A method and system for unstructured environment saliency semantic segmentation

Abstract

The application provides a non-structured environment saliency semantic segmentation method and system, and belongs to the technical field of image semantic segmentation. In the original feature extraction stage, more ghost feature maps are generated through low-cost linear transformation by using existing feature maps, so as to improve the calculation efficiency of the network; affinity relationship channel attention and affinity relationship space attention are introduced to enhance the feature expression of different channels and extract key information at different positions in the space, so as to improve the performance without increasing the complexity of the network; the features output by the network are expanded by using differences, low-probability pixels are suppressed, high-probability pixels are enhanced, the concentration value of gray or haze is expressed by a probability value, similar dust or haze probability values are mapped to a more reasonable gray interval, the subtle difference of the concentration of dust or haze is enhanced, and the segmentation precision is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image semantic segmentation, and particularly relates to a saliency semantic segmentation method and system in an unstructured environment. BACKGROUND

[0002] Image semantic segmentation is an important research direction in the field of computer vision. Traditional semantic segmentation methods based on deep learning mainly rely on stacked convolutional layers for feature extraction. Although this feature extraction method is simple to build a model, the convolutional layers usually only take a simple stacking or splicing approach, without considering the weight of the features, making it difficult for the model to learn accurate image detail features. When dealing with shape changes or small targets, especially dust and haze in unstructured environments, important features may be lost, resulting in reduced segmentation accuracy.

[0003] Therefore, it is necessary to provide a saliency semantic segmentation method and system in an unstructured environment to solve the above problems. SUMMARY

[0004] The application provides a saliency semantic segmentation method and system in an unstructured environment. By introducing an affinity relationship attention mechanism in the semantic segmentation network, the generated segmentation map can effectively reflect the concentration difference of dust and haze in an unstructured environment, improving the segmentation accuracy and the generalization ability of the model, thereby solving at least one technical problem involved in the background art.

[0005] To solve the above technical problems, the application is implemented as follows:

[0006] A saliency semantic segmentation method in an unstructured environment includes the following steps:

[0007] Step S1, obtaining an unstructured environment image to be segmented, and generating an original feature map of the unstructured environment image using ghost convolution;

[0008] Step S2, learning the importance of different positions and different channels in the original feature map using affinity relationship spatial attention and affinity relationship channel attention, respectively, to generate a feature map considering affinity relationship attention;

[0009] Step S3, mapping the feature map considering affinity relationship attention to a class space to generate a class probability distribution of each pixel, and obtaining a segmentation result;

[0010] Step S4, sequentially performing Fourier transform, band-pass filtering, and inverse Fourier transform on the segmentation result to remove noise, and introducing a difference expansion mapping relationship function to expand the image gray difference in the segmentation result.

[0011] As a preferred improvement, the generation process of the original feature map specifically includes the following steps:

[0012] Step S11: performing convolution operation on the unstructured environment image to extract features and generate a feature map with a size of , where , , respectively represent the number of channels, height and width of the feature map ;

[0013] Step S12: performing ghost convolution on the feature map to generate a feature map ;

[0014] Step S13: performing one convolution operation on the feature map to generate a feature map ;

[0015] Step S14: splicing the feature map and the feature map along the channel dimension to obtain an original feature map with a size of ;

[0016] Step S15: performing batch normalization on the original feature map and adopting a hardswish function for activation.

[0017] As a preferred improvement, the processing procedure of affinity relationship space attention specifically includes the following steps:

[0018] Step S211: performing global maximum pooling and global average pooling on the original feature map respectively along the channel dimension to obtain two channel attention vectors with a size of , ;

[0019] Step S212: dividing the original feature map into feature units and flattening and reshaping them into a feature map ; where , represents the set of real numbers;

[0020] Step S213: converting the feature map through two different conversion functions and respectively to generate feature maps and , where the conversion functions and Each is composed of a convolutional layer with a convolution kernel of 1, a batch normalization layer, and a ReLU activation function;

[0021] Step S214, the feature map and the feature map are multiplied to obtain the affinity relationship matrix ;

[0022] Step S215, the affinity relationship matrix is reshaped into two relationship feature maps and respectively according to the row and column elements;

[0023] Step S216, the channel attention vectors , and the relationship feature maps and are spliced in the channel direction to obtain the feature map ;

[0024] Step S217, the feature map is extracted through two conversion functions and to generate the affinity relationship space attention weight matrix with a size of , wherein is composed of a convolutional layer with a convolution kernel of 1, a batch normalization layer, and a ReLU activation function, is composed of a convolutional layer with a convolution kernel of 1, a batch normalization layer, and a Sigmoid activation function.

[0025] As a preferred improvement, the processing process of the affinity relationship channel attention specifically includes the following steps:

[0026] Step S221, the original feature map is respectively subjected to global maximum pooling and global average pooling in the channel dimension, so as to obtain two channel attention vectors and ;

[0027] Step S222, the feature map is divided into feature units in the channel direction, and is tiled and reshaped into a feature map ;

[0028] Step S223, the feature map is converted into feature maps and respectively through two different conversion functions and ​​Transformation function and Each consists of a convolutional layer with a kernel of 1, batch normalization, and a ReLU activation function;

[0029] Step S224, transfer the feature map and Perform matrix multiplication to obtain the affinity matrix. ;

[0030] Step S225, generate the affinity matrix Reshape the shape by row and column elements respectively. Two relation feature diagrams and ;

[0031] Step S226, convert the channel attention vector , Relationship Feature Map and A new feature map is obtained by stitching along the channel direction. ;

[0032] Step S227, transfer the feature map Through conversion function and Extract global information and generate a size of Affinity space attention weight matrix .

[0033] As a preferred improvement, the feature map generation process considering affinity attention specifically includes the following steps:

[0034] Step S231, convert the original feature map Spatial attention weights related to affinity The feature map is obtained by multiplying element by element along the channel direction. ;

[0035] Step S232, then the feature map Attention weights for affinity channels Element-wise multiplication along spatial locations yields the feature map. ;

[0036] Step S233, finally, the original feature map With feature map The feature map considering affinity attention is obtained by adding elements together. .

[0037] As a preferred improvement, step S3 specifically includes the following process: adjusting the feature map using a 1×1 convolution. The number of channels is made equal to the number of classes, and then the output value of each pixel is compressed to [0,1] using a Sigmoid function, representing the probability of belonging to dust or haze.

[0038] As a preferred improvement, the Fourier transform is performed as follows:

[0039]

[0040] wherein, represents the complex-valued Fourier coefficient at frequency point , represents the vertical spatial frequency component index in the frequency domain, represents the horizontal spatial frequency component index in the frequency domain; represents the intensity of the pixel at spatial domain coordinate , represents the row index of the pixel in the spatial domain, represents the column index of the pixel in the spatial domain; represents the imaginary unit;

[0041] The bandpass filter is represented as:

[0042]

[0043] wherein, represents the response value of the bandpass filter at frequency point ; represents the Euclidean distance from the frequency point to the center frequency point; represents the center frequency; represents the band bandwidth;

[0044] The inverse Fourier transform is represented as:

[0045]

[0046] As a preferred improvement, the difference amplification mapping function is represented as:

[0047] ,

[0048] wherein, represents an intermediate parameter; represents the difference amplification mapping function; represents the amplification factor; represents the maximum value of the image gray pixel value, taking the value 255; represents the inverse Fourier transform operator; represents the Fourier transform operator; A rounding function.

[0049] A system for performing the above-mentioned unstructured environment saliency semantic segmentation method, comprising:

[0050] A semantic segmentation network, which is obtained by replacing the feature extraction module in a traditional Unet network with a ghost affinity relation attention module, the ghost affinity relation attention module comprising a ghost convolution layer, an affinity relation attention layer and an output layer connected in sequence, wherein:

[0051] The ghost convolution layer is used to obtain an unstructured environment image to be segmented, and generate an original feature map of the unstructured environment image by using ghost convolution;

[0052] The affinity relation attention layer is used to learn the importance of different positions and different channels in the original feature map by using affinity relation spatial attention and affinity relation channel attention respectively, and generate a feature map considering affinity relation attention;

[0053] The output layer is used to map the feature map considering affinity relation attention to a category space, generate a category probability distribution of each pixel, and obtain a segmentation result;

[0054] A difference amplification module is used to sequentially perform Fourier transform, band-pass filtering and inverse Fourier transform on the segmentation result to remove noise, and introduce a difference amplification mapping relationship function to amplify the image gray scale difference in the segmentation result.

[0055] As a preferred improvement, the semantic segmentation network comprises 6 layers of encoders and 5 layers of decoders, each layer of the encoder and the decoder uses the ghost affinity relation attention module as a basic module for feature extraction, the encoder layer extracts abstract features of the input image by downsampling, the decoder reconstructs the details of the image by upsampling, and outputs a saliency map, the saliency map uses a predicted probability value as a dust concentration value, and is used to reflect the concentration difference of dust; the encoder and the decoder are connected by a skip connection, the local features of the encoder are directly transmitted to the decoder, and the loss of spatial information in the downsampling is compensated; each decoder of the semantic segmentation network outputs a segmentation result, and the output results of all decoders are weighted and summed to obtain a final segmentation result map.

[0056] The beneficial effects of the present application are:

[0057] (1) In the original feature extraction stage, more ghost feature maps are generated by low-cost linear transformation using the existing feature maps, thereby improving the computational efficiency of the network;

[0058] (2) Introducing affinity channel attention and affinity spatial attention to enhance the expression of features of different channels and extract key information at different positions in the space, and improve the performance without increasing the complexity of the network;

[0059] (3) Expanding the difference of the features output by the network, suppressing the pixels with lower probability and enhancing the pixels with higher probability, expressing the concentration value of the gray or haze with the probability value, mapping the similar dust or haze probability value to a more reasonable gray interval, enhancing the subtle difference of the dust or haze concentration, and improving the segmentation precision. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0061] Figure 1 It is the framework diagram of the semantic segmentation network in embodiment 2.

[0062] Figure 2 It is the framework diagram of the ghost affinity attention module in embodiment 2.

[0063] Figure 3 It is the flowchart of the affinity spatial attention processing process in the present application.

[0064] Figure 4 It is the flowchart of the affinity channel attention processing process in the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0066] Embodiment 1

[0067] The present embodiment provides a non-structured environment saliency semantic segmentation method, comprising the following steps:

[0068] Step S1, obtaining a non-structured environment image to be segmented, and generating an original feature map of the non-structured environment image by using ghost convolution.

[0069] The unstructured environment image is acquired by using the prior art in the field, for example, collected from existing image data or photographed by a camera, a camera device, etc.

[0070] The generation process of the original feature map specifically includes the following steps:

[0071] Step S11: performing convolution operation on the unstructured environment image to extract features and generate a feature map with a size of , wherein , , respectively represent the channel number, height and width of the feature map ;

[0072] Step S12: performing ghost convolution on the feature map to generate a feature map ;

[0073] Step S13: performing one convolution operation on the feature map to generate a feature map ;

[0074] Step S14: splicing the feature map and the feature map along the channel dimension to obtain an original feature map with a size of ;

[0075] Step S15: performing batch normalization on the original feature map and adopting a hardswish function for activation.

[0076] In the original feature extraction stage, more ghost feature maps are generated by low-cost linear transformation using the existing feature map, thereby improving the calculation efficiency of the network.

[0077] Step S2: learning the importance of different positions and different channels in the original feature map by using affinity relationship spatial attention and affinity relationship channel attention respectively, to generate a feature map considering affinity relationship attention.

[0078] The processing process of the affinity relationship spatial attention specifically includes the following steps:

[0079] Step S211: performing global maximum pooling and global average pooling on the original feature map respectively along the channel dimension to obtain two channel attention vectors with a size of , ;

[0080] Step S212: performing ghost convolution on the original feature map Division Each feature unit is tiled and reshaped into a feature map. ;in, , Represents the set of real numbers;

[0081] Step S213, transfer the feature map They were processed through two different transformation functions. and Perform the transformation to generate feature maps. and , where the transformation function and Each consists of a convolutional layer with a kernel of 1, batch normalization, and a ReLU activation function;

[0082] Step S214, transfer the feature map and feature map Perform matrix multiplication to obtain the affinity matrix. ;

[0083] Step S215, generate the affinity matrix Reshape the shape by row and column elements respectively. Two relation feature diagrams and ;

[0084] Step S216, convert the channel attention vector , and relation feature graph and Feature maps are obtained by stitching along the channel direction. ;

[0085] Step S217, transfer the feature map Through two conversion functions respectively and Extract global information and generate a size of Affinity space attention weight matrix ,in, It consists of convolutional layers with a kernel of 1, batch normalization layers, and ReLU activation functions. It consists of convolutional layers with a kernel of 1, batch normalization layers, and a sigmoid activation function.

[0086] The common learning feature vector attention weight method is divided into local attention and global attention. For local attention, a feature unit is calculated by a common conversion function from other local feature units to form a weight, which ignores global information. For global attention, a feature unit is jointly learned from all feature units to form an attention weight, but this strategy is inefficient and consumes a large amount of computing resources when the number of feature units is large. The present application considers the affinity relationship between feature units, explores global feature information and performs knowledge mining, and fuses the features and the relationship between the features to jointly learn attention, which can better extract the features of dust and haze areas.

[0087] The processing process of the affinity relationship channel attention includes the following steps:

[0088] In step S221, the original feature map is subjected to global maximum pooling and global average pooling in the channel dimension, respectively, to obtain two feature maps . , ;

[0089] In step S222, the feature map is divided into feature units in the channel direction, and is tiled and reshaped into a feature map .

[0090] In step S223, the feature map is converted into feature maps and by two different conversion functions and , respectively. The conversion functions and each consist of a convolution layer with a convolution kernel of 1, batch normalization, and a ReLU activation function.

[0091] In step S224, the feature maps and are multiplied to obtain an affinity relationship matrix .

[0092] In step S225, the affinity relationship matrix is reshaped into two relationship feature maps and by row and column elements, respectively.

[0093] In step S226, the channel attention vectors , and the relationship feature maps and are spliced in the channel direction to obtain a new feature map​ ;

[0094] Step S227, the feature map is converted by the conversion function and global information is extracted to generate an affinity relationship space attention weight matrix with a size of .

[0095] The feature map generation process considering affinity relationship attention specifically includes the following steps:

[0096] Step S231, the original feature map is element-wise multiplied with the affinity relationship space attention weight in the channel direction to obtain a feature map ;

[0097] Step S232, the feature map is element-wise multiplied with the affinity relationship channel attention weight in the spatial position to obtain a feature map ;

[0098] Step S233, finally, the original feature map is element-wise added with the feature map to obtain a feature map considering affinity relationship attention .

[0099] Step S3, the feature map considering affinity relationship attention is mapped to the category space to generate a category probability distribution of each pixel to obtain a segmentation result.

[0100] Step S3 specifically includes the following process: using a 1x1 convolution to adjust the channel number of the feature map to be equal to the number of categories, and then using a Sigmoid function to compress the output value of each pixel to [0,1] to represent the probability of belonging to dust or haze.

[0101] Step S4, the segmentation result is sequentially subjected to Fourier transform, band-pass filtering and inverse Fourier transform for denoising, and a difference expansion mapping relationship function is introduced to expand the image gray difference in the segmentation result.

[0102] The process of Fourier transform is as follows:

[0103]

[0104] In the formula, represents a complex Fourier coefficient at a frequency point , represents a vertical spatial frequency component index in the frequency domain, ​represents a horizontal spatial frequency component index in the frequency domain; represents a spatial domain coordinate represents an intensity of a pixel at a position, represents a row index of a pixel in the spatial domain, represents a column index of a pixel in the spatial domain; represents an imaginary unit;

[0105] The band-pass filter is represented as:

[0106]

[0107] wherein, represents a response value of the band-pass filter at a frequency point ; represents a Euclidean distance from the frequency point to a frequency center point; represents a center frequency; represents a frequency band bandwidth;

[0108] The inverse Fourier transform is represented as:

[0109]

[0110] In order to further expand the image gray scale difference in the segmentation result, the application defines a difference expansion mapping relationship function, which is represented as:

[0111] ,

[0112] wherein, represents an intermediate parameter; represents a difference expansion mapping relationship function; represents an amplification factor; represents a maximum value of an image gray scale pixel value, which is 255; represents an inverse Fourier transform operator; represents a Fourier transform operator; represents a rounding up function.

[0113] The difference expansion mapping relationship function is, The greater the value is, the more significant the difference in the area with high dust density is, but at the same time, the gray scale value in the area with low dust density is lower, so that the difference processing is realized.

[0114] The final segmentation result image is obtained by weighting each segmentation image processed by the above image processing.

[0115] Embodiment 2

[0116] As Figures 1-4As shown, the embodiment provides a system for performing the unstructured environment saliency semantic segmentation method of embodiment 1, comprising:

[0117] a semantic segmentation network, which is obtained by replacing the feature extraction module in the traditional Unet network with a ghost affinity relation attention module, the ghost affinity relation attention module comprising a ghost convolution layer, an affinity relation attention layer and an output layer connected in sequence, wherein:

[0118] the ghost convolution layer is used to obtain an unstructured environment image to be segmented, and generate an original feature map of the unstructured environment image using ghost convolution;

[0119] the affinity relation attention layer is used to learn the importance of different positions and different channels in the original feature map respectively using affinity relation spatial attention and affinity relation channel attention, and generate a feature map considering affinity relation attention;

[0120] the output layer is used to map the feature map considering affinity relation attention to a category space, generate a category probability distribution of each pixel, and obtain a segmentation result;

[0121] a difference amplification module, which is used to sequentially perform Fourier transform, band-pass filtering and inverse Fourier transform on the segmentation result to denoise and amplify the image gray scale difference in the segmentation result.

[0122] The semantic segmentation network retains the nested U-shaped structure of the Unet network, comprising 6 layers of encoders and 5 layers of decoders, and each layer of the encoder and the decoder uses the ghost affinity relation attention module as a basic module for feature extraction. The encoder layer extracts abstract features of the input image through downsampling, and the decoder reconstructs the details of the image through upsampling, and outputs a saliency map, which uses predicted probability values as dust concentration values to reflect the concentration difference of dust. The encoder and the decoder are connected by a skip connection, which directly transmits the local features of the encoder to the decoder to make up for the loss of spatial information in downsampling.

[0123] Each decoder of the semantic segmentation network outputs a segmentation result, and the final segmentation result map is obtained by weighted sum of the output results of all decoders.

[0124] The embodiments of the application are described above in combination with the drawings, but the application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the application without departing from the purpose of the application and the scope protected by the claims, which are all within the protection of the application.

Claims

1. A method for salient semantic segmentation in unstructured environments, the method comprising: The method comprises the following steps: Step S1, obtaining an unstructured environment image to be segmented, and generating an original feature map of the unstructured environment image by using ghost convolution; Step S2, learning the importance of different positions and different channels in the original feature map by using affinity relationship spatial attention and affinity relationship channel attention respectively, and generating a feature map considering affinity relationship attention; Step S3, mapping the feature map considering affinity relationship attention to a category space, generating a category probability distribution of each pixel, and obtaining a segmentation result; Step S4, sequentially performing Fourier transform, band pass filtering and inverse Fourier transform on the segmentation result for denoising, and introducing a difference expansion mapping relationship function to expand the image gray difference in the segmentation result. The difference expansion mapping relationship function is represented as: ， wherein denotes an intermediate parameter; denotes a difference amplification mapping function; denotes an amplification factor; denotes a maximum value of image gray pixel value, taking a value of 255; denotes an inverse Fourier transform operator; denotes a Fourier transform operator; denotes a ceiling function; denotes a frequency point in a band-pass filter Euclidean distance from a frequency center point; denotes a frequency band bandwidth.

2. The unstructured environment saliency semantic segmentation method according to claim 1, characterized in that, The generation process of the original feature map specifically comprises the following steps: Step S11: performing convolution operation on the unstructured environment image to extract features, to generate a feature map with a size of , wherein, , , respectively represent the number of channels, height and width of the feature map .​ Step S12, performing ghost convolution on the feature map to generate a feature map ; Step S13, generating a feature map for the feature performing a convolution operation once to generate a feature map ; Step S14, the feature map with the feature map along the channel dimension to obtain a feature map with a size of original feature map ; Step S15, the original feature map batch normalization is performed and a hardswish function is activated.

3. The unstructured environment saliency semantic segmentation method according to claim 2, characterized in that, The processing process of the affinity relationship spatial attention specifically comprises the following steps: Step S211, the original feature map Global maximum pooling and global average pooling are respectively performed on the channel dimension to obtain two spatial attention vectors with sizes of , ; Step S212, the original feature map partitioning into individual feature units and tiling and reshaping into a feature map ; wherein , denotes the set of real numbers; Step S213, the feature map respectively through two different conversion functions and conversion, generate feature map and , wherein the conversion function and consists of a convolution kernel of 1, batch normalization and ReLU activation function; Step S214, the feature map and the feature map do matrix multiplication to get the affinity matrix ; Step S215, reshape the affinity matrix respectively as row and column elements two relational feature maps and ; Step S216, concatenating the channel attention vectors , and the relationship feature map and to obtain a feature map in the channel direction; Step S217, the feature map respectively through two conversion functions and extract global information, generate affinity relationship space attention weight matrix with size , wherein consisting of a convolution layer with a convolution kernel of 1, a batch normalization layer and a ReLU activation function, consisting of a convolution layer with a convolution kernel of 1, a batch normalization layer and a Sigmoid activation function.

4. The unstructured environment saliency semantic segmentation method according to claim 3, characterized in that, The processing process of the affinity relationship channel attention specifically comprises the following steps: Step S221, the original feature map Global maximum pooling and global average pooling are respectively performed on the spatial dimension, so that two feature maps with sizes of Channel attention vector , ; Step S222, divide the feature map into C feature units in the channel direction, and tile and reshape the feature map ; Step S223, the feature map respectively through two different conversion functions and converted into a feature map and , the conversion functions and are composed of a convolution layer with a convolution kernel of 1, batch normalization and ReLU activation function; Step S224, the feature map and Matrix multiplication is done to get the affinity matrix ; Step S225, reshaping the affinity matrix respectively as row and column elements two relational feature maps and ; Step S226, the channel attention vector is obtained by 、 and the relationship feature map and splicing in the channel direction to obtain a new feature map ; Step S227, the feature map By the conversion function And Extract global information to generate an affinity relationship channel attention weight matrix with a size of . .

5. The unstructured environment saliency semantic segmentation method according to claim 4, characterized in that, The generation process of the feature map considering affinity relationship attention specifically comprises the following steps: Step S231, the original feature map with the affinity relationship space attention weight matrix element-wise multiplication along the channel direction to obtain a feature map ; Step S232, the feature map with the affinity channel attention weight matrix element-wise multiplication along the spatial positions, obtaining a feature map ; Step S233, finally, the original feature map with the feature map element-wise addition to obtain the feature map considering the attention of the affinity relationship .

6. The unstructured environment saliency semantic segmentation method according to claim 5, characterized in that, Step S3 specifically includes the following process: adjusting the channel number of the feature map using a 1x1 convolution, so that it is equal to the class number, and then using a Sigmoid function to compress the output value of each pixel to [0,1], representing the probability of belonging to dust or haze. Step S3 specifically includes the following process: adjusting the channel number of the feature map using a 1x1 convolution, so that it is equal to the class number, and then using a Sigmoid function to compress the output value of each pixel to [0,1], representing the probability of belonging to dust or haze.

7. The unstructured environment saliency semantic segmentation method according to claim 6, characterized in that, The process of the Fourier transform is as follows: wherein represents a complex-valued Fourier coefficient at a frequency point represents a complex-valued Fourier coefficient at a frequency point represents a vertical spatial frequency component index in the frequency domain, represents a horizontal spatial frequency component index in the frequency domain; represents an intensity of a pixel at a spatial domain coordinate represents an intensity of a pixel at a spatial domain coordinate represents a row index of a pixel in the spatial domain, represents a column index of a pixel in the spatial domain; represents the imaginary unit; The band pass filter is represented as: wherein represents a response value of the band-pass filter at a frequency point ; and represents a center frequency; The inverse Fourier transform is represented as: 。 8. A system for performing the unstructured environment salient semantic segmentation method of any one of claims 1-7, characterized in that, It comprises: The semantic segmentation network is obtained by replacing the feature extraction module in the traditional Unet network with a ghost affinity relationship attention module, and the ghost affinity relationship attention module comprises a ghost convolution layer, an affinity relationship attention layer and an output layer connected in sequence, wherein: The ghost convolution layer is used to obtain an unstructured environment image to be segmented, and generate an original feature map of the unstructured environment image by using ghost convolution; The affinity relationship attention layer is used to learn the importance of different positions and different channels in the original feature map by using affinity relationship spatial attention and affinity relationship channel attention respectively, and generate a feature map considering affinity relationship attention; The output layer is used to map the feature map considering affinity relationship attention to a category space, generate a category probability distribution of each pixel, and obtain a segmentation result; The difference expansion module is used to sequentially perform Fourier transform, band pass filtering and inverse Fourier transform on the segmentation result for denoising, and introduce a difference expansion mapping relationship function to expand the image gray difference in the segmentation result.

9. The system of claim 8, wherein, The semantic segmentation network comprises 6 layers of encoders and 5 layers of decoders, each layer of the encoder and the decoder takes the ghost affinity relationship attention module as a basic module for feature extraction, the encoder layer extracts abstract features of the input image through downsampling, the decoder reconstructs the details of the image through upsampling, and outputs a saliency map, the saliency map takes a predicted probability value as a dust concentration value, and is used to reflect the concentration difference of the dust; the encoder and the decoder are connected by a skip connection, the local features of the encoder are directly transmitted to the decoder, and the loss of spatial information in the downsampling is compensated; each decoder of the semantic segmentation network outputs a segmentation result, and the output results of all the decoders are weighted and summed to obtain a final segmentation result map.

Images