Frequency domain semantic segmentation method based on heterogeneous distillation

Abstract

The invention relates to the technical field of semantic segmentation, in particular to a frequency domain semantic segmentation method based on heterogeneous distillation, and the method comprises the following steps: S1, constructing a local teacher model and a global student model, and aligning a feature map of the local teacher model with a feature map of the global student model; s2, based on the aligned feature map of the local teacher model and the feature map of the global student model, capturing high-frequency features of the local teacher model and high-frequency features of the global student model; s3, calculating the characteristic distillation loss of the frequency domain guide boundary; s4, calculating the response distillation loss of the gating equilibrium entropy; and S5, comprehensively evaluating the semantic segmentation capability of the global student model based on the characteristic distillation loss of the frequency domain guide boundary and the response distillation loss of the gating balance entropy. The method solves the problems that an existing semantic segmentation technology is low in operation speed and poor in segmentation integrity, and is suitable for semantic segmentation.

Description

Technical Field

[0001] The present invention relates to the field of semantic segmentation technology, and in particular to a frequency domain semantic segmentation method based on heterogeneous distillation. Background Art

[0002] Semantic segmentation is one of the core tasks in computer vision, which aims to assign precise semantic labels to each pixel in an image. As an important supporting technology in fields such as autonomous driving, environmental monitoring, and urban planning, semantic segmentation plays a vital role in traffic and remote sensing scenarios, especially in processing large-scale, high-dimensional image data and accurately extracting semantic information. However, in practical applications, the existing semantic segmentation technology has the following problems due to its own limitations: First, the existing semantic segmentation technology has a large number of model parameters in the segmentation process, resulting in high computational complexity and thus slow running speed. Second, the existing semantic segmentation technology is prone to information loss during the segmentation process, resulting in poor segmentation integrity. Based on this, it is necessary to invent a frequency domain semantic segmentation method based on heterogeneous distillation to solve the problems of slow running speed and poor segmentation integrity of the existing semantic segmentation technology. Summary of the Invention

[0003] In order to solve the problems of slow running speed and poor segmentation integrity of existing semantic segmentation technologies, the present invention provides a frequency domain semantic segmentation method based on heterogeneous distillation.

[0004] The present invention is achieved by adopting the following technical solutions:

[0005] A frequency domain semantic segmentation method based on heterogeneous distillation, the method comprising the following steps:

[0006] Step S1: Construct a local teacher model and a global student model, and align the feature map of the local teacher model with the feature map of the global student model;

[0007] Step S2: Based on the aligned feature maps of the local teacher model and the global student model, the high-frequency features of the local teacher model and the high-frequency features of the global student model are captured;

[0008] Step S3: Based on the high-frequency features of the local teacher model and the high-frequency features of the global student model, the boundary features of each are fused and the feature distillation loss of the frequency-domain guided boundary is calculated;

[0009] Step S4: Calculate the response distillation loss of the gated balanced entropy based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model;

[0010] Step S5: Based on the feature distillation loss of frequency-domain guided margin and the response distillation loss of gated balanced entropy, the semantic segmentation ability of the global student model is comprehensively evaluated.

[0011] Furthermore, the step S1 is specifically as follows:

[0012] Step S1.1: Construct a local teacher model based on a convolutional neural network and a global student model based on a Transformer network;

[0013] Step S1.2: Calculate the optical flow field between the feature map of the local teacher model and the feature map of the global student model; the specific formula is as follows:

[0014] ;

[0015] Where: represents the optical flow field; represents computational optical flow; represents the feature map of the local teacher model, ; represents the feature map of the global student model, ; Represents the displacement information of the optical flow field; and Represent the positions in the optical flow field Horizontal and vertical displacements;

[0016] Step S1.3: Apply the displacement information of the optical flow field to each pixel of the feature map of the global student model, and perform horizontal and vertical movement operations through horizontal and vertical depth strip convolutions, thereby aligning the feature map of the local teacher model and the feature map of the global student model.

[0017] Furthermore, the step S2 is specifically as follows:

[0018] Step S2.1: Using the matrix Detect the aligned feature maps of the local teacher model and the global student model to obtain the local teacher model matrix and the global student model matrix. Then, use the two-dimensional Fourier transform to map the local teacher model matrix and the global student model matrix to the frequency domain to obtain the frequency matrix of the local teacher model and the frequency matrix of the global student model. The specific formula is as follows:

[0019] ;

[0020] ;

[0021] Where: The frequency matrix representing the local teacher model; The frequency matrix representing the global student model; represents the two-dimensional Fourier transform; represents the local teacher model matrix; represents the global student model matrix;

[0022] Step S2.2: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use a circular mask to extract the features of the central area of the spectrum, thereby obtaining the low-frequency features of the local teacher model and the low-frequency features of the global student model; the specific formula is as follows:

[0023] ;

[0024] ;

[0025] Where: Represents the low-frequency features of the local teacher model; Represents the low-frequency features of the global student model; represents a circular mask;

[0026] Step S2.3: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use the complement of the circular mask to extract the features of the peripheral area of the spectrum, thereby capturing the high-frequency features of the local teacher model and the high-frequency features of the global student model; the specific formula is as follows:

[0027] ;

[0028] ;

[0029] Where: Represents the high-frequency features of the local teacher model; Represents the high-frequency features of the global student model; represents the complement of the circular mask, .

[0030] Furthermore, the step S3 is specifically as follows:

[0031] Step S3.1: Using the class probability matrix Detecting the boundary points of the local teacher model and the global student model, thereby obtaining the boundary matrix of the local teacher model and the boundary matrix of the global student model;

[0032] Step S3.2: The high-frequency features of the local teacher model and the high-frequency features of the global student model are mapped using a smooth activation function to obtain the high-frequency weight matrix of the local teacher model and the high-frequency weight matrix of the global student model. The specific formula is as follows:

[0033] ;

[0034] ;

[0035] Where: Represents the high-frequency weight matrix of the local teacher model; represents the high-frequency weight matrix of the global student model; represents the adjustment coefficient;

[0036] Then, on the one hand, the high-frequency weight matrix of the local teacher model is fused with the boundary matrix of the local teacher model to obtain the frequency-domain guided boundary features of the local teacher model. On the other hand, the high-frequency weight matrix of the global student model is fused with the boundary matrix of the global student model to obtain the frequency-domain guided boundary features of the global student model. The specific formula is as follows:

[0037] ;

[0038] ;

[0039] Where: Represents the frequency domain guided boundary features of the local teacher model; Represents the frequency domain guided boundary characteristics of the global student model; represents the boundary matrix of the local teacher model; represents the boundary matrix of the global student model;

[0040] Then, based on the frequency domain guided boundary features of the local teacher model and the frequency domain guided boundary features of the global student model, an adaptive threshold is used to filter out low response noise; the specific formula is as follows:

[0041] ;

[0042] ;

[0043] Where: represents the adaptive threshold, ; and represent the mean and standard deviation of the frequency domain weighted graph respectively; represents the adjustment coefficient;

[0044] Step S3.3: Calculate the feature distillation loss based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model; the specific formula is as follows:

[0045] ;

[0046] Where: represents the feature distillation loss; Represents a set of boundary pixels; Represents the pixel in the frequency boundary response map The weighted value of .

[0047] Furthermore, the step S4 is specifically as follows:

[0048] Step S4.1: Calculate the category probability of the local teacher model and the category probability of the global student model using the Softmax function for the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model. Then, calculate the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model based on the category probabilities of the local teacher model and the category probabilities of the global student model. The specific formula is as follows:

[0049] ;

[0050] ;

[0051] Where: represents the pixel-wise entropy of the feature map of the local teacher model; Pixel-wise entropy of the feature map representing the global student model; represents the category probability of the local teacher model, ; represents the class probability of the global student model, ; , used to prevent numerical overflow; represents the channel index, ; Represents the spatial position index, ;

[0052] Step S4.2: Calculate the entropy difference based on the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model; the specific formula is as follows:

[0053] ;

[0054] Where: represents the entropy difference;

[0055] Then, based on the entropy difference, the balanced gating weight is calculated; the specific formula is as follows:

[0056] ;

[0057] Where: represents the balanced gating weight; is a hyperparameter used to adjust the effect of entropy difference on fusion. ;

[0058] Step S4.3: Perform weighted fusion of the feature maps of the local teacher model and the global student model by balancing the gating weights and calculate the response distillation loss; the specific formula is as follows:

[0059] ;

[0060] Where: Represents the response distillation loss.

[0061] Furthermore, the step S5 is specifically as follows:

[0062] Step S5.1: Calculate the total distillation loss based on the characteristic distillation loss and the response distillation loss. The specific formula is as follows:

[0063] ;

[0064] Where: represents the total distillation loss; and Both represent weight hyperparameters;

[0065] Step S5.2: Train and validate a global student model based on the feature distillation loss and the response distillation loss; then, use the global student model to perform semantic segmentation on the image and assign an accurate semantic label to each pixel in the image;

[0066] Step S5.3: Use evaluation indicators to measure the segmentation accuracy and overall performance of the global student model in each category, thereby comprehensively evaluating the semantic segmentation ability of the global student model.

[0067] Furthermore, the evaluation indicators include but are not limited to: mIoU indicator, IoU indicator, ACC indicator, Parameter indicator and FLOPs indicator.

[0068] Compared with existing semantic segmentation technologies, the frequency-domain semantic segmentation method based on heterogeneous distillation described in this invention has the following advantages: First, it can effectively reduce the number of model parameters during the segmentation process, thereby effectively reducing computational complexity and significantly accelerating the operation speed. Second, it can effectively avoid information loss during the segmentation process, thereby effectively enhancing segmentation integrity.

[0069] The present invention effectively solves the problems of slow running speed and poor segmentation integrity of existing semantic segmentation technologies and is suitable for semantic segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic flow diagram of the present invention.

[0071] Figure 2 It is a principle block diagram of the present invention.

[0072] Figure 3 It is a schematic diagram of step S1.3 in the present invention. DETAILED DESCRIPTION

[0073] A frequency domain semantic segmentation method based on heterogeneous distillation, the method comprising the following steps:

[0074] Step S1: Construct a local teacher model and a global student model, and align the feature map of the local teacher model with the feature map of the global student model;

[0075] Step S2: Based on the aligned feature maps of the local teacher model and the global student model, the high-frequency features of the local teacher model and the high-frequency features of the global student model are captured;

[0076] Step S3: Based on the high-frequency features of the local teacher model and the high-frequency features of the global student model, the boundary features of each are fused and the feature distillation loss of the frequency-domain guided boundary is calculated;

[0077] Step S4: Calculate the response distillation loss of the gated balanced entropy based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model;

[0078] Step S5: Based on the feature distillation loss of frequency-domain guided margin and the response distillation loss of gated balanced entropy, the semantic segmentation ability of the global student model is comprehensively evaluated.

[0079] The step S1 is specifically as follows:

[0080] Step S1.1: Construct a local teacher model based on a convolutional neural network and a global student model based on a Transformer network;

[0081] Step S1.2: Calculate the optical flow field between the feature map of the local teacher model and the feature map of the global student model; the specific formula is as follows:

[0082] ;

[0083] Where: represents the optical flow field; represents computational optical flow; represents the feature map of the local teacher model, ; represents the feature map of the global student model, ; Represents the displacement information of the optical flow field; and Represent the positions in the optical flow field Horizontal and vertical displacements;

[0084] Step S1.3: Apply the displacement information of the optical flow field to each pixel of the feature map of the global student model, and perform horizontal and vertical movement operations through horizontal and vertical depth strip convolutions, thereby aligning the feature map of the local teacher model and the feature map of the global student model.

[0085] The step S2 is specifically as follows:

[0086] Step S2.1: Using the matrix Detect the aligned feature maps of the local teacher model and the global student model to obtain the local teacher model matrix and the global student model matrix. Then, use the two-dimensional Fourier transform to map the local teacher model matrix and the global student model matrix to the frequency domain to obtain the frequency matrix of the local teacher model and the frequency matrix of the global student model. The specific formula is as follows:

[0087] ;

[0088] ;

[0089] Where: The frequency matrix representing the local teacher model; The frequency matrix representing the global student model; represents the two-dimensional Fourier transform; represents the local teacher model matrix; represents the global student model matrix;

[0090] Step S2.2: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use a circular mask to extract the features of the central area of the spectrum, thereby obtaining the low-frequency features of the local teacher model and the low-frequency features of the global student model; the specific formula is as follows:

[0091] ;

[0092] ;

[0093] Where: Represents the low-frequency features of the local teacher model; Represents the low-frequency features of the global student model; represents a circular mask;

[0094] Step S2.3: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use the complement of the circular mask to extract the features of the peripheral area of the spectrum, thereby capturing the high-frequency features of the local teacher model and the high-frequency features of the global student model; the specific formula is as follows:

[0095] ;

[0096] ;

[0097] Where: Represents the high-frequency features of the local teacher model; Represents the high-frequency features of the global student model; represents the complement of the circular mask, .

[0098] The step S3 is specifically as follows:

[0099] Step S3.1: Using the class probability matrix Detecting the boundary points of the local teacher model and the global student model, thereby obtaining the boundary matrix of the local teacher model and the boundary matrix of the global student model;

[0100] Step S3.2: The high-frequency features of the local teacher model and the high-frequency features of the global student model are mapped using a smooth activation function to obtain the high-frequency weight matrix of the local teacher model and the high-frequency weight matrix of the global student model. The specific formula is as follows:

[0101] ;

[0102] ;

[0103] Where: Represents the high-frequency weight matrix of the local teacher model; represents the high-frequency weight matrix of the global student model; represents the adjustment coefficient;

[0104] Then, on the one hand, the high-frequency weight matrix of the local teacher model is fused with the boundary matrix of the local teacher model to obtain the frequency-domain guided boundary features of the local teacher model. On the other hand, the high-frequency weight matrix of the global student model is fused with the boundary matrix of the global student model to obtain the frequency-domain guided boundary features of the global student model. The specific formula is as follows:

[0105] ;

[0106] ;

[0107] Where: Represents the frequency domain guided boundary features of the local teacher model; Represents the frequency domain guided boundary characteristics of the global student model; represents the boundary matrix of the local teacher model; represents the boundary matrix of the global student model;

[0108] Then, based on the frequency domain guided boundary features of the local teacher model and the frequency domain guided boundary features of the global student model, an adaptive threshold is used to filter out low response noise; the specific formula is as follows:

[0109] ;

[0110] ;

[0111] Where: represents the adaptive threshold, ; and represent the mean and standard deviation of the frequency domain weighted graph respectively; represents the adjustment coefficient;

[0112] Step S3.3: Calculate the feature distillation loss based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model; the specific formula is as follows:

[0113] ;

[0114] Where: represents the feature distillation loss; Represents a set of boundary pixels; Represents the pixel in the frequency boundary response map The weighted value of .

[0115] The step S4 is specifically as follows:

[0116] Step S4.1: Calculate the category probability of the local teacher model and the category probability of the global student model using the Softmax function for the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model. Then, calculate the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model based on the category probabilities of the local teacher model and the category probabilities of the global student model. The specific formula is as follows:

[0117] ;

[0118] ;

[0119] Where: represents the pixel-wise entropy of the feature map of the local teacher model; Pixel-wise entropy of the feature map representing the global student model; represents the category probability of the local teacher model, ; represents the class probability of the global student model, ; , used to prevent numerical overflow; represents the channel index, ; Represents the spatial position index, ;

[0120] Step S4.2: Calculate the entropy difference based on the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model; the specific formula is as follows:

[0121] ;

[0122] Where: represents the entropy difference;

[0123] Then, based on the entropy difference, the balanced gating weight is calculated; the specific formula is as follows:

[0124] ;

[0125] Where: represents the balanced gating weight; is a hyperparameter used to adjust the effect of entropy difference on fusion. ;

[0126] Step S4.3: Perform weighted fusion of the feature maps of the local teacher model and the global student model by balancing the gating weights and calculate the response distillation loss; the specific formula is as follows:

[0127] ;

[0128] Where: Represents the response distillation loss.

[0129] The step S5 is specifically as follows:

[0130] Step S5.1: Calculate the total distillation loss based on the characteristic distillation loss and the response distillation loss. The specific formula is as follows:

[0131] ;

[0132] Where: represents the total distillation loss; and Both represent weight hyperparameters;

[0133] Step S5.2: Train and validate a global student model based on the feature distillation loss and the response distillation loss; then, use the global student model to perform semantic segmentation on the image and assign an accurate semantic label to each pixel in the image;

[0134] Step S5.3: Use evaluation indicators to measure the segmentation accuracy and overall performance of the global student model in each category, thereby comprehensively evaluating the semantic segmentation ability of the global student model.

[0135] The evaluation indicators include but are not limited to: mIoU indicator, IoU indicator, ACC indicator, Parameter indicator and FLOPs indicator.

[0136] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these are merely illustrative and that the scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, and such changes and modifications are intended to fall within the scope of the present invention.

Claims

1. A frequency domain semantic segmentation method based on heterogeneous distillation, characterized by: The method comprises the following steps: Step S1: Construct a local teacher model and a global student model, and align the feature map of the local teacher model with the feature map of the global student model; Step S2: Based on the aligned feature maps of the local teacher model and the global student model, the high-frequency features of the local teacher model and the high-frequency features of the global student model are captured; Step S3: Based on the high-frequency features of the local teacher model and the high-frequency features of the global student model, the boundary features of each are fused and the feature distillation loss of the frequency-domain guided boundary is calculated; Step S4: Calculate the response distillation loss of the gated balanced entropy based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model; Step S5: Based on the feature distillation loss of frequency-domain guided margin and the response distillation loss of gated balanced entropy, the semantic segmentation ability of the global student model is comprehensively evaluated.

2. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 1, characterized in that: The step S1 is specifically as follows: Step S1.1: Construct a local teacher model based on a convolutional neural network and a global student model based on a Transformer network; Step S1.2: Calculate the optical flow field between the feature map of the local teacher model and the feature map of the global student model; the specific formula is as follows: ； Where: represents the optical flow field; represents computational optical flow; represents the feature map of the local teacher model, ; represents the feature map of the global student model, ; Represents the displacement information of the optical flow field; and Represent the positions in the optical flow field Horizontal and vertical displacements; Step S1.3: Apply the displacement information of the optical flow field to each pixel of the feature map of the global student model, and perform horizontal and vertical movement operations through horizontal and vertical depth strip convolutions, thereby aligning the feature map of the local teacher model and the feature map of the global student model.

3. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 2, characterized in that: The step S2 is specifically as follows: Step S2.1: Using the matrix Detect the aligned feature maps of the local teacher model and the global student model to obtain the local teacher model matrix and the global student model matrix. Then, use the two-dimensional Fourier transform to map the local teacher model matrix and the global student model matrix to the frequency domain to obtain the frequency matrix of the local teacher model and the frequency matrix of the global student model. The specific formula is as follows: ； ； Where: The frequency matrix representing the local teacher model; The frequency matrix representing the global student model; represents the two-dimensional Fourier transform; represents the local teacher model matrix; represents the global student model matrix; Step S2.2: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use a circular mask to extract the features of the central area of the spectrum, thereby obtaining the low-frequency features of the local teacher model and the low-frequency features of the global student model; The specific formula is as follows: ； ； Where: Represents the low-frequency features of the local teacher model; Represents the low-frequency features of the global student model; represents a circular mask; Step S2.3: For the frequency matrix of the local teacher model and the frequency matrix of the global student model, use the complement of the circular mask to extract the features of the peripheral area of the spectrum, thereby capturing the high-frequency features of the local teacher model and the high-frequency features of the global student model; the specific formula is as follows: ； ； Where: Represents the high-frequency features of the local teacher model; Represents the high-frequency features of the global student model; represents the complement of the circular mask, .

4. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 3, characterized in that: The step S3 is specifically as follows: Step S3.1: Using the class probability matrix Detecting the boundary points of the local teacher model and the global student model, thereby obtaining the boundary matrix of the local teacher model and the boundary matrix of the global student model; Step S3.2: The high-frequency features of the local teacher model and the high-frequency features of the global student model are mapped using a smooth activation function to obtain the high-frequency weight matrix of the local teacher model and the high-frequency weight matrix of the global student model. The specific formula is as follows: ； ； Where: Represents the high-frequency weight matrix of the local teacher model; represents the high-frequency weight matrix of the global student model; represents the adjustment coefficient; Then, on the one hand, the high-frequency weight matrix of the local teacher model is fused with the boundary matrix of the local teacher model to obtain the frequency-domain guided boundary features of the local teacher model. On the other hand, the high-frequency weight matrix of the global student model is fused with the boundary matrix of the global student model to obtain the frequency-domain guided boundary features of the global student model. The specific formula is as follows: ； ； Where: Represents the frequency domain guided boundary features of the local teacher model; Represents the frequency domain guided boundary characteristics of the global student model; represents the boundary matrix of the local teacher model; represents the boundary matrix of the global student model; Then, based on the frequency domain guided boundary features of the local teacher model and the frequency domain guided boundary features of the global student model, an adaptive threshold is used to filter out low response noise; the specific formula is as follows: ； ； Where: represents the adaptive threshold, ; and represent the mean and standard deviation of the frequency domain weighted graph respectively; represents the adjustment coefficient; Step S3.3: Calculate the feature distillation loss based on the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model; the specific formula is as follows: ； Where: represents the feature distillation loss; Represents a set of boundary pixels; Represents the pixel in the frequency boundary response map The weighted value of .

5. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 4, characterized in that: The step S4 is specifically as follows: Step S4.1: Calculate the category probability of the local teacher model and the category probability of the global student model using the Softmax function for the frequency-domain guided boundary features of the local teacher model and the frequency-domain guided boundary features of the global student model. Then, calculate the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model based on the category probabilities of the local teacher model and the category probabilities of the global student model. The specific formula is as follows: ； ； Where: represents the pixel-wise entropy of the feature map of the local teacher model; Pixel-wise entropy of the feature map representing the global student model; represents the category probability of the local teacher model, ; represents the class probability of the global student model, ; , used to prevent numerical overflow; represents the channel index, ; Represents the spatial position index, ; Step S4.2: Calculate the entropy difference based on the pixel-by-pixel entropy of the feature map of the local teacher model and the pixel-by-pixel entropy of the feature map of the global student model; the specific formula is as follows: ； Where: represents the entropy difference; Then, based on the entropy difference, the balanced gating weight is calculated; the specific formula is as follows: ； Where: represents the balanced gating weight; is a hyperparameter used to adjust the effect of entropy difference on fusion. ; Step S4.3: Perform weighted fusion of the feature maps of the local teacher model and the global student model by balancing the gating weights and calculate the response distillation loss; the specific formula is as follows: ； Where: Represents the response distillation loss.

6. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 5, characterized in that: The step S5 is specifically as follows: Step S5.1: Calculate the total distillation loss based on the characteristic distillation loss and the response distillation loss. The specific formula is as follows: ； Where: represents the total distillation loss; and Both represent weight hyperparameters; Step S5.2: Train and validate a global student model based on the feature distillation loss and the response distillation loss; then, use the global student model to perform semantic segmentation on the image and assign an accurate semantic label to each pixel in the image; Step S5.3: Use evaluation indicators to measure the segmentation accuracy and overall performance of the global student model in each category, thereby comprehensively evaluating the semantic segmentation ability of the global student model.

7. The frequency domain semantic segmentation method based on heterogeneous distillation according to claim 6, characterized in that: The evaluation indicators include but are not limited to: mIoU indicator, IoU indicator, ACC indicator, Parameter indicator and FLOPs indicator.

Images