A chemical plant area intelligent safety detection method and system based on deep learning

Abstract

The present invention discloses a chemical plant intelligent safety detection method based on deep learning, wherein multiple cameras are installed in the area where there are safety hazards in the chemical plant to obtain images of workers on site; the original image is processed by improved adaptive wavelet threshold denoising; then image data is generated by CycleGAN to solve the problem of sample imbalance; the pre-processed image is labeled to construct a data set; the processed data set is input into the improved EfficientDet target detection model; the improved EfficientDet model includes: introducing the PSA attention mechanism in MBConv, improving the bidirectional feature pyramid network BiFPN structure; inputting the image to be detected into the trained improved EfficientDet model, obtaining the wearing condition and position information of the workers' protective equipment, and further judging whether the relevant personnel have dangerous behaviors; finally, the detection results are displayed on the front-end interface and the corresponding early warning signal is issued. The present invention provides a system capable of real-time, accurate and automatic detection of safety hazards in chemical plant areas.

Description

Technical Field

[0001] The present invention relates to an image processing and target detection technology based on deep learning, and in particular to an intelligent safety detection method and system for a chemical plant area based on deep learning. Background Art

[0002] In a high-risk environment like a chemical plant, work safety is crucial to protecting the lives of personnel and ensuring stable production. Chemical plants often face the storage and use of flammable, explosive and toxic chemicals, with high accident risks. Once a safety accident occurs, the consequences are serious, which may not only cause huge economic losses, but also cause irreversible impacts on the environment and personnel safety.

[0003] With the development of science and technology, especially the advancement of artificial intelligence and computer vision technology, it has become feasible to use intelligent monitoring to achieve automatic monitoring and control of safe production. However, most of the existing monitoring systems rely on traditional models, such as scheduled inspections and manual monitoring, which not only consumes a lot of manpower and material resources, but also makes it difficult to detect all potential safety hazards in real time.

[0004] Among the current technical solutions, although some safety detection systems based on deep learning have been proposed and applied in the field of public safety, in the complex and changeable chemical production environment, these systems often fail to meet actual needs due to problems such as poor image quality, slow model detection speed or low accuracy. In addition, these systems often fail to effectively identify and provide real-time feedback on dangerous behaviors and states specific to chemical plants, such as improper use of personal protective equipment. Summary of the invention

[0005] Purpose of the invention: To provide a system that can detect safety hazards in chemical plant areas in real time, accurately and automatically, especially a system that can effectively identify the wearing of workers' protective equipment and their behavioral status.

[0006] Technical solution: The present invention discloses a method and system for intelligent safety detection of chemical plant areas based on deep learning. The system mainly includes image acquisition, image preprocessing, target detection model and front-end interface. The method includes the following steps:

[0007] S1. Install multiple cameras in areas of the chemical plant where there are potential safety hazards to obtain images of workers on site;

[0008] S2, using improved wavelet threshold denoising to process the original image;

[0009] S3. Generate image data through CycleGAN to solve the problem of sample imbalance;

[0010] S4. Label the head, torso, and limbs in the preprocessed images and construct a dataset.

[0011] S5. Build an improved EfficientDet chemical plant safety detection model. The improvements include introducing the PSA attention mechanism in MBConv and improving the Bidirectional Feature Pyramid Network (BiFPN) structure.

[0012] S6. Input the data set into the improved EfficientDet model for training and test the model performance;

[0013] S7. Use the trained model to perform target detection on the chemical plant images collected on site to obtain the workers’ protective equipment wearing and personal behavior;

[0014] S8. Further judge whether dangerous behavior occurs based on the result of target detection, and display the safety status of the chemical plant in real time through the front-end interface. When there is an abnormal situation, the early warning device is automatically triggered and an alarm signal is issued.

[0015] Furthermore, the implementation process of step S2 is as follows:

[0016] S21. Let s(x,y) be the original image signal, n(x,y) be the Gaussian white noise signal, and f(x,y) be the image signal interfered by the noise. We can get:

[0017] f(x,y)=s(x,y)+n(x,y)

[0018] Where x, y = 1, 2, 3…N + .

[0019] S22, selecting a suitable wavelet basis to perform multi-scale wavelet decomposition on the image, and obtaining a series of wavelet coefficients of high-frequency and low-frequency components at different scales;

[0020] S23. Select appropriate threshold λ for different decomposition scales:

[0021]

[0022] Among them, σ is the standard deviation of the noise; N is the image length; o is the number of decomposition layers.

[0023] S24, threshold function for high frequency coefficient w x,y Perform threshold quantization to obtain high-frequency estimation coefficients

[0024] S25. The discontinuity of the hard threshold function will cause oscillation during reconstruction, and the final image will be blurred; the soft threshold function may cause the image edge information to be lost. For this reason, an adaptive threshold function is proposed:

[0025]

[0026] in, is the high-frequency estimated coefficient; w x,y is the high frequency coefficient; λ is the threshold; sgn() is the sign function; R is the adjustment factor.

[0027] S26, the adjustment factor R in the new threshold function affects the denoising effect, so the optimal adjustment factor R is found through the predator optimization algorithm;

[0028] S27, reconstructing the low-frequency coefficients after wavelet decomposition and the high-frequency coefficients after threshold quantization of each layer to obtain a denoised image.

[0029] The predator optimization algorithm described in step S26 searches for the optimal adjustment factor R, and the specific process is as follows:

[0030] S261, randomly initialize the population position, that is, the adjustment factor R in the threshold function;

[0031] S262. Define the fitness function as F:

[0032] F=SNR+SSIM

[0033] In the formula, SNR is the signal-to-noise ratio and SSIM is the structural similarity.

[0034] S263, calculating the fitness value and recording the global optimum;

[0035] S264, updating the location of prey populations;

[0036] S265, determine whether the maximum number of iterations has been reached, if not, repeat steps S262 and S263 until the termination condition is reached;

[0037] S266. Output the particle corresponding to the global optimal solution, and substitute the optimal population position as the adjustment factor R into the new threshold function.

[0038] Furthermore, in step S3, CycleGAN generates an image, and the implementation process is as follows:

[0039] S31, CycleGAN consists of two generative adversarial networks (GANs), which learn the mapping from source image to target image and the mapping from target image to source image respectively. The collected images are divided into two domains, A and B. The original images in domains A and B are converted by generators G and H respectively, and the discriminator is used to judge the authenticity of the generated images. The loss function of the two generators G and H can be defined as:

[0040]

[0041] Where E represents the expected value, and a and b are images from domains A and B respectively. data(a) and Pdata(b) They represent the probability distribution of image domain A and image domain B respectively.

[0042] S32. A single adversarial loss function cannot guarantee that a single input can be mapped. It may happen that the generator converts an image in domain A into an image in domain B, making the adversarial loss function invalid. To avoid this possibility, a cycle consistency loss function Loss is introduced. cycle :

[0043]

[0044] S33. Through equations (4), (5), and (6), we can get the total loss function of CycleGAN: full :

[0045] Loss full =Loss GAN (G,D A )+Loss GAN (F,D B )+λLoss cycle

[0046] Among them, the parameter λ is the weight of the cycle consistency loss.

[0047] S34, cyclic training generator and discriminator, optimize the quality and cycle consistency of generated images. CycleGAN is more stable in the actual training process, and the quality of generated images is higher.

[0048] Furthermore, in step S5, an improved EfficientDet chemical plant safety detection model is constructed, and the model construction process is as follows:

[0049] S511, EfficientDet is composed of three parts: backbone network EfficientNet, feature extraction module BiFPN and prediction network. EfficientNet uses MBConv as the backbone structure of the model. MBConv consists of dilated convolution, depthwise separable convolution, SE attention mechanism module and linear residual connection. SE only considers channel attention and ignores spatial attention. In order to fully extract the multi-scale spatial information and important cross-dimensional features in the channel attention vector, the SE attention mechanism in MBConv is replaced with the improved PSA attention mechanism, which can improve the ability to detect small-scale target objects and improve the overall accuracy of the model.

[0050] S521, dividing the channel by squeezing and splicing SPC modules, and performing multi-scale feature extraction on the spatial information on each channel feature map;

[0051] S522, using the SEWeight module to extract the channel attention of feature maps of different scales, and obtain the channel attention vector at each different scale;

[0052] S523. The original Softmax function is replaced by the Sparsemax function. Sparsemax generates a sparser attention distribution. The multi-scale channel attention vector is recalibrated using Sparsemax to obtain the new attention weight after multi-scale channel interaction.

[0053] S524, performing a dot multiplication operation on the recalibrated weights and the corresponding feature map element by element, and outputting a feature map weighted by multi-scale feature information attention.

[0054] S531, in order to enhance the utilization of intermediate information at different levels and enhance network performance, the BiFPN network structure is improved. The improved BiFPN integrates intermediate feature information and improves the detection capability of the target. In the top-down path, intermediate information is obtained. The process is as follows:

[0055]

[0056] Among them, P j is the input feature of the jth layer, is the intermediate feature information of the jth layer, ω is the weight, Res is the upsampling or downsampling operation matching the current resolution, Conv is the convolution operation used for feature processing, and e=0.0001 is a small value to reduce the instability of the calculation.

[0057] S532. In the bottom-up path, make full use of intermediate information of different scales to obtain the output of the network:

[0058]

[0059] in, is the output feature of the jth layer.

[0060] Furthermore, the acquisition of workers' protective equipment wearing and personal behavior mentioned in step S7 mainly includes the following situations: not wearing personal protective equipment such as safety helmets, safety glasses, protective gloves, safety shoes, mistakenly entering dangerous areas to perform high-risk operations, and smoking in non-smoking areas.

[0061] Furthermore, the front-end interface mentioned in step S8 displays the safety status of the chemical plant in real time, and the test results will be displayed in real time on the front-end interface, including the wearing status of the workers' protective equipment and their specific location information. The system further determines the safety status of the relevant personnel by analyzing this information, and immediately issues a warning signal to notify the relevant person in charge and safety management personnel when dangerous behavior is found.

[0062] Beneficial effects:

[0063] 1) The introduction of wavelet threshold denoising and Cycle Generative Adversarial Network (CycleGAN) technology significantly improves the quality of the dataset. This improvement in quality helps deep learning models detect and analyze targets more accurately, especially in poor lighting or obstructed vision conditions.

[0064] 2) By introducing the PSA attention mechanism in MBConv and improving the bidirectional feature pyramid network (BiFPN), the performance of EfficientDet in multi-scale image processing is enhanced. This provides the system with better generalization capabilities and can adapt to a variety of different monitoring environments and conditions.

[0065] 3) The system can not only detect the current use of safety equipment, but also prevent potential safety accidents. Once irregular behavior or improper wearing of equipment is detected, the system will automatically trigger an alarm and quickly notify relevant personnel through the front-end interface so that necessary safety measures can be taken. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is the flow chart of the intelligent safety detection system for chemical plants;

[0067] Figure 2 This is a schematic diagram of the improved MBconv structure;

[0068] Figure 3 Schematic diagram of the improved BiFPN structure. DETAILED DESCRIPTION

[0069] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0070] The present invention provides a chemical plant area intelligent safety detection method and system based on deep learning, such as Figure 1 As shown, the following steps are included:

[0071] S1. Install multiple cameras in areas of the chemical plant where there are potential safety hazards to obtain images of workers on site;

[0072] S2, using improved wavelet threshold denoising to process the original image;

[0073] S21. Let s(x,y) be the original image signal, n(x,y) be the Gaussian white noise signal, and f(x,y) be the image signal interfered by the noise. We can get:

[0074] f(x,y)=s(x,y)+n(x,y)

[0075] Where x, y = 1, 2, 3…N + .

[0076] S22, selecting a suitable wavelet basis to perform multi-scale wavelet decomposition on the image, and obtaining a series of wavelet coefficients of high-frequency and low-frequency components at different scales;

[0077] S23. Select appropriate threshold λ for different decomposition scales:

[0078]

[0079] Among them, σ is the standard deviation of the noise; N is the image length; o is the number of decomposition layers.

[0080] S24, threshold function for high frequency coefficient w x,y Perform threshold quantization to obtain high-frequency estimation coefficients

[0081] S25. The discontinuity of the hard threshold function will cause oscillation during reconstruction, and the final image will be blurred; the soft threshold function may cause the image edge information to be lost. For this reason, an adaptive threshold function is proposed:

[0082]

[0083] in, is the high-frequency estimated coefficient; w x,y is the high frequency coefficient; λ is the threshold; sgn() is the sign function; R is the adjustment factor.

[0084] S26, reconstructing the low-frequency coefficients after wavelet decomposition and the high-frequency coefficients after threshold quantization of each layer to obtain a denoised image.

[0085] S3. Generate image data through CycleGAN to solve the problem of sample imbalance;

[0086] S31, CycleGAN consists of two generative adversarial networks (GANs), which learn the mapping from source image to target image and the mapping from target image to source image respectively. The collected images are divided into two domains, A and B. The original images in domains A and B are converted by generators G and H respectively, and the discriminator is used to judge the authenticity of the generated images. The loss function of the two generators G and H can be defined as:

[0087]

[0088] Where E represents the expected value, and a and b are images from domains A and B respectively. data(a) and P data(b) They represent the probability distribution of image domain A and image domain B respectively.

[0089] S32. A single adversarial loss function cannot guarantee that a single input can be mapped. It may happen that the generator converts an image in domain A into an image in domain B, making the adversarial loss function invalid. To avoid this possibility, a cycle consistency loss function Loss is introduced. cycle :

[0090]

[0091] S33. Through equations (4), (5), and (6), we can get the total loss function of CycleGAN: full :

[0092] Loss full =Loss GAN (G,D A )+Loss GAN (F,D B )+λLoss cycle

[0093] Among them, the parameter λ is the weight of the cycle consistency loss.

[0094] S34, cyclic training generator and discriminator, optimize the quality and cycle consistency of generated images. CycleGAN is more stable in the actual training process, and the quality of generated images is higher.

[0095] S4. Label the head, torso, and limbs in the preprocessed images and construct a dataset.

[0096] S5. Build an improved EfficientDet chemical plant safety detection model. The improvements include introducing the PSA attention mechanism in MBConv and improving the Bidirectional Feature Pyramid Network (BiFPN) structure.

[0097] S511 uses EfficientNets as its backbone network, which consists of 7 MBConv stacks; introduces a weighted bidirectional feature pyramid network (BiFPN) as the feature network, extracts features from the 3rd to 7th layers, and repeatedly applies top-down and bottom-up bidirectional feature fusion; the output features are sent to the classification and box prediction network to output the object category and bounding box prediction respectively.

[0098] S521. First, the input tensor is processed using the Squeeze and Concatenate module (SPC) to obtain a multi-scale feature map at the channel level. This module extracts spatial information of different scales from the feature map of each channel more effectively by compressing the channel dimension. The multi-scale feature map generation function is given by the following formula:

[0099] F i =Conv(k i ×ki ,G i )(X)i=0,1,2…S-1

[0100] Among them, the i-th kernel size is k i =2×(i+1)+1, the size of the i-th group F i ∈R C'×H×W Represents feature maps with different scales.

[0101] S522, the entire multi-scale preprocessing feature map can be obtained by cascading:

[0102] F=Concat([F 0 ,F 1 ,…,F S-1 ])

[0103] Among them, F∈R C×H×W is the obtained multi-scale feature map.

[0104] S523. Secondly, the SEWeight module is used to extract the channel attention weights from the feature maps of different scales. This step generates a channel attention vector for each scale, which can gradually capture the contextual features of different spatial sizes. Mathematically, the vector of attention weights can be expressed as:

[0105] Z i =SEWeight(F i ),i=0,1,2…S-1

[0106] Among them, Z i ∈R C'×1×1 is the attention weight.

[0107] S524. In order to further realize the interaction of attention information, the cross-dimensional vectors are fused without destroying the original channel attention vector. The entire multi-scale channel attention vector is obtained in a cascade manner:

[0108]

[0109] Among them, Z i is the concat operator, Z i It is from F i The attention value of , Z is the multi-scale attention weight vector.

[0110] S525. Recalibrate the extracted channel attention vector using the Softmax operation to obtain the calibration weight of the multi-scale channel. The weight is given by the following formula:

[0111]

[0112] Among them, Softmax is used to obtain the recalibrated weights att of the multi-scale channel i .

[0113] S526: perform an element-wise multiplication operation on the obtained recalibrated weights and the corresponding feature maps. This step further refines the feature maps, so that the output feature maps are rich in more detailed multi-scale information.

[0114] Y i =F i ⊙att i i=1,2,3,…S-1

[0115] Among them, ⊙ represents channel-by-channel multiplication, Y i refers to the feature map with the obtained multi-scale channel-wise attention weights.

[0116] S527, the concatenation operator is more effective than the summation operator because it can completely maintain the feature representation without destroying the information of the original feature map. In summary, the process of obtaining the refined output can be written as:

[0117] Out = Concat([Y 0 ,Y 1 ,…,Y S-1 ])

[0118] S531, in order to enhance the utilization of intermediate information at different levels and enhance network performance, the BiFPN network structure is improved. The improved BiFPN integrates intermediate feature information and improves the detection capability of the target. In the top-down path, intermediate information is obtained. The process is as follows:

[0119]

[0120] Among them, P j is the input feature of the jth layer, is the intermediate feature information of the jth layer, ω is the weight, Res is the upsampling or downsampling operation matching the current resolution, Conv is the convolution operation used for feature processing, and e=0.0001 is a small value to reduce the instability of the calculation.

[0121] S532. In the bottom-up path, make full use of intermediate information of different scales to obtain the output of the network:

[0122]

[0123] in, is the output feature of the jth layer.

[0124] S6. Input the data set into the improved EfficientDet model for training and test the model performance;

[0125] S7. Use the trained model to perform target detection on the chemical plant images collected on site to obtain the workers’ protective equipment wearing and personal behavior;

[0126] Failure to wear personal protective equipment such as safety helmets, safety glasses, protective gloves, and safety shoes; mistakenly entering dangerous areas to perform high-risk operations; and smoking in non-smoking areas.

[0127] S8. Further judge whether dangerous behavior occurs based on the result of target detection, and display the safety status of the chemical plant in real time through the front-end interface. When there is an abnormal situation, the early warning device is automatically triggered and an alarm signal is issued.

[0128] The test results will be displayed in real time on the front-end interface, including the workers' protective equipment wearing status and specific location information. The system will further judge the safety status of relevant personnel by analyzing this information, and immediately issue a warning signal to notify relevant persons in charge and safety management personnel when dangerous behavior is found.

Claims

1. A chemical plant intelligent safety detection method based on deep learning, characterized in that: The following steps are involved: S1. Install multiple cameras in areas of the chemical plant where there are potential safety hazards to obtain images of workers on site; S2, using improved adaptive wavelet threshold denoising to process the original image; S3, generate image data through CycleGAN; S4. Label the head, torso, and limbs in the preprocessed images and construct a dataset. S5. Build an improved EfficientDet chemical plant safety detection model. The improvements include introducing an improved PSA attention mechanism in MBConv and improving the bidirectional feature pyramid network BiFPN structure. S6. Input the data set into the improved EfficientDet chemical plant safety detection model for training and testing the model performance; S7. Use the trained model to perform target detection on the chemical plant images collected on site to obtain the workers’ protective equipment wearing and personal behavior; S8. Determine whether dangerous behavior occurs based on the target detection results, and display the safety status of the chemical plant in real time through the front-end interface. When there is an abnormal situation, the early warning device is automatically triggered to send out an alarm signal; The improved adaptive wavelet threshold denoising described in step S2 has the following specific process: S21. Let s(x,y) be the original image signal, n(x,y) be the Gaussian white noise signal, and f(x,y) be the image signal interfered by the noise. We can get: f(x,y)=s(x,y)+n(x,y) Where x, y = 1, 2, 3…N + ; S22, selecting a wavelet basis to perform multi-scale wavelet decomposition on the image, and obtaining a series of wavelet coefficients of high-frequency and low-frequency components at different scales; S23. Select threshold λ for different decomposition scales: Where σ is the standard deviation of the noise; N is the image length; o is the number of decomposition layers; S24, threshold function for high frequency coefficient w x,y Perform threshold quantization to obtain high-frequency estimation coefficients S25. Propose an adaptive threshold function: in, is the high-frequency estimated coefficient; w x,y is the high frequency coefficient; λ is the threshold; sgn() is the sign function; R is the adjustment factor; S26, the adjustment factor R in the new threshold function affects the denoising effect, so the optimal adjustment factor R is found through the predator optimization algorithm; S27, reconstructing the low-frequency coefficients after wavelet decomposition and the high-frequency coefficients after threshold quantization of each layer to obtain a denoised image; The improved PSA attention mechanism has the following specific process: S521, dividing the channel by squeezing and splicing SPC modules, and performing multi-scale feature extraction on the spatial information on each channel feature map; S522, using the SEWeight module to extract the channel attention of feature maps of different scales, and obtain the channel attention vector at each different scale; S523, the original Softmax function is replaced by the Sparsemax function; the multi-scale channel attention vector is recalibrated using Sparsemax to obtain the new attention weight after multi-scale channel interaction; S524, performing a dot multiplication operation on the recalibrated weights and the corresponding feature map element by element, and outputting a feature map after multi-scale feature information attention weighting; The improved bidirectional feature pyramid network BiFPN structure described in step S5 has the following specific process: S531, improved BiFPN fusion intermediate feature information, in the top-down path, will obtain intermediate information S532. In the bottom-up path, use intermediate information of different scales to obtain the output of the network: in, is the output feature of the jth layer, P j is the input feature of the jth layer, is the intermediate feature information of the jth layer, ω is the weight, Res is the upsampling or downsampling operation matching the current resolution, Conv is the convolution operation used for feature processing, and ε is the correction value.

2. According to claim 1, a chemical plant area intelligent safety detection method based on deep learning is characterized in that: The predator optimization algorithm described in step S26 searches for the optimal adjustment factor R, and the specific process is as follows: S261, randomly initialize the population position, that is, the adjustment factor R in the threshold function; S262. Define the fitness function as F: F=SNR+SSIM In the formula, SNR is the signal-to-noise ratio, SSIM is the structural similarity; S263, calculating the fitness value and recording the global optimum; S264, updating the location of prey populations; S265, determine whether the maximum number of iterations has been reached, if not, repeat steps S262 and S263 until the termination condition is reached; S266. Output the particle corresponding to the global optimal solution, and substitute the optimal population position as the adjustment factor R into the new threshold function.

3. According to the deep learning-based intelligent safety detection method for chemical plants in claim 1, it is characterized in that: The step S7 obtains the worker's protective equipment wearing status and personal behavior, including: not wearing personal protective equipment, mistakenly entering a dangerous area to perform high-risk operations, and smoking in a non-smoking area; the personal protective equipment includes a safety helmet, safety glasses, protective gloves, and safety shoes.

4. According to the deep learning-based intelligent safety detection method for chemical plants in claim 1, it is characterized in that: The front-end interface in step S8 displays the safety status of the chemical plant in real time, and the detection results will be displayed in real time on the front-end interface, including the workers' protective equipment wearing status and their specific location information.

5. A chemical plant intelligent safety detection system based on deep learning, characterized in that: It includes acquisition module, preprocessing module, EfficientDet chemical plant safety detection model and detection module; The acquisition module installs multiple cameras in areas of the chemical plant where there are potential safety hazards to obtain images of workers on site; The preprocessing module uses improved adaptive wavelet threshold denoising to process the original image; generates image data through CycleGAN; annotates the head, torso, and limbs in the preprocessed image to construct a data set; The EfficientDet chemical plant safety detection model includes building an improved EfficientDet chemical plant safety detection model, wherein the improvement includes introducing an improved PSA attention mechanism in MBConv and improving a bidirectional feature pyramid network BiFPN structure; inputting a data set into the improved EfficientDet chemical plant safety detection model for training and testing model performance; The detection module uses the trained model to perform target detection, detects the chemical plant images collected on site, and obtains the workers' protective equipment wearing and personal behavior; determines whether dangerous behavior occurs based on the target detection results, and displays the safety status of the chemical plant in real time through the front-end interface. When there is an abnormal situation, the early warning device is automatically triggered to send out an alarm signal; The improved adaptive wavelet threshold denoising is used to process the original image, and the specific process is as follows: Assume s(x,y) is the original image signal, n(x,y) is the Gaussian white noise signal, and f(x,y) is the image signal interfered by the noise. We can get: f(x,y)=s(x,y)+n(x,y) Where x, y = 1, 2, 3…N + ; Select wavelet basis to perform multi-scale wavelet decomposition on the image, and obtain a series of wavelet coefficients of high-frequency and low-frequency components at different scales; For different decomposition scales, select the threshold λ: Where σ is the standard deviation of the noise; N is the image length; o is the number of decomposition layers; Threshold function for high frequency coefficient w x,y Perform threshold quantization to obtain high-frequency estimation coefficients Propose an adaptive threshold function: in, is the high-frequency estimated coefficient; w x,y is the high frequency coefficient; λ is the threshold; sgn() is the sign function; R is the adjustment factor; The adjustment factor R in the new threshold function affects the denoising effect, so the optimal adjustment factor R is found through the predator optimization algorithm; Reconstruct the low-frequency coefficients after wavelet decomposition and the high-frequency coefficients after threshold quantization of each layer to obtain the denoised image; The predator optimization algorithm described above searches for the optimal adjustment factor R. The specific process is as follows: Randomly initialize the population position, that is, the adjustment factor R in the threshold function; Define the fitness function as F: F=SNR+SSIM In the formula, SNR is the signal-to-noise ratio, SSIM is the structural similarity; Calculate the fitness value and record the global optimum; Update the location of prey populations; Determine whether the maximum number of iterations has been reached. If not, repeat the optimization until the termination condition is reached. Output the particles corresponding to the global optimal solution, and substitute the optimal population position as the adjustment factor R into the new threshold function; The improved PSA attention mechanism has the following specific process: S521, dividing the channel by squeezing and splicing SPC modules, and performing multi-scale feature extraction on the spatial information on each channel feature map; S522, using the SEWeight module to extract the channel attention of feature maps of different scales, and obtain the channel attention vector at each different scale; S523, the original Softmax function is replaced by the Sparsemax function; the multi-scale channel attention vector is recalibrated using Sparsemax to obtain the new attention weight after multi-scale channel interaction; S524, performing a dot multiplication operation on the recalibrated weights and the corresponding feature map element by element, and outputting a feature map after multi-scale feature information attention weighting; The improved bidirectional feature pyramid network BiFPN structure has the following specific process: S531, improved BiFPN fusion intermediate feature information, in the top-down path, will obtain intermediate information S532. In the bottom-up path, use intermediate information of different scales to obtain the output of the network: in, is the output feature of the jth layer, P j is the input feature of the jth layer, is the intermediate feature information of the jth layer, ω is the weight, Res is the upsampling or downsampling operation matching the current resolution, Conv is the convolution operation used for feature processing, and ε is the correction value.

Images