image and the real low-resolution image of the jth sample; Therefore, the overall loss function is defined as follows: L total =L recon +λ deg L deg (8) Among them, λ deg is a weight hyperparameter used to balance the contribution of the two losses during training.

Images