Frequency-height Map Denoising and Recognition Method Based on DnCNN and YOLOv8

Abstract

The present invention provides a frequency-height map denoising and recognition method based on DnCNN and YOLOv8. The method includes: inputting an ionospheric sounding image into the DnCNN model to obtain a denoised clear image; inputting the denoised clear image into the YOLOv8 model to obtain recognized objects and output ionospheric parameters. Through the precisely adjusted DnCNN model, the noise in the ionospheric image is effectively removed, and the key echo information is retained, providing high-quality input for subsequent image processing; through the fine-tuning of the YOLOv8 model, the resource consumption is reduced, and the efficiency of model training and the accuracy of object recognition are improved; the automation and intelligence of ionospheric sounding image processing are realized, and the processing efficiency and practicality are improved; it is applicable to scientific research and monitoring applications that require high-precision ionospheric image analysis, and has important practical value and scientific significance.

Description

Technical Field

[0001] The present invention belongs to the field of ionogram processing, and particularly relates to a method for ionogram denoising and recognition based on DnCNN and YOLOv8. Background Art

[0002] The ionosphere is a region in the Earth's atmosphere from about 60 kilometers to 1000 kilometers in altitude, where the air is ionized by solar radiation to form charged particles. This layer has an important impact on the propagation of radio waves and is an important channel for electromagnetic communication between the Earth and space. Disturbances in the ionosphere can directly affect the positioning accuracy of global navigation satellite systems, as well as the stability and reliability of shortwave communication.

[0003] The ionogram is an important tool for studying the ionosphere, obtained through an ionospheric vertical sounder. The horizontal and vertical coordinates represent frequency and virtual height respectively, and the signal value at each coordinate point represents the echo intensity. The ionogram provides a basic description of ionospheric characteristics such as electron density and height, and shows the relationship between the reflection layer height and frequency. However, there is a large amount of noise in the ionogram, resulting in a low signal-to-noise ratio and affecting the accurate analysis and application of ionospheric characteristics.

[0004] Existing ionogram processing methods mainly include traditional image processing and deep learning methods. Traditional methods such as the ARTIST algorithm represent the electron density profile of each layer as a shifted Chebyshev polynomial. Although certain results have been achieved, it is difficult to adapt to the complexity and variability of the ionospheric structure. Other methods use digital image processing, mathematical morphology, and graph theory techniques to measure the ionogram, but these methods rely on geographical location information, and the processing rules vary greatly in different regions. Some techniques such as Steger curve tracking and statistical feature recognition, although not dependent on time, location, and season, require complex adjustments when extended to other layers and are vulnerable to noise. Generally speaking, traditional methods often rely on manual feature selection, are difficult to adapt to the complex structure of the ionogram, have weak generalization performance, and are easily affected by noise. With the development of deep learning technology, various image classification, semantic segmentation, and object detection algorithms have been applied to the ionogram measurement task. However, the performance of these methods is limited by the quantity and quality of the labeled data, and it is difficult to obtain valuable results for images with excessive noise.

[0005] Common problems of traditional and deep learning methods include: difficulty in processing complex structures and diverse datasets, vulnerability to noise interference, and dependence on manual feature selection or specific statistical information, with weak generalization performance. Summary of the Invention

[0006] The object of the present invention is to overcome the above problems and provide a frequency-height map denoising and recognition method based on DnCNN and YOLOv8. The present invention solves the problems of large noise interference and inaccurate extraction of key parameters in the existing ionospheric sounding image processing technology.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A frequency-height map denoising and recognition method based on DnCNN and YOLOv8, the method comprising the following steps:

[0009] Input the ionospheric sounding image into the DnCNN model to obtain a denoised clear image;

[0010] Among them, the training method of the DnCNN model comprises the following steps: training the DnCNN basic model using a denoising dataset, and using the mean square error as the loss function;

[0011] Input the denoised clear image into the YOLOv8 model to obtain the recognized object and output the ionospheric parameters;

[0012] Among them, the training method of the YOLOv8 model comprises the following steps: pre-training the YOLOv8 basic model using natural images, and fine-tuning the pre-trained YOLOv8 basic model using the cross-entropy loss function and the bounding box regression loss function.

[0013] As a further improvement, the construction of the denoising dataset comprises the following steps:

[0014] Collect ionospheric sounding images under different geographical locations and various climate conditions, and divide the ionospheric sounding images into a first group of ionospheric sounding images and a second group of ionospheric sounding images;

[0015] Perform artificial denoising and echo information retention processing on the first group of ionospheric sounding images to obtain a third group of ionospheric sounding images;

[0016] Pair the second group of ionospheric sounding images and the third group of ionospheric sounding images to construct a denoising dataset for model supervised training; among them, the denoising dataset is 80% training set and 20% validation set for model training and performance evaluation.

[0017] As a further improvement, the training method of the YOLOv8 model further comprises the following steps:

[0018] Adopt the Adam optimizer and early stopping method for DnCNN model training;

[0019] Among them, during the training of the DnCNN model, when the DnCNN model has completed consecutive full traversals of the dataset and the loss on the validation set no longer decreases, the training stops; among them, after each full traversal of the dataset by the DnCNN model, the denoising performance of the DnCNN model is evaluated on the validation set.

[0020] As a further improvement, during the training of the DnCNN model, the following steps are also included:

[0021] Save the parameters of the DnCNN model that performs best on the validation set;

[0022] Load the parameters of the best DnCNN model on the validation set and evaluate it on the test set.

[0023] As a further improvement, after obtaining the denoised clear image, the denoised clear image needs to be labeled, including the following steps:

[0024] Label the metrics of the denoised clear image;

[0025] Pair the results after manual metric labeling with the denoised clear image to create an object detection dataset for the supervised training of the YOLOv8 model; among them, the object detection dataset is 60% training set, 20% validation set, and 20% test set, which is used for the training and performance evaluation of the model.

[0026] As a further improvement, the training method of the YOLOv8 model also includes the following steps:

[0027] During the fine-tuning of the YOLOv8 model, when the YOLOv8 model has completed consecutive full traversals of the dataset and the loss on the validation set no longer decreases, the training stops; among them, after each full traversal of the dataset by the YOLOv8 model, the metric accuracy of the YOLOv8 model is evaluated on the validation set.

[0028] As a further improvement, during the fine-tuning of the YOLOv8 model, the following steps are also included:

[0029] Save the parameters of the YOLOv8 model that performs best on the validation set;

[0030] Load the parameters of the best YOLOv8 model and evaluate it on the test set.

[0031] As a further improvement, the training method of the YOLOv8 model further includes the following steps:

[0032] Based on the characteristics of the ionospheric image, adjust the output layer of the YOLOv8 model to align the number of categories in the classification head with the number of categories in the ionospheric image layer.

[0033] As a further improvement, the following steps are further included:

[0034] Convert the ionospheric parameters of the YOLOv8 model into critical frequency and virtual height;

[0035] Among them, the formula for converting the critical frequency and virtual height is:

[0036]

[0037] In the formula, f represents the critical frequency, h represents the virtual height, δf represents the unit measure of the horizontal axis of the ionogram, δh represents the unit measure of the vertical axis of the ionogram, f0 represents the starting frequency of the horizontal and vertical axes of the ionogram, and h0 represents the starting height of the horizontal and vertical axes of the ionogram. represents the abscissa of the center of the bounding box relative to the upper left corner of the image, represents the abscissa of the center of the bounding box relative to the upper left corner of the image, represents the width of the bounding box, represents the height of the bounding box.

[0038] On the other hand, the present invention also provides a frequency-height map denoising and recognition system based on DnCNN and YOLOv8. The system includes at least one processor and a memory that stores instructions. When the instructions are executed by at least one processor, the method in the above technical solution is implemented.

[0039] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method described in the steps of the above technical solution.

[0040] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in the steps of the above technical solution is implemented.

[0041] The beneficial effects of the present invention are as follows: Through the precisely adjusted DnCNN model, the noise in the ionospheric image is effectively removed, and the key echo information is retained, providing high-quality input for subsequent image processing; By fine-tuning the YOLOv8 model, resource consumption is reduced, and the efficiency of model training and the accuracy of object recognition are improved; The automation and intelligence of ionospheric detection image processing are realized, improving the processing efficiency and practicality; It is applicable to scientific research and monitoring applications that require high-precision ionospheric image analysis, and has important practical value and scientific significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention has the following drawings:

[0043] Figure 1It is a flowchart of the frequency-height map denoising and recognition method based on DnCNN and YOLOv8 in some specific embodiments of the present invention;

[0044] Figure 2 It is a flowchart of manual denoising in some specific embodiments of the present invention;

[0045] Figure 3 It is a flowchart of manual measurement in some specific embodiments of the present invention;

[0046] Figure 4 Schematic diagram of the system for frequency-height map denoising and recognition based on DnCNN and YOLOv8 of the present invention. Specific embodiments

[0047] The following embodiments further illustrate the content of the present invention, but should not be construed as limiting the present invention. Without departing from the spirit and essence of the present invention, any modification or replacement of the methods, steps or conditions of the present invention shall fall within the scope of the present invention.

[0048] Embodiment 1:

[0049] As Figures 1-3 shown, the present invention provides a frequency-height map denoising and recognition method based on DnCNN and YOLOv8, and the method includes the following steps:

[0050] Step S101: Input the ionospheric sounding image into the DnCNN model to obtain a denoised clear image;

[0051] Among them, the training method of the DnCNN model includes the following steps: training the DnCNN basic model with a denoising data set, and using the mean square error as the loss function;

[0052] Step S102: Input the denoised clear image into the YOLOv8 model to obtain the recognition object and output the ionospheric parameters;

[0053] For example, input the denoised clear image into the fine-tuned YOLOv8 model, and the model not only recognizes the target object in the image, but also accurately measures and outputs the required ionospheric parameters.

[0054] Among them, the training method of the YOLOv8 model includes the following steps: pre-training the YOLOv8 basic model with natural images, and fine-tuning the pre-trained YOLOv8 basic model using the cross-entropy loss function and the bounding box regression loss function.

[0055] Among them, the formula of the loss function is

[0056] is the total loss, is the classification loss, is the localization loss, is the object loss; λ CIOU , λ obj , λ cls are the weight coefficients used to balance the contributions of each part of the loss function.

[0057] b is the ground truth bounding box, the predicted bounding box CIOU: represents the complete intersection over union between the predicted bounding box and the ground truth bounding box.

[0058] y represents the true label, which is 1 if there is an object, otherwise 0; represents the probability that the model predicts the current box contains an object.

[0059] C is the total number of classes; if the true class of the object is c, it is 1, otherwise 0; The probability that the model predicts the current box belongs to class c.

[0060] For example, fine-tuning the pre-trained YOLOv8 model: Use the YOLOv8 model pre-trained on natural images and fine-tune it based on the characteristics of ionospheric images; When fine-tuning, use the cross-entropy loss function and the bounding box regression loss to optimize the model to ensure that the model can accurately identify and locate the key features in ionospheric images.

[0061] The construction of the denoising dataset includes the following steps:

[0062] Step S1011: Collect ionospheric sounding images under different geographical locations and various climate conditions, and divide the ionospheric sounding images into a first group of ionospheric sounding images and a second group of ionospheric sounding images;

[0063] For example, collecting and selecting ionograms: Collect ionospheric sounding images under different geographical locations and various climate conditions, and select some images as the basis of the denoising dataset.

[0064] Step S1012: Perform artificial denoising and echo information retention processing on the first group of ionospheric sounding images to obtain a third group of ionospheric sounding images;

[0065] For example, artificial processing and annotation: Perform artificial processing on the selected images, including removing noise and retaining echo information; This can ensure that the model can focus on noise features in subsequent training.

[0066] Step S1013: Pair the second group of ionospheric sounding images and the third group of ionospheric sounding images to construct a denoising dataset for model supervised training; wherein, the denoising dataset is 80% training set and 20% validation set for model training and performance evaluation.

[0067] For example, establish a training and validation dataset: Pair the clean images after manual processing with the original noisy images to create a denoising dataset for model supervised training; the dataset is divided into 80% training set and 20% validation set for model training and performance evaluation.

[0068] The training method of the YOLOv8 model further includes the following steps:

[0069] Use the Adam optimizer and early stopping method to train the DnCNN model;

[0070] Wherein, during the training of the DnCNN model, when the DnCNN model traverses the dataset completely for multiple consecutive times and the loss on the validation set no longer decreases, stop the training; wherein, after each complete traversal of the dataset by the DnCNN model, evaluate the denoising performance of the DnCNN model on the validation set.

[0071] For example, use the Adam optimizer to train the model, set the learning rate to 0.001, and use batches of size 32 for iteration; use the early stopping method during training to avoid overfitting, and regularly evaluate the model performance on the validation set to adjust the learning rate.

[0072] During the training of the DnCNN model, the following steps are further included:

[0073] During the training of the DnCNN model, save the parameters of the DnCNN model that performs best on the validation set;

[0074] Load the best DnCNN model parameters on the validation set and evaluate them on the test set.

[0075] After obtaining the denoised clear images, the denoised clear images need to be labeled, including the following steps:

[0076] Label the metrics of the denoised clear images;

[0077] For example, label the denoised images: Further label all the output denoised images to clarify the key objects in the images and their corresponding ionospheric parameters, such as virtual height and critical frequency.

[0078] Pair the results after manual annotation measurement with the denoised clear image (referring to the image output by DnCNN. Since DnCNN is a denoising model, the image it outputs is called the denoised clear image. Pair the denoised clear image after annotation with the denoised clear image before annotation) to create an object detection dataset for the supervised training of the YOLOv8 model; among them, the object detection dataset is 60% training set, 20% validation set, and 20% test set, which is used for the training and performance evaluation of the model.

[0079] The training method of the YOLOv8 model further includes the following steps:

[0080] During the fine-tuning of the YOLOv8 model, when the loss on the validation set no longer decreases after the YOLOv8 model has completed multiple consecutive full traversals of the dataset; among them, after each full traversal of the dataset by the YOLOv8 model, the measurement accuracy of the YOLOv8 model is evaluated on the validation set.

[0081] During the fine-tuning of the YOLOv8 model, it further includes the following steps:

[0082] During the fine-tuning of the YOLOv8 model, save the parameters of the YOLOv8 model that performs best on the validation set;

[0083] Load the parameters of the best YOLOv8 model and evaluate them on the test set.

[0084] The training method of the YOLOv8 model further includes the following steps:

[0085] Based on the characteristics of the ionospheric image, adjust the output layer of the YOLOv8 model to align the number of categories in the classification head with the number of categories in the ionospheric image layer, where the output of the YOLOv8 model is

[0086]

[0087] In the formula, represents the abscissa of the center of the bounding box relative to the upper left corner of the picture, represents the ordinate of the center of the bounding box relative to the upper left corner of the picture, represents the width size of the bounding box, represents the height size of the bounding box, represents the object confidence, represents the classification confidence of each class, and C represents the total number of classes.

[0088] In some embodiments of the present invention, based on the measurement method in the frequency-height map denoising and recognition method of DnCNN and YOLOv8, it further includes the following steps:

[0089] Convert the ionospheric parameters of the YOLOv8 model into critical frequency and virtual height;

[0090] Among them, the formula for converting the critical frequency and virtual height is:

[0091]

[0092] In the formula, f represents the critical frequency, h represents the virtual height, δf represents the unit measure of the horizontal axis of the ionogram, δh represents the unit measure of the vertical axis of the ionogram, f0 represents the starting frequency of the horizontal and vertical axes of the ionogram, and h[0] represents the starting height of the horizontal and vertical axes of the ionogram. represents the abscissa of the center of the bounding box relative to the upper left corner of the picture, represents the abscissa of the center of the bounding box relative to the upper left corner of the picture, represents the width of the bounding box, represents the height of the bounding box.

[0093] Example 2:

[0094] The present invention provides a method for ionogram denoising and recognition based on the DnCNN model and the YOLOv8 model. The method includes the following steps:

[0095] Step S1: Input the original ionogram into the pre-established and trained DnCNN model. The DnCNN model learns the noise characteristics of the ionogram through multiple convolutional layers and performs denoising processing to output a high-quality denoised ionogram;

[0096] Step S2: Use the YOLOv8 model to perform object detection on the denoised ionogram;

[0097] Step S3: The YOLOv8 model uses convolutional layers to extract features and locate objects in the ionogram;

[0098] Step S4: Use the YOLOv8 model to extract different layer parameters in the ionogram, including virtual height and critical frequency, etc. The model effectively identifies the key features and parameters in the ionogram through accurate bounding box positioning.

[0099] In the present invention, a method combining the DnCNN and YOLOv8 models is proposed. The DnCNN is used for denoising the frequency-height diagram to improve the image quality and signal-to-noise ratio. Then, the YOLOv8 model is used for target detection and parameter extraction to enhance the detection accuracy and efficiency. The YOLOv8 model can accurately identify and extract different layer parameters in the frequency-height diagram through precise bounding box localization, and is suitable for the analysis and application of complex frequency-height diagrams. The method of the present invention is applicable to various ionospheric monitoring scenarios, has broad application prospects and practical value. Through the method of the present invention, the signal-to-noise ratio and detection accuracy of the frequency-height diagram can be effectively improved, providing reliable technical support for ionospheric monitoring and communication system optimization.

[0100] Embodiment 3:

[0101] The present invention provides a method for denoising, identifying and measuring frequency-height diagrams based on the DnCNN model and the YOLOv8 model. The specific flowchart is as Figure 1 shown, including the training processes of the DnCNN denoising model and the YOLOv8 target detection model, as well as the denoising and target detection tasks of the frequency-height diagram using these models:

[0102] I. The training process of the DnCNN denoising model specifically includes the following steps:

[0103] Step S11: Collect ionospheric frequency-height diagrams obtained at different geographical locations around the world under various climate conditions to ensure the diversity and extensiveness of the data.

[0104] Step S12: Select representative images from the collected data, especially those with a high noise level, for training the denoising model.

[0105] Step S13: Artificially add different types of noise (such as Gaussian noise) to the selected images to form more noisy images.

[0106] Step S14: Perform artificial denoising on the selected images, fill in the breakpoints of discontinuous signals, and retain key echo information for supervising the training of the denoising model DnCNN. For specific denoising examples, refer to Figure 2 .

[0107] Step S15: Establish a training and validation dataset. Pair the artificially denoised images with the original noisy images (I, I denoise ), and create a denoising dataset for the supervised training of the DnCNN model. The dataset is divided into a 60% training set, a 20% validation set, and a 20% test set for model training and performance evaluation.

[0108] Step S16: Select the DnCNN model as the denoising basis. This model effectively learns the noise distribution through a deep convolutional network and performs denoising. Among them, the output of the convolutional layer is

[0109] h l = ReLU(BN(W l * h l-1 + b l )): ReLU(x) = max(0, x), and h l is the output of the l

[0110] -th layer of convolution, W l is the convolution kernel of the l-th layer of convolution, b l is the bias of the l-th layer of convolution, and * is the convolution operation; μ and σ 2 are the mean and variance of this batch, and γ and β are the scaling and offset factors.

[0111] Step S17: Adopt the mean squared error (MSE) as the loss function

[0112] N is the batch size, y i is the original clean image, and f(x i ; θ) is the image after denoising by the model.

[0113] Directly measure the difference between the denoised image and the original clean image.

[0114] Step S18: Use the Adam optimizer for model training. Its adaptive learning rate can accelerate the convergence speed and improve the denoising effect. The specific method of the Adam optimizer is as follows: First, θ0 = 0, m0 = 0, v0 = 0, where θ0 is the initialization parameter of the model, and m0 and v0 are the first-order and second-order moment estimates of the initial gradient; then, calculate the gradient of the current loss function and update the moment estimate, β1 and β2 are the decay rates of the first-order and second-order moment estimates; correct the moment estimate, Use the moment estimate of the gradient to update the parameters of the model α is the learning rate, and ∈ avoids the denominator being zero.

[0115] Step S19: Implement early stopping during training to prevent overfitting. Stop training when the loss on the validation set no longer decreases for multiple consecutive epochs (the number of times the model completely traverses the dataset).

[0116] Step S20: After each epoch of training (the number of times the model completely traverses the dataset), evaluate the denoising performance of the model on the validation set to ensure the generalization ability of the model, and save the model parameters that perform best on the validation set during the training process.

[0117] Step S21: Load the best model parameters on the validation set and evaluate them on the test set.

[0118] Step S22: Input all the original ionospheric frequency-height maps in Step S11 into the model to obtain high-quality automatically denoised frequency-height maps corresponding to all the original ionospheric frequency-height maps.

[0119] II. Training and fine-tuning of the YOLOv8 model object detection model:

[0120] Step S23: Perform further annotation and measurement on all the denoised images output in Step S22 to clarify the key objects in the images and their corresponding ionospheric parameters, such as virtual height and critical frequency. For specific annotation examples, refer to Figure 3 .

[0121] Step S24: Establish the training and validation datasets. Pair the results after manual annotation and measurement with the denoised clear images to create an object detection dataset for the supervised training of the YOLOv8 model. The dataset is divided into a 60% training set, a 20% validation set, and a 20% test set for the training and performance evaluation of the model.

[0122] Step S25: Download the YOLOv8 model.

[0123] Step S26: According to the characteristics of the ionospheric images, adjust the output layer of the YOLOv8 model to align the number of categories in the classification head with the number of categories in the ionospheric images.

[0124] Output of the YOLOv8 model

[0125]

[0126] Among them: are the coordinates of the center of the bounding box relative to the upper left corner of the image;

[0127] are the width and height of the bounding box;

[0128] is the object confidence;

[0129] are the classification confidences for each class;

[0130] C is the total number of classes.

[0131] Step S27: Download and load the pre-trained weights on the COCO natural image dataset (natural images refer to images that actually exist in nature, which are generated by photographing natural processes or phenomena and are not created by artificial or technical means), and use the complex visual features it has learned on a wide range of datasets to provide a powerful starting point for specific tasks of ionospheric images.

[0132] Step S28: Fine-tune the pre-trained YOLOv8 model, and use the cross-entropy loss function and bounding box regression loss to optimize the model. The cross-entropy loss function and bounding box regression loss function are as follows:

[0133] is the total loss, is the classification loss, is the localization loss, is the object loss; λ CIOU , λ obj , λ cls are weight coefficients used to balance the contributions of each part of the loss function.

[0134] b is the ground truth bounding box, the predicted bounding box CIOU: represents the complete intersection over union between the predicted bounding box and the ground truth bounding box.

[0135] y represents the true label, which is 1 if there is an object and 0 otherwise; represents the probability that the model predicts that the current box contains an object.

[0136] C is the total number of classes; it is 1 if the true class of the object is c and 0 otherwise; the probability that the model predicts that the current box belongs to class c.

[0137] Step S29: Implement early stopping during the fine-tuning process to prevent overfitting, and stop training when the loss on the validation set no longer decreases for multiple consecutive epochs (the number of times the model completely traverses the dataset).

[0138] Step S30: Evaluate the metric accuracy of the YOLOv8 model on the validation set after each epoch (the number of times the model completely traverses the dataset) of training, ensure the generalization ability of the model, and save the model parameters that perform best on the validation set during the fine-tuning process.

[0139] Step S31: Load the best model parameters and evaluate on the test set.

[0140] Step S32: Train using the SGD optimizer with momentum. Initialize the learning rate to 0.0001 and the momentum factor to 0.9. The specific method of the SGD optimizer is as follows: First, calculate the gradient of the loss function at the parameter θ t-1 (the parameter value at the previous moment). of the loss function Then, use the momentum V t-1 from the previous time step t and the current gradient g t to update the momentum V according to the momentum factor γ and the learning rate α t-1 = γV t + αg t ; Finally, use the updated momentum V t to update the parameter θ t-1 = θ t - V

[0141] Step S33: Input all the denoised ionogram frequency-height maps in Step S22 into the model to obtain the automatic measurement results of all ionogram frequency-height maps.

[0142] Step S34: Through mathematical calculations, convert the object detection results of YOLOv8 into numerical results with physical significance.

[0143]

[0144] where f is the critical frequency, h is the virtual height, δf and δh are the unit measurements of the horizontal and vertical axes of the ionogram frequency-height map respectively, f0 and h0 are the starting frequencies and starting heights of the horizontal and vertical axes of the ionogram frequency-height map respectively. For obtaining the calculation parameters, refer to Figure 2 .

[0145] Through the precisely tuned DnCNN model, the noise in the ionosphere image is effectively removed, and the key echo information is retained, providing high-quality input for subsequent image processing; through the fine-tuning of the YOLOv8 model, the resource consumption is reduced, and the efficiency of model training and the accuracy of object recognition are improved; the automation and intelligence of ionosphere detection image processing are realized, and the processing efficiency and practicality are improved; it is applicable to scientific research and monitoring applications that require high-precision ionosphere image analysis, and has important practical value and scientific significance.

[0146] In some specific embodiments, as Figure 4 shown, a system for denoising and identifying frequency-height maps based on the DnCNN model and the YOLOv8 model includes at least one processor and a memory that stores instructions to implement the method for denoising and identifying frequency-height maps based on DnCNN and YOLOv8 in any of the above technical solutions.

[0147] In some specific embodiments, a computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method of the steps in the above technical solution.

[0148] In some specific embodiments, a computer program product includes a computer program, and when the computer program is executed by a processor, it implements the method of the steps in the above technical solution.

[0149] Embodiments and functional operations of the subject matter described in this specification can be implemented in: digital electronic circuits, tangible computer software or firmware, computer hardware, including the structures disclosed in this specification and their structural equivalents, or a combination of one or more of the above. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, that is, one or more modules of computer program instructions encoded on one or more tangible non-transitory program carriers, for being executed by a data processing device or controlling the operation of a data processing device.

[0150] As an alternative or in addition, program instructions can be encoded on an artificially generated propagated signal, for example, a machine-generated electrical signal, optical signal, or electromagnetic signal, which is generated as encoded information to be transmitted to an appropriate receiver device for execution by a data processing device. A computer storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of the above devices.

[0151] A computer program (which can also be referred to as or described as a program, software, software application, module, software module, script, or code) can be written in any form of programming language, including a compiled language, an interpreted language, a declarative language, or a procedural language, and the computer program can be deployed in any form, including as an independent program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program may or may not correspond to a file in a file system. The program can be stored as part of a file that stores other programs or data, for example, stored in one or more of the following scripts: in a markup language document; in a single file dedicated to the relevant program; or in multiple cooperating files, for example, files that store one or more modules, subroutines, or code portions. A computer program can be deployed to execute on one computer or multiple computers, which are located in one place, or distributed to multiple locations and interconnected through a communication network.

[0152] The content not described in detail in this specification belongs to the prior art well known to those skilled in the art.

[0153] Although the present invention has been described in detail above with general descriptions, specific embodiments and experiments, modifications or improvements can be made thereto based on the present invention, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of the present invention claimed.

Claims

1. A method for frequency-height map denoising and recognition based on DnCNN and YOLOv8, characterized in that, It includes the following steps: Input the ionospheric sounding image into the DnCNN model to obtain a denoised clear image; wherein, the training method of the DnCNN model includes the following steps: training the DnCNN basic model using a denoising dataset, and using the mean square error as the loss function; Input the denoised clear image into the YOLOv8 model to obtain the recognized object and output the ionospheric parameters; Wherein, the training method of the YOLOv8 model includes the following steps: pre-training the YOLOv8 basic model using natural images, and fine-tuning the pre-trained YOLOv8 basic model using the cross-entropy loss function and the bounding box regression loss function; wherein, The construction of the denoising dataset includes the following steps: Collect ionospheric sounding images under different geographical locations and various climate conditions, and divide the ionospheric sounding images into a first group of ionospheric sounding images and a second group of ionospheric sounding images; Perform artificial denoising and echo information retention processing on the first group of ionospheric sounding images to obtain a third group of ionospheric sounding images; Pair the second group of ionospheric sounding images and the third group of ionospheric sounding images to construct a denoising dataset for model supervised training; wherein, the denoising dataset is 80% training set and 20% validation set, which is used for model training and performance evaluation; It further includes the following steps: Convert the ionospheric parameters of the YOLOv8 model into critical frequency and virtual height; Wherein, the formula for converting the critical frequency and virtual height is: In the formula, f represents the critical frequency, h represents the virtual height, δf represents the unit measure of the horizontal axis of the ionogram, δh represents the unit measure of the vertical axis of the ionogram, f0 represents the starting frequency of the horizontal and vertical axes of the ionogram, and h0 represents the starting height of the horizontal and vertical axes of the ionogram. represents the abscissa of the center of the bounding box relative to the upper left corner of the picture. represents the abscissa of the center of the bounding box relative to the upper left corner of the picture. represents the width of the bounding box. represents the height of the bounding box.

2. The frequency-height map denoising and recognition method based on DnCNN and YOLOv8 according to claim 1, wherein The training method of the YOLOv8 model further includes the following steps: Use the Adam optimizer and early stopping method to train the DnCNN model; wherein, during the training of the DnCNN model, when the loss on the validation set no longer decreases after the DnCNN model has completed multiple complete traversals of the dataset, stop training; wherein, after each complete traversal of the dataset by the DnCNN model, evaluate the denoising performance of the DnCNN model on the validation set.

3. The frequency-height map denoising and recognition method based on DnCNN and YOLOv8 according to claim 2, wherein, During the training of the DnCNN model, it further includes the following steps: Save the DnCNN model parameters that perform best on the validation set; Load the best DnCNN model parameters on the validation set and evaluate them on the test set.

4. The frequency-height map denoising and recognition method based on DnCNN and YOLOv8 according to claim 1, characterized in that, After obtaining the denoised clear image, it is necessary to label the denoised clear image, including the following steps: Label the denoised clear image metrics; Pair the results after artificial annotation of the metrics with the denoised clear image to create an object detection dataset for YOLOv8 model supervised training; wherein, the object detection dataset is 60% training set, 20% validation set and 20% test set, which is used for model training and performance evaluation.

5. The frequency-height map denoising and recognition method based on DnCNN and YOLOv8 according to claim 1, characterized in that, The training method of the YOLOv8 model further includes the following steps: During the fine-tuning of the YOLOv8 model, when the loss on the validation set no longer decreases after the YOLOv8 model has completed multiple complete traversals of the dataset, stop training; wherein, after each complete traversal of the dataset by the YOLOv8 model, evaluate the metric accuracy of the YOLOv8 model on the validation set.

6. The frequency-height map denoising and recognition method based on DnCNN and YOLOv8 according to claim 5, characterized in that, During the fine-tuning of the YOLOv8 model, it further includes the following steps: Save the parameters of the YOLOv8 model that performs best on the validation set; Load the parameters of the best YOLOv8 model and evaluate them on the test set.

7. The denoising and recognition method of frequency-height map based on DnCNN and YOLOv8 according to claim 1, characterized in that, The training method of the YOLOv8 model further includes the following steps: Based on the characteristics of the ionospheric image, adjust the output layer of the YOLOv8 model to align the number of categories in the classification head with the number of categories in the ionospheric image layer.

8. Frequency-height map denoising and recognition system based on DnCNN and YOLOv8, the system includes at least one processor, and a memory storing instructions, characterized in that, When the instruction is executed by at least one processor, the ionogram denoising and recognition method based on DnCNN and YOLOv8 as described in any one of claims 1-7 is implemented.

Images