A method for enhancing the definition of images of endoscopic examinations of the respiratory tract

Abstract

This invention provides a method for enhancing the clarity of respiratory endoscopic images, belonging to the field of image processing. The method includes: downsampling the respiratory endoscopic image, inputting it into a constructed multi-scale feature extraction module for feature extraction and fusion, then inputting it into a constructed frequency-enhanced blur perception module for enhancement, and finally performing transposed convolution and upsampling step by step to fuse the feature information of the shallow and deep layers of the network to obtain the enhanced respiratory endoscopic image. The method achieves efficient noise suppression and accurate detail restoration in respiratory endoscopic images, providing clearer and more accurate image support for clinical diagnosis, and has significant clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and specifically relates to a method for enhancing the clarity of respiratory endoscopic images. Background Technology

[0002] Breathing endoscopy is widely used in the diagnosis and treatment of respiratory diseases. The quality of the images directly affects the accuracy of diagnosis and treatment. However, in actual clinical applications, due to uneven lighting of the endoscope imaging equipment, scattering of biological tissues, and noise of the equipment itself, the obtained images often have serious noise interference and blurring distortion, which seriously affects the doctor's observation and judgment of lesion details and reduces the accuracy of diagnosis and treatment.

[0003] Traditional methods for enhancing endoscopic respiratory images typically employ convolutional neural networks, classical Fourier transforms, or wavelet analysis. These methods suffer from several drawbacks: convolution operations have a fixed local receptive field, easily leading to excessive smoothing of texture details; wavelet analysis can only capture specific frequency components and cannot effectively adapt to complex multi-scale random textures in images; traditional Fourier transforms only process integer-order spectra and cannot finely perceive harmonic attenuation characteristics caused by blurring. In contrast, prior art document CN118134950A discloses a semi-supervised bronchial image segmentation method and device based on dual-perturbation consistency. This method preserves phase information related to high-frequency structural features in the image frequency domain using FFT and alters the amplitude related to low-frequency semantics. While information enhancement is a powerful data augmentation method, it fails to consider the correlation between multi-scale frequency domain features, which can easily lead to the loss of structural information. A prior art document with publication number CN120526063B discloses a three-dimensional bronchial image generation device that outputs global features through multi-level self-attention calculation, outputs local features through convolution and wavelet transform, and finally performs feature fusion. However, this method requires high-quality input images and does not consider the complexities commonly encountered in actual respiratory endoscopic images, such as large variations in lighting conditions and interference from secretions and mucus. Therefore, it is necessary to develop a respiratory endoscopic image denoising and enhancement method that combines non-local multi-scale feature capture with fine frequency perception to effectively improve image quality and diagnostic efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for enhancing the clarity of respiratory endoscopic images. This method combines a multi-scale feature extraction module and a frequency-enhanced blur perception module to achieve noise reduction and enhancement of respiratory endoscopic images, comprising the following steps: S1. Collect respiratory endoscopic images and create a respiratory endoscopic image dataset; S2. The endoscopic images of the respiratory tract are downsampled by 2x and 4x respectively to obtain endoscopic images of the respiratory tract at 2x and 4x downsampled scales. S3. Construct a multi-scale feature extraction module FSDO, including: for any pixel in the input image, calculate the difference between the pixel and other pixels in the neighborhood, combine fractional distance weights and spatial random perturbation factors for weighted integration, perform normalization processing to obtain feature extraction operators, and generate output feature maps by nonlinearly weighted fusion of feature extraction operators at multiple scales. S4. Use FSDO to process respiratory endoscopic images. After processing the lower-scale respiratory endoscopic images, upsample them and fuse them with the features of the upper-scale respiratory endoscopic images to obtain a respiratory endoscopic image feature map. S5. Construct a frequency-enhanced fuzzy perception module AFFRHO, including: performing a two-dimensional fractional Fourier transform on the input feature map to obtain a spectral feature map, enhancing the spectral feature map separately using random harmonic enhancement and frequency enhancement weighting, and then fusing them, converting them into an enhanced feature map through inverse fractional Fourier transform; S6. The AFFRHO is used to process the feature map of the respiratory tract endoscopy image. Then, the transposed convolution operation is used to perform stepwise upsampling, and the feature information of the shallow and deep layers of the skip connection fusion network is combined to obtain the enhanced image of the respiratory tract endoscopy image.

[0005] Preferably, in step S1, a total of 5,000 respiratory endoscopic images are collected, and the images are manually labeled and graded using the labeling tool LabelMe to form a respiratory endoscopic image dataset. Finally, the dataset is divided into a training set and a validation set in a 7:3 ratio.

[0006] Preferably, in step S2, the respiratory endoscopic image is Use it as the original image Then, the images were downsampled by 2x and 4x respectively to obtain 2x downsampled respiratory endoscopic images. and 4x downsampled respiratory endoscopic images The specific process is as follows: , .

[0007] Preferably, in S3, the respiratory endoscopic images are often affected by complex imaging environments such as large changes in lighting conditions and interference from secretions and mucus, resulting in blurred tissue edges and texture features and significant noise interference; at the same time, the internal structures of cavities such as the trachea and bronchi have obvious gray-scale abrupt changes and irregular surface texture distribution.

[0008] Preferably, in step S3, the construction of the multi-scale feature extraction module FSDO specifically includes: S31. For any pixel in the input image In its neighborhood Within the range, the difference between this pixel and other pixels is calculated, and a weighted integral is performed combining fractional distance weights and spatial random perturbation factors. After normalization, the feature extraction operator is obtained, with the specific formula as follows: , in This is a difference term, reflecting the local difference at position (u,v) near point (x,y). If the local pixel changes drastically, this difference term is larger and can capture edges and textures. This is a fractional-order decay term related to distance; the greater the distance, the smaller the contribution. The fractional-order scaling parameter controls the scale of action of the nonlocal operator, with a value range of [0.5, 1.5]. For large-scale features, The value range is [0.5, 0.8], which is suitable for capturing the overall structure of an image. For medium-scale features, The value range is [0.8, 1.2], which is suitable for capturing image texture. For small-scale features, The value range is [1.2, 1.5], which is suitable for capturing image details and edges. To use a small constant, to avoid singularities when the distance approaches 0. It is a spatial random function, which functions to make the multi-scale feature extraction operator... It is more sensitive to local detail changes in an image, especially when the image texture is highly random. It can better magnify detailed features. The normalization coefficient is calculated as follows: ; S32. By nonlinearly weighting and fusing feature extraction operators at various scales, an output feature map is generated. The specific formula is as follows: , in The weights are obtained through training. The `arctan()` function can effectively suppress abrupt changes in features, maintain feature stationarity, and facilitate subsequent processing. For the tanh function, A spatial random signal is used for the feature extraction operator. Random enhancement.

[0009] Preferably, in S3, S31 emphasizes local structural changes through pixel difference terms, and combines fractional distance attenuation kernels to weight and enhance nearby details while weakening distant interference, supporting the restoration of blurred boundaries and enhancement of texture edges in small and medium-scale images. Simultaneously, a spatial random function is introduced to randomly assign perturbation weights to different neighboring sampling points, improving resistance to grayscale unevenness, increasing the diversity and random robustness of feature expression, and avoiding overfitting of features to a particular type of tissue structure. This solves the problem of unstable response in traditional local convolution, which cannot distinguish between real features and noise. The multi-scale feature extraction module FSDO adopts a fractional random differential operator as a whole. By combining fractional distance weights and a random perturbation mechanism, it effectively captures multi-scale non-local texture and random detail features in respiratory endoscopy images. Furthermore, the use of random enhancement and nonlinear fusion strategies significantly improves the sensitivity to local abnormal details and the ability to express features, achieving noise suppression and texture restoration in respiratory endoscopy images, thus improving diagnostic accuracy and clinical applicability.

[0010] Preferably, in step S4, the multi-scale feature extraction module FSDO is used to process the endoscopic images of the respiratory tract at three different scales, using the following formula: , in This refers to the multi-scale feature extraction module. Used to represent different scales of endoscopic images of the respiratory tract, with values ​​ranging from [1, 2, 3]. The feature maps of the lower-scale respiratory endoscopic images are upsampled and then fused with the feature maps of the upper-scale respiratory endoscopic images to obtain three different scales of respiratory endoscopic image feature maps. The specific formula is as follows: , in This indicates an upsampling operation.

[0011] Preferably, in step S4, this method uses fractional-order stochastic differential operators to extract effective features for different scales, and effectively integrates information between different scales through upsampling and cross-scale feature fusion, significantly improving the detail recognition and noise suppression capabilities of lesion areas in the image, providing clearer and more stable image support for accurate clinical diagnosis.

[0012] Preferably, in S5, respiratory endoscopic images often contain a large number of features such as blurred tissue boundaries, weak blood vessel contours, and sparse textures in inflammatory areas. These detailed areas appear as low-energy spectra or indistinct transition frequencies in the original image. At the same time, the image may introduce pseudo-high-frequency information due to factors such as lens fog, reflection, and tissue dynamics.

[0013] Preferably, in step S5, the process of constructing the frequency-enhanced fuzzy perception module AFFRHO includes: S51. For the input feature map Construct the kernel function for the two-dimensional fractional Fourier transform. and The specific formula is as follows: , , in and These are the fractional Fourier transform rotation parameters in the x and y directions, respectively, used to adjust the angle of frequency domain analysis. Represents an exponential function. This refers to the cot() function. This refers to the csc() function. Represents the complex frequency domain, with values ​​ranging from 1 to 10. , Then, a two-dimensional fractional Fourier transform is performed, with the specific formula as follows: , in The spectral feature map is obtained after the two-dimensional fractional Fourier transform; S52. A random harmonic perturbation mechanism is introduced to enhance the spectral characteristics and simulate the detailed changes in blurred areas of the image. The specific formula is as follows: , in The random harmonic disturbance factor follows a zero-mean Gaussian distribution. The local spectral frequency estimation is expressed by the following formula: , in Indicates the imaginary part. Indicates the real part; S53. A frequency enhancement weighting strategy is adopted, with weaker enhancement in the low-frequency region and stronger enhancement in the high-frequency region. The enhancement weight is adaptively adjusted according to the magnitude of the spectral amplitude, as shown in the following formula: , in and To adjust the parameters for frequency enhancement, The value range is [0.1, 1], which is used to control the enhancement amplitude. The value range is [0.5, 1.5], which is used to control the degree of nonlinear enhancement. The amplitude is in the frequency domain, representing the frequency response intensity. S54. The spectral features after random harmonic perturbation and frequency enhancement weighting are fused by element-wise multiplication to obtain the enhanced spectral feature map. The specific formula is as follows: , Then, the enhanced spectral feature map is transformed back to the spatial domain using the inverse fractional Fourier transform to generate the enhanced feature map. The specific formula is as follows: , in and This is the kernel function for the inverse fractional Fourier transform. and The rotation angle parameter is the inverse fractional Fourier transform.

[0014] Preferably, in step S5, S51 uses fractional-order transform to extract frequency domain features, which has mid-frequency enhancement capabilities and is more suitable for capturing irregular details and excessive frequency features. Steps S52 and S53 use phase perturbation and weight control to enhance weak structural regions in the spectrum without erasing the original low-frequency structure. At the same time, S53 adopts an adaptive spectrum enhancement strategy, which makes the frequency enhancement fuzzy perception module AFFRHO more effective attenuation in high-frequency energy regions, while preserving and enhancing mid- and low-frequency regions. It has a segmented frequency enhancement control capability. The frequency enhancement fuzzy perception module AFFRHO as a whole uses two-dimensional fractional-order Fourier transform to flexibly capture spectrum information and introduces random harmonic perturbation mechanism and adaptive spectrum enhancement strategy, which effectively restores high-frequency detail information and lesion structural features in the image, thereby significantly improving the image clarity, texture recognition and diagnostic accuracy of lesion regions, providing clinicians with more reliable diagnostic basis and improving the efficiency and accuracy of medical diagnosis.

[0015] Preferably, in step S6, the frequency enhancement blur perception module AFFRHO is used to process the feature maps of three different scales of respiratory endoscopic images to obtain three different scales of enhanced feature maps of respiratory endoscopic images. The specific process is as follows: , in This indicates the frequency-enhanced fuzzy perception module AFFRHO. The value range is [1,2,3], used to represent the enhancement feature maps of respiratory endoscopic images at different scales. The enhanced feature maps of respiratory endoscopic images are progressively upsampled using transposed convolution operations, and shallow and deep features are fused through skip connections to obtain more detailed enhanced feature maps of respiratory endoscopic images step by step. The specific process is as follows: , in For transposed convolution upsampling, low-resolution feature maps are upsampled to a high-resolution scale before fusion, resulting in the final enhanced image of the respiratory endoscopic image. For the highest-scale decoding output, the specific formula is: .

[0016] Preferably, in step S6, for clinical scenarios where respiratory endoscopic images have large scale differences, blurred lesion features, and are easily affected by noise, the AFFRHO module significantly enhances spectral details at each scale. Then, by employing a stepwise upsampling and cross-scale skip connection fusion mechanism, deep semantic features and shallow detail texture information are accurately integrated, effectively achieving image noise suppression and lesion area detail enhancement. This significantly improves the accuracy and visual clarity of image diagnosis, meeting the actual needs of clinical diagnosis of respiratory endoscopic images.

[0017] Compared with existing technologies, this invention has the following technical advantages: It utilizes fractional-order stochastic differential operators to achieve precise extraction and stochastic enhancement of multi-scale texture details in respiratory endoscopic images, overcoming the limitations of traditional convolution methods such as limited receptive field and insufficient detail extraction; through the frequency enhancement fuzzy perception module AFFRHO, it flexibly captures frequency domain features using two-dimensional fractional Fourier transform, and combines random harmonic perturbation and adaptive frequency weighting strategies to effectively recover lost high-frequency details in blurred areas, achieving refined spectral enhancement; a multi-scale decoding mechanism with stepwise upsampling fusion is designed, which can further enhance the transmission of detail information between different scales, effectively improving the clarity and diagnostic accuracy of lesion areas; it achieves efficient noise suppression and precise detail recovery in respiratory endoscopic images, providing clearer and more accurate image support for clinical diagnosis, and has significant clinical application value. Attached Figure Description

[0018] Figure 1 The present invention provides a flowchart of a method for denoising and enhancing respiratory endoscopic images.

[0019] Figure 2 This is a flowchart illustrating the image processing procedure for respiratory endoscopic images in a method for denoising and enhancing respiratory endoscopic images provided by the present invention.

[0020] Figure 3 The structure diagram of the multi-scale feature extraction module FSDO provided by this invention.

[0021] Figure 4 The structure diagram of the frequency enhancement fuzzy perception module AFFRHO provided by the present invention.

[0022] Figure 5 This invention provides a grayscale image of a respiratory endoscopic image and an image of the image after processing by the multi-scale feature extraction module FSDO.

[0023] Figure 6 This is a comparison image of the endoscopic respiratory tract image before and after enhancement, provided in one embodiment of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1 to 6 This invention provides a method for denoising and enhancing respiratory endoscopic images, which aims to combine a multi-scale feature extraction module and a frequency enhancement blur perception module to achieve denoising and enhancement of respiratory endoscopic images, including the following steps.

[0026] S1. Collect respiratory endoscopic images and create a respiratory endoscopic image dataset.

[0027] Furthermore, in S1, under the premise of ethical approval and informed consent from patients, a high-definition respiratory endoscopy system that meets clinical use standards is selected. The system is connected to a video capture card and a computer to collect respiratory endoscopy video data during actual surgery or examination. The collected respiratory endoscopy video data is analyzed frame by frame to extract static image frames. Representative and diagnostically significant respiratory endoscopy images are selected, totaling 5,000 images. Professional respiratory physicians are organized to manually annotate and classify the respiratory endoscopy images using the LabelMe annotation tool to form a respiratory endoscopy image dataset. Finally, the dataset is divided into a training set and a validation set in a 7:3 ratio for subsequent model training and performance evaluation.

[0028] S2. The endoscopic images of the respiratory tract are downsampled by 2x and 4x respectively to obtain endoscopic images of the respiratory tract at 2x and 4x downsampled scales.

[0029] Furthermore, in S2, the respiratory endoscopic image is Use it as the original image Then, the images were downsampled by 2x and 4x respectively to obtain 2x downsampled respiratory endoscopic images. and 4x downsampled respiratory endoscopic images The specific process is as follows: , .

[0030] S3. Construct a multi-scale feature extraction module FSDO, including: for any pixel in the input image, calculate its difference with other pixels in the neighborhood, perform weighted integration by combining fractional distance weights and spatial random perturbation factors, and obtain the feature extraction operator after normalization. Generate the output feature map by nonlinearly weighted fusion of feature extraction operators at multiple scales.

[0031] Furthermore, in step S3, the multi-scale feature extraction module FSDO is constructed, and the specific process includes: S31. For any pixel in the input image In its neighborhood Within the range, the difference between this pixel and other pixels is calculated, and a weighted integral is performed combining fractional distance weights and spatial random perturbation factors. After normalization, the feature extraction operator is obtained, with the specific formula as follows: , in This is a difference term, reflecting the local difference at position (u,v) near point (x,y). If the local pixel changes drastically, this difference term is larger and can capture edges and textures. This is a fractional-order decay term related to distance; the greater the distance, the smaller the contribution. The fractional-order scaling parameter controls the scale of action of the nonlocal operator, with a value range of [0.5, 1.5]. For large-scale features, The value range is [0.5, 0.8], which is suitable for capturing the overall structure of an image. For medium-scale features, The value range is [0.8, 1.2], which is suitable for capturing image texture. For small-scale features, The value range is [1.2, 1.5], which is suitable for capturing image details and edges. To use a small constant, to avoid singularities when the distance approaches 0. It is a spatial random function, which functions to make the multi-scale feature extraction operator... It is more sensitive to local detail changes in an image, especially when the image texture is highly random. It can better magnify detailed features. The normalization coefficient is calculated as follows: ; S32. By nonlinearly weighting and fusing feature extraction operators at various scales, an output feature map is generated. The specific formula is as follows: , in The weights are obtained through training. The `arctan()` function can effectively suppress abrupt changes in features, maintain feature stationarity, and facilitate subsequent processing. For the tanh function, A spatial random signal is used for the feature extraction operator. Random enhancement.

[0032] Furthermore, in step S31, for this embodiment, the feature map is a small-scale feature, and a 7×7 neighborhood window is used. The initial value is 1.0, a small constant. The initial value is Normalize the coefficients The calculation process is as follows: .

[0033] Furthermore, in S31, for this embodiment, the spatial random function Gaussian random field Initial setting mean The value is 0, and the standard deviation is 0. Setting it to 0.05 ensures enhanced detail features without compromising the overall image structure; spatial random function. The specific settings .

[0034] Furthermore, in S32, the spatial random signal For the feature extraction operator The random enhancement, in this embodiment, adopts a setting symmetrical to a range of 0, specifically as follows: This setup is computationally simple, easy to implement, and relatively stable in its sensitivity to image structure, allowing for the fusion of weights. The initial setting is to maintain equal weights for features at all scales, ensuring that features at each scale contribute equally in the early stages of training. This applies to feature maps at all three scales. The value is set to 1 / 3, and a hybrid loss of cross-entropy and structural similarity is used for update iteration. Meanwhile, to avoid... The divergence is then addressed by performing softmax normalization during training, specifically as follows: .

[0035] S4. Use FSDO to process respiratory endoscopic images. After processing the lower-scale respiratory endoscopic images, upsample them and fuse them with the features of the upper-scale respiratory endoscopic images to obtain a respiratory endoscopic image feature map.

[0036] Furthermore, in step S4, the multi-scale feature extraction module FSDO is used to process the endoscopic images of the respiratory tract at three different scales, using the following formula: , in This refers to the multi-scale feature extraction module. Used to represent different scales of endoscopic images of the respiratory tract, with values ​​ranging from [1, 2, 3]. The feature maps of the lower-scale respiratory endoscopic images are upsampled and then fused with the feature maps of the upper-scale respiratory endoscopic images to obtain three different scales of respiratory endoscopic image feature maps. The specific formula is as follows: , in This indicates an upsampling operation.

[0037] S5. Construct a frequency-enhanced fuzzy perception module AFFRHO, including: performing a two-dimensional fractional Fourier transform on the input feature map to obtain a spectral feature map, enhancing the spectral feature map using random harmonic enhancement and frequency enhancement weighting respectively, and then fusing them, converting them into an enhanced feature map through inverse fractional Fourier transform.

[0038] Furthermore, in step S5, the frequency-enhanced fuzzy perception module AFFRHO is constructed, and the specific process includes: S51. For the input feature map Construct the kernel function for the two-dimensional fractional Fourier transform. and The specific formula is as follows: , , in and These are the fractional Fourier transform rotation parameters in the x and y directions, respectively, used to adjust the angle of frequency domain analysis. In this embodiment... and Set to respectively Set as It is beneficial for capturing mid-frequency and transitional frequency information of respiratory endoscopic image feature maps. Represents an exponential function. This refers to the cot() function. This refers to the csc() function. Represents the complex frequency domain, with values ​​ranging from 1 to 10. , Then, a two-dimensional fractional Fourier transform is performed, with the specific formula as follows: , in The spectral feature map is obtained after the two-dimensional fractional Fourier transform; S52. A random harmonic perturbation mechanism is introduced to enhance the spectral characteristics and simulate the detailed changes in blurred areas of the image. The specific formula is as follows: , in The random harmonic disturbance factor follows a zero-mean Gaussian distribution. In this example, , The local spectral frequency estimation is expressed by the following formula: , in Indicates the imaginary part. Indicates the real part; S53. A frequency enhancement weighting strategy is adopted, with weaker enhancement in the low-frequency region and stronger enhancement in the high-frequency region. The enhancement weight is adaptively adjusted according to the magnitude of the spectral amplitude, as shown in the following formula: , in and To adjust the parameters for frequency enhancement, The value range is [0.1, 1], which is used to control the enhancement magnitude. In this embodiment, the initial value is set to 0.5, and gradient optimization is performed using the L1 loss function. The value range is [0.5, 1.5], which controls the degree of nonlinearity of the enhancement. In this embodiment, the initial value is set to 1.0, and gradient optimization is performed using the L1 loss function. The amplitude is in the frequency domain, representing the frequency response intensity. S54. The spectral features after random harmonic perturbation and frequency enhancement weighting are fused by element-wise multiplication to obtain the enhanced spectral feature map. The specific formula is as follows: , Then, the enhanced spectral feature map is transformed back to the spatial domain using the inverse fractional Fourier transform to generate the enhanced feature map. The specific formula is as follows: , in and This is the kernel function for the inverse fractional Fourier transform. and The rotation angle parameter of the inverse fractional Fourier transform is set to [value] in this embodiment. .

[0039] S6. The AFFRHO is used to process the feature map of the respiratory tract endoscopy image. Then, the transposed convolution operation is used to perform stepwise upsampling, and the feature information of the shallow and deep layers of the skip connection fusion network is combined to obtain the enhanced image of the respiratory tract endoscopy image.

[0040] Furthermore, in step S6, the frequency enhancement blur perception module AFFRHO is used to process the feature maps of the three different scales of the respiratory endoscopic images to obtain three different scales of enhanced feature maps of the respiratory endoscopic images. The specific process is as follows: , in This indicates the frequency-enhanced fuzzy perception module AFFRHO. The value range is [1,2,3], used to represent the enhancement feature maps of respiratory endoscopic images at different scales. The enhanced feature maps of respiratory endoscopic images are progressively upsampled using transposed convolution operations, and shallow and deep features are fused through skip connections to obtain more detailed enhanced feature maps of respiratory endoscopic images step by step. The specific process is as follows: , in For transposed convolution upsampling, low-resolution feature maps are upsampled to a high-resolution scale before fusion, resulting in the final enhanced image of the respiratory endoscopic image. For the highest-scale decoding output, the specific formula is: .

[0041] Furthermore, this method uses the PyCharm application and Python language to write code, and uses the PyTorch framework. It is trained by inputting 640×640×3 respiratory endoscopic images, and the model is trained for 300 epochs, of which 50 epochs are frozen.

[0042] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for enhancing the clarity of respiratory endoscopic images, characterized in that, Includes the following steps: S1. Collect respiratory endoscopic images and create a respiratory endoscopic image dataset; S2. The endoscopic images of the respiratory tract are downsampled by 2x and 4x respectively to obtain endoscopic images of the respiratory tract at 2x and 4x downsampled scales. S3. Construct a multi-scale feature extraction module FSDO, including: for any pixel in the input image, calculate the difference between the pixel and other pixels in the neighborhood, combine fractional distance weights and spatial random perturbation factors for weighted integration, perform normalization processing to obtain feature extraction operators, and generate output feature maps by nonlinearly weighted fusion of feature extraction operators at multiple scales. S4. Use FSDO to process respiratory endoscopic images. After processing the lower-scale respiratory endoscopic images, upsample them and fuse them with the features of the upper-scale respiratory endoscopic images to obtain a respiratory endoscopic image feature map. S5. Constructing the frequency-enhanced fuzzy perception module AFFRHO, including: performing a two-dimensional fractional Fourier transform on the input feature map to obtain a spectral feature map; enhancing the spectral feature map using random harmonic enhancement and frequency enhancement weighting methods respectively, then fusing them; and converting them into an enhanced feature map through an inverse fractional Fourier transform; the process of performing a two-dimensional fractional Fourier transform on the input feature map to obtain the spectral feature map is as follows: for the input feature map of the frequency-enhanced fuzzy perception module... First, construct the kernel function of the two-dimensional fractional Fourier transform. and ,in and Let x and y be the fractional Fourier transform rotation parameters, respectively, and then a two-dimensional fractional Fourier transform is performed. The specific formula is as follows: ， Where (u,v) represents the neighborhood centered at the current pixel (x,y). pixel coordinates within, The spectral feature map is obtained after a two-dimensional fractional Fourier transform. The process of enhancing the spectral feature map using the random harmonics is as follows: a random harmonic perturbation mechanism is introduced to enhance the spectral features, simulating the detail changes in blurred areas of the image. The specific formula is: ， in For random harmonic disturbance factor, For local spectral frequency estimation, Represents the complex frequency domain, with values ​​ranging from 1 to 10. The process of enhancing the spectral feature map using the aforementioned frequency enhancement weighting method is as follows: A frequency enhancement weighting strategy is employed, with weaker enhancement applied to the low-frequency region and stronger enhancement applied to the high-frequency region. The enhancement weights are adaptively adjusted based on the calculated spectral amplitude, using the following formula: ， in and To adjust the parameters for frequency enhancement, The amplitude is in the frequency domain. For the random harmonic enhancement and frequency enhancement weighting methods, the spectral feature maps are enhanced separately and then fused. The specific process of converting the enhanced feature map using the inverse fractional Fourier transform is as follows: The spectral features after random harmonic perturbation and frequency enhancement weighting are fused by element-wise multiplication to obtain the enhanced spectral feature map. Then, the enhanced spectral feature map is converted back to the spatial domain using the inverse fractional Fourier transform to generate the enhanced feature map. The specific formula is: ， in This represents the enhanced spectral feature map obtained by fusing the random harmonic perturbation enhancement and frequency enhancement weighted processing. The specific formula is as follows: , and This is the kernel function for the inverse fractional Fourier transform. and The rotation angle parameter is the inverse fractional Fourier transform. S6. The AFFRHO is used to process the feature map of the respiratory tract endoscopy image. Then, the transposed convolution operation is used to perform stepwise upsampling, and the feature information of the shallow and deep layers of the skip connection fusion network is combined to obtain the enhanced image of the respiratory tract endoscopy image.

2. The method for enhancing the clarity of respiratory endoscopic images according to claim 1, characterized in that, In step S3, the construction of the multi-scale feature extraction module FSDO includes the following specific steps: S31. For any pixel in the input image In its neighborhood Within the function, the difference between the pixel and other pixels is calculated, and a weighted integral is performed by combining fractional distance weights and spatial random perturbation factors. After normalization, the feature extraction operator is obtained. The specific formula is as follows: ， in This represents the input image pixel value located at pixel coordinates (u,v), where (u,v) is the neighborhood centered at the current pixel (x,y). pixel coordinates within, For the difference term, This is the fractional attenuation term for the distance. For fractional scale parameters, It is a small constant. It is a spatial random function. The normalization coefficient is calculated as follows: ； S32. By nonlinearly weighting and fusing feature extraction operators at various scales, an output feature map is generated. The specific formula is as follows: ， in To integrate weights, For the arctan() function, For the tanh function, It is a spatially random signal.

3. The method for enhancing the clarity of respiratory endoscopic images according to claim 2, characterized in that, In step S4, the multi-scale feature extraction module FSDO is used to process the respiratory endoscopic images at three different scales. The feature map of the lower scale respiratory endoscopic image is upsampled and then fused with the feature map of the upper scale respiratory endoscopic image to finally obtain the feature maps of the three different scales respiratory endoscopic images.

4. The method for enhancing the clarity of respiratory endoscopic images according to claim 3, characterized in that, In step S6, the frequency-enhanced fuzzy perception module AFFRHO is used to process the feature maps of three different scales of respiratory endoscopic images to obtain three different scales of enhanced respiratory endoscopic image feature maps. Then, the transposed convolution operation is used to upsample the enhanced respiratory endoscopic image feature maps step by step, and shallow and deep features are fused through skip connections to obtain the final enhanced respiratory endoscopic image. This is the highest-scale decoding output.

Images