A texture-aware image smoke detection method

Abstract

This paper proposes a texture-aware image-based fine-grained smoke detection method. The method comprises: a feature extraction module for acquiring low-level, mid-level, and high-level local features; a self-attention expression module for deriving global smoke attention features from high-level local features; a global convolutional texture perception module for acquiring smoke object-level features and internal detail features from mid-level local features; and a feature decoding module for fusing all of these features and mapping them into a smoke probability map as the result of fine-grained smoke detection. This method can accurately capture smoke locations in images and reflect internal details of smoke puffs in an end-to-end manner, providing clues for subsequent tasks in early fire detection, such as analyzing smoke concentration and puff motion trends. It has promising application prospects in the field of fire safety.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning and multimedia data analysis, and in particular to a texture perception-based image smoke fine detection method. Background Art

[0002] Preventing fires and detecting and assessing them early can not only effectively reduce or even avoid significant economic and plant and animal losses, but also minimize or prevent casualties and conserve social resources dedicated to police response. Fires typically produce large amounts of smoke in their early stages, and smoke detection can provide early warning of fires.

[0003] The dynamic characteristics of smoke are unique, but over-reliance on these characteristics can easily lead to false positives. Research has shown that the texture characteristics of smoke are unique and robust, and leveraging them can improve the accuracy of smoke detection in images and effectively reduce false positives. However, objects with similar textures to smoke, such as clouds and fog, can still lead to false positives, posing a challenge in further improving the robustness of smoke detection that relies on texture features.

[0004] Among existing smoke detection methods, those based on semantic segmentation models offer the highest accuracy. However, they all rely on convolution operations, making them susceptible to false positives and false negatives due to the influence of local fields of view. Furthermore, unique smoke semantics, such as the internal characteristics of puffs and the blurred edges of smoke, are not fully captured, compromising the robustness of smoke detection. Furthermore, attention mechanisms can capture long-range information, overcoming the limitations of local features, and attention maps can reflect blurred edges of smoke. Therefore, combining attention with convolution helps integrate object-level smoke features with local-level features of internal smoke structure. However, global attention is computationally expensive, limiting its applicability to smoke detection.

[0005] Based on this, it is necessary to propose a texture-aware image smoke fine detection method to solve the above technical problems. Summary of the Invention

[0006] In view of the above situation, the main purpose of the present invention is to propose a texture-based image smoke fine detection method to solve the above technical problems.

[0007] An embodiment of the present invention provides a texture-based image smoke detection method, characterized in that the method is implemented using a smoke detection network, which includes a feature extraction module, a self-attention expression module, a global convolution texture perception module, and a feature decoding module. The method includes the following steps:

[0008] Step 1: Obtain a colored smoke composite image as input;

[0009] Step 2: Based on the feature extraction module, the colored smoke composite image is received and extracted to obtain low-level local features, mid-level local features, and high-level local features;

[0010] Step 3: Based on the self-attention expression module, the high-level local features are used as input to extract the global smoke attention features from the high-level local features;

[0011] Step 4: Based on the global convolution texture perception module, the middle-layer local features are used as input to extract the smoke object-level features and the smoke internal detail features from the middle-layer local features;

[0012] Step 5: Based on the feature decoding module, the low-level local features, mid-level local features, high-level local features, global smoke attention features, smoke object-level features, and smoke internal detail features are used as input, and feature fusion is performed and mapped into a smoke probability map, which is the result of smoke detection.

[0013] This paper proposes a texture-aware image-based fine-grained smoke detection method. The method comprises: a feature extraction module for acquiring low-level, mid-level, and high-level local features; a self-attention expression module for deriving global smoke attention features from high-level local features; a global convolutional texture perception module for acquiring smoke object-level features and internal detail features from mid-level local features; and a feature decoding module for fusing all of these features and mapping them into a smoke probability map as the result of fine-grained smoke detection. This method can accurately capture smoke locations in images and reflect internal details of smoke puffs in an end-to-end manner, providing clues for subsequent tasks in early fire detection, such as analyzing smoke concentration and puff motion trends. It has promising application prospects in the field of fire safety.

[0014] In step 1, the colored smoke composite image is represented as:

[0015] I(x)=αS(x)+(1-α)B(x)

[0016] Where I(x) represents the color smoke composite image, S(x) represents the smoke pixel at coordinate x, B(x) represents the background pixel at coordinate x, and α represents the composite coefficient at coordinate x, α∈[0,1].

[0017] In the step 2, the feature extraction module uses MobildNetV2 as the backbone network for extracting smoke features. The backbone network has 28 layers, including a shallow network, a middle network, and a deep network. The shallow network corresponds to the 7th layer, the middle network corresponds to the 11th layer, and the deep network corresponds to the 28th layer. The shallow network is used to output low-level local features x low , the middle layer network is used to output the middle layer local features x mid , the deep network is used to output high-level local features x high .

[0018] In step 3, the method of extracting the global smoke attention feature from the high-level local features using the high-level local features as input based on the self-attention expression module includes the following steps:

[0019] Step 3.1: transform the high-level local features x high Input to the self-attention expression module, and the high-level local features x are transformed into high Decompose into a first smoke sub-feature graph Q, a second smoke sub-feature graph K, and a third smoke sub-feature graph V;

[0020] Step 3.2: Calculate the texture attention map A based on the first smoke sub-feature map Q and the second smoke sub-feature map K;

[0021] Step 3.3, apply the texture attention map A to the third smoke sub-feature map V and the high-level local feature x high Perform weighted fusion to obtain the intermediate global smoke attention feature;

[0022] Step 3.4: Repeat steps 3.1 to 3.3 above, and input the intermediate global smoke attention feature into the self-attention expression module again. After three steps of sub-feature map decomposition, texture attention map calculation, and feature fusion, the global smoke attention feature x is finally obtained. ga .

[0023] In step 3.2, the calculation formula of the texture attention map A is expressed as:

[0024]

[0025] Among them, Q i Represents the vector corresponding to the i-th spatial position of the first smoke sub-feature map Q, K j，l represents the j-th component of the vector at the i-th spatial position corresponding to the second smoke sub-feature map K and the first smoke sub-feature map Q, T represents the transpose operation, and softmax(·) represents the normalization operation.

[0026] In step 3.3, the texture attention map A is applied to the third smoke sub-feature map V and the high-level local feature x high In the method of performing weighted fusion to obtain the intermediate global smoke attention feature, the calculation formula corresponding to the intermediate global smoke attention feature is expressed as:

[0027]

[0028] in, represents the intermediate global smoke attention feature, V n Indicates the channel range corresponding to the nth spatial coordinate, A m，n Represents the attention value corresponding to the m-th channel index at the n-th spatial coordinate, V m，n Represents the sub-eigenvalue corresponding to the m-th channel index at the n-th spatial coordinate.

[0029] In step 4, based on the global convolution texture perception module, the method of taking the middle-layer local features as input and extracting smoke object-level features and smoke internal detail features from the middle-layer local features includes the following steps:

[0030] Step 4.1: Use one-dimensional global convolution to obtain the first smoke sub-feature map Q and the second smoke sub-feature map K from the middle-level local features, and use matrix multiplication to calculate the texture attention map A;

[0031] Step 4.2, calculate the internal detail features x of the smoke fuse ;

[0032] Step 4.3, based on the smoke internal detail features x fuse , texture attention map A and mid-level local features x mid Calculate the smoke object-level feature x te .

[0033] In step 4.1, the corresponding calculation formula is as follows:

[0034] Q=f reshape (f dv (x mid ))

[0035] K=f reshape (f dh (σ id ))

[0036] A=softmax(Q i K i )

[0037] Among them, f reshape (·) represents the operation of reorganizing the matrix into vectors by rows, f dv(·) represents the local feature x in the middle layer mid A vertical one-dimensional convolution operation with equal length and width, f dh (·) represents the local feature x in the middle layer mid Horizontal one-dimensional convolution operation with equal length and width, K i Represents the vector at the i-th spatial position corresponding to the second smoke sub-feature image K and the first smoke sub-feature image Q;

[0038] In step 4.2, the calculation formula corresponding to the detailed features inside the smoke is expressed as:

[0039] x fuse =f fuse (f inp (σ(f dv (x mid )+f dh (x mid ))))

[0040] Among them, f fuse (·) represents a 1×1 convolution operation, f inp (·) represents the interpolation operation, σ(·) represents the nonlinear activation operation;

[0041] In step 4.3, the smoke object-level feature x te The corresponding calculation formula is expressed as:

[0042] x te =A⊙(x fuse +x mid )

[0043] Here, ⊙ represents the Hadamard product.

[0044] In step 5, based on the feature decoding module, the low-level local features, mid-level local features, high-level local features, global smoke attention features, smoke object-level features, and smoke internal detail features are used as inputs, and feature fusion is performed and mapped into a smoke probability map. The smoke probability map is the result of smoke detection. The method includes the following steps:

[0045] Step 5.1, transform the smoke object level feature x te With the global smoke attention feature x ga Perform fusion to obtain the first fusion feature x1;

[0046] Step 5.2: Combine the first fusion feature x1 with the low-level local feature x low Perform fusion to obtain the channel feature map x2;

[0047] In step 5.3, the channel feature map x2 is subjected to a channel-by-channel nonlinear normalization operation to obtain a smoke probability map.

[0048] In step 5.1, the calculation formula corresponding to the first fusion feature x1 is expressed as:

[0049] x1=f res (f cat (x te , x ga ))

[0050] Among them, f rcs (·) represents the residual convolution operation, f cat (·) indicates the splicing operation along the channel direction;

[0051] In step 5.2, the calculation formula corresponding to the channel feature map x2 is expressed as:

[0052] x2=f out (f cat (f duc (x low ), x1))

[0053] Among them, f out (·) represents the normal convolution operation, f duc (·) indicates an upsampling operation where the number of channels is divided by 4 and the spatial resolution is doubled;

[0054] In step 5.3, the corresponding calculation formula is expressed as:

[0055]

[0056] in, Represents the smoke probability map x out The cth channel of Represents the c-th channel of the channel feature map x2.

[0057] Additional aspects and advantages of the present invention will be set forth in part in the following description and, in part, will be obvious from the following description, or may be learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the texture-aware image smoke fine detection method proposed by the present invention;

[0059] Figure 2 This is a schematic diagram of the structure of the smoke detection network proposed in the present invention;

[0060] Figure 3 Schematic diagram of the structure of the global convolution texture perception module in the present invention. DETAILED DESCRIPTION

[0061] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0062] These and other aspects of the embodiments of the present invention will become apparent with reference to the following description and accompanying drawings. While some specific implementations of the embodiments of the present invention are disclosed in detail in these descriptions and accompanying drawings to illustrate some ways of implementing the principles of the embodiments of the present invention, it should be understood that the scope of the embodiments of the present invention is not limited thereby. On the contrary, the embodiments of the present invention encompass all variations, modifications, and equivalents that fall within the spirit and scope of the appended claims.

[0063] See also Figure 1 The present invention proposes a texture-based image smoke fine detection method, which is characterized by being implemented through a smoke detection network, wherein the smoke detection network (see Figure 2 ) includes a feature extraction module, a self-attention expression module, a global convolution texture perception module and a feature decoding module, and the method includes the following steps:

[0064] Step 1: Obtain a colored smoke composite image as input.

[0065] In step 1, the colored smoke composite image is represented as:

[0066] I(x)=αS(x)+(1-α)B(x)

[0067] Where I(x) represents the color smoke composite image, S(x) represents the smoke pixel at coordinate x, B(x) represents the background pixel at coordinate x, and α represents the composite coefficient at coordinate x, α∈[0,1].

[0068] As a supplementary note, both smoke pixels and background pixels are from public datasets. The background pixels can be replaced by any natural image, and the smoke pixels can be replaced by any virtual color image with RGBA channels.

[0069] Step 2: Based on the feature extraction module, the colored smoke composite image is received and extracted to obtain low-level local features, middle-level local features and high-level local features.

[0070] In step 2, the feature extraction module uses MobileNetV2 as the backbone network to extract smoke features. The backbone network has 28 layers, including shallow network, middle network and deep network. The shallow network corresponds to the 7th layer, the middle network corresponds to the 11th layer, and the deep network corresponds to the 28th layer. In this embodiment, the shallow network is used to output low-level local features x low , the middle layer network is used to output the middle layer local features x mid , the deep network is used to output high-level local features x high .

[0071] Step 3: Based on the self-attention expression module, the high-level local features are used as input, and the global smoke attention features are extracted from the high-level local features.

[0072] In step three, the following steps are specifically included:

[0073] Step 3.1: transform the high-level local features x high Input to the self-attention expression module, and the high-level local features x are transformed into high Decomposed into the first smoke sub-feature graph Q, the second smoke sub-feature graph K and the third smoke sub-feature graph V.

[0074] Step 3.2: Calculate the texture attention map A based on the first smoke sub-feature map Q and the second smoke sub-feature map K.

[0075] In step 3.2, the calculation formula of the texture attention map A is expressed as:

[0076]

[0077] Among them, Q i Represents the vector corresponding to the i-th spatial position of the first smoke sub-feature map Q, K j，i represents the j-th component of the vector at the i-th spatial position corresponding to the second smoke sub-feature map K and the first smoke sub-feature map Q, T represents the transpose operation, and softmax(·) represents the normalization operation.

[0078] Step 3.3, apply the texture attention map A to the third smoke sub-feature map V and the high-level local feature x high Weighted fusion is performed to obtain the intermediate global smoke attention features.

[0079] In step 3.3, the calculation formula corresponding to the intermediate global smoke attention feature is expressed as:

[0080]

[0081] in, represents the intermediate global smoke attention feature, V n Indicates the channel range corresponding to the nth spatial coordinate, A m，n Represents the attention value corresponding to the m-th channel index at the n-th spatial coordinate, V m，n Represents the sub-eigenvalue corresponding to the m-th channel index at the n-th spatial coordinate.

[0082] Step 3.4: Repeat steps 3.1 to 3.3 above, and input the intermediate global smoke attention feature into the self-attention expression module again. After three steps of sub-feature map decomposition, texture attention map calculation, and feature fusion, the global smoke attention feature x is finally obtained. ga .

[0083] Step 4: Based on the global convolution texture perception module, the middle-layer local features are used as input, and the smoke object-level features and the smoke internal detail features are extracted from the middle-layer local features.

[0084] In step 4 (see Figure 3 ), specifically including the following steps:

[0085] In step 4.1, one-dimensional global convolution is used to obtain the first smoke sub-feature map Q and the second smoke sub-feature map K from the middle-level local features, and the texture attention map A is obtained by matrix multiplication.

[0086] In step 4.1, the corresponding calculation formula is as follows:

[0087] Q=f reshape (f dv (x mid ))

[0088] K=f reshape (f dh (x mid ))

[0089] A=softmax(Q i K i )

[0090] Among them, f reshape (·) represents the operation of reorganizing the matrix into vectors by rows, f dv (·) represents the local feature x in the middle layer mid A vertical one-dimensional convolution operation with equal length and width, f dh (·) represents the local feature x in the middle layer mid Horizontal one-dimensional convolution operation with equal length and width, K i Represents the vector at the i-th spatial position corresponding to the second smoke sub-feature image K and the first smoke sub-feature image Q.

[0091] Step 4.2, calculate the internal detail features x of the smoke fuse .

[0092] The calculation formula corresponding to the detailed features inside the smoke is expressed as:

[0093] x fuse =f fuse (f inp (σ(f dv (x mid )+f dh (x mid ))))

[0094] Among them, f fuse (·) represents a 1×1 convolution operation, f inp (·) represents the interpolation operation, and σ(·) represents the nonlinear activation operation.

[0095] Step 4.3, based on the smoke internal detail features x fuse , texture attention map A and mid-level local features x mid Calculate the smoke object-level feature x te .

[0096] In step 4.3, the smoke object-level feature x te The corresponding calculation formula is expressed as:

[0097] x te =A⊙(x fuse +x mid )

[0098] Here, ⊙ represents the Hadamard product.

[0099] In the present invention, the smoke internal detail feature x fuse , middle-level local features x mid , texture attention map A and the final smoke object-level feature x te Have the same number of channels and spatial dimensions.

[0100] Step 5: Based on the feature decoding module, the low-level local features, mid-level local features, high-level local features, global smoke attention features, smoke object-level features, and smoke internal detail features are used as input, and feature fusion is performed and mapped into a smoke probability map, which is the result of smoke detection.

[0101] In step five, the following steps are specifically included:

[0102] Step 5.1, transform the smoke object level feature x te With the global smoke attention feature x ga The fusion is performed to obtain the first fused feature x1.

[0103] Among them, the calculation formula corresponding to the first fusion feature x1 is expressed as:

[0104] x1=f res (f cat (x te , x ga ))

[0105] Among them, f res (·) represents the residual convolution operation, f cat (·) indicates the splicing operation along the channel direction.

[0106] Step 5.2: Combine the first fusion feature x1 with the low-level local feature x low Fusion is performed to obtain the channel feature map x2.

[0107] Among them, the calculation formula corresponding to the channel feature map x2 is expressed as:

[0108] x2=f out (f cat (f duc (x low ), x1))

[0109] Among them, f out (·) represents the normal convolution operation, f duc (·) represents an upsampling operation where the number of channels is divided by 4 and the spatial resolution is doubled. out (·) takes 1 as the convolution kernel number, then the channel feature map x2 is a single channel feature map; if f out The number of convolution kernels of (·) is 4, so the channel feature map x2 is a four-channel feature map.

[0110] In step 5.3, the channel feature map x2 is subjected to a channel-by-channel nonlinear normalization operation to obtain a smoke probability map.

[0111] The calculation formula corresponding to the smoke probability map is expressed as:

[0112]

[0113] in, Represents the smoke probability map x out The cth channel of Represents the c-th channel of the channel feature map x2.

[0114] If the channel feature map x2 is a single-channel feature map, then the obtained x out is the smoke probability map; if the channel feature map x2 is a four-channel feature map, then the obtained x outis a normalized RGBA image. The fourth channel is the resulting smoke probability map. Regardless of whether the number of channels is 1 or 4, each element in the smoke probability map between 0 and 1 reflects the probability of smoke appearing or the concentration of smoke.

[0115] This paper proposes a texture-aware image-based fine-grained smoke detection method. The method comprises: a feature extraction module for acquiring low-level, mid-level, and high-level local features; a self-attention expression module for deriving global smoke attention features from high-level local features; a global convolutional texture perception module for acquiring smoke object-level features and internal detail features from mid-level local features; and a feature decoding module for fusing all of these features and mapping them into a smoke probability map as the result of fine-grained smoke detection. This method can accurately capture smoke locations in images and reflect internal details of smoke puffs in an end-to-end manner, providing clues for subsequent tasks in early fire detection, such as analyzing smoke concentration and puff motion trends. It has promising application prospects in the field of fire safety.

[0116] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0117] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0118] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A texture-based image smoke detection method, characterized in that: This is achieved through a smoke detection network, which includes a feature extraction module, a self-attention expression module, a global convolution texture perception module, and a feature decoding module. The method includes the following steps: Step 1: Obtain a colored smoke composite image as input; Step 2: Based on the feature extraction module, the colored smoke composite image is received and extracted to obtain low-level local features, mid-level local features, and high-level local features; Step 3: Based on the self-attention expression module, the high-level local features are used as input to extract the global smoke attention features from the high-level local features; Step 4: Based on the global convolution texture perception module, the middle-layer local features are used as input to extract the smoke object-level features and the smoke internal detail features from the middle-layer local features; Step 5: Based on the feature decoding module, the low-level local features, global smoke attention features, and smoke object-level features are used as input to perform feature fusion and map them into a smoke probability map, which is the result of smoke detection; In step 4, based on the global convolution texture perception module, the method of taking the middle-layer local features as input and extracting smoke object-level features and smoke internal detail features from the middle-layer local features includes the following steps: Step 4.1: Use one-dimensional global convolution to obtain the first smoke sub-feature map from the middle-level local features Q And the second smoke sub-characteristic map K , and use matrix multiplication to calculate the texture attention map A ; Step 4.2: Calculate the internal details of the smoke ; Step 4.3, based on the detailed features of the smoke , texture attention map A and mid-level local features Calculate the smoke object-level features ; In step 4.1, the corresponding calculation formula is as follows: ； ； ； in, Indicates the operation of reorganizing the matrix into vectors by rows. Representation and mid-level local features A vertical one-dimensional convolution operation with equal length and width, Representation and mid-level local features Horizontal one-dimensional convolution operation with equal length and width, Represents the second smoke sub-feature map K and the first smoke sub-characteristic map Q Corresponding to i A vector at a spatial position; In step 4.2, the calculation formula corresponding to the detailed features inside the smoke is expressed as: ； in, express Convolution operation, represents the interpolation operation, represents a nonlinear activation operation; In step 4.3, the smoke object-level features The corresponding calculation formula is expressed as: ； in, Represents the Hadamard product.

2. The texture-aware image smoke detection method according to claim 1, characterized in that: In step 1, the colored smoke composite image is represented as: ； in, represents a composite image of colored smoke, Indicates that the coordinates are x Smoke pixels at Indicates that the coordinates are x The background pixels at Indicates that the coordinates are x The composite coefficient at .

3. The texture-aware image smoke detection method according to claim 2, characterized in that: In the step 2, the feature extraction module uses MobileNetV2 as the backbone network for extracting smoke features. The backbone network has 28 layers, including shallow network, middle network and deep network. The shallow network corresponds to the 7th layer, the middle network corresponds to the 11th layer, and the deep network corresponds to the 28th layer. The shallow network is used to output low-level local features. , the middle layer network is used to output the middle layer local features , the deep network is used to output high-level local features .

4. The texture-aware image smoke detection method according to claim 3, characterized in that: In step 3, the method of extracting the global smoke attention feature from the high-level local features using the high-level local features as input based on the self-attention expression module includes the following steps: Step 3.1: high-level local features Input to the self-attention expression module, through three The convolution kernel transforms high-level local features Decomposed into the first smoke sub-feature map Q , Second smoke sub-characteristic map K And the third smoke sub-characteristic map V ; Step 3.2, based on the first smoke sub-feature map Q and the second smoke sub-characteristic map K Calculate the texture attention map A ; Step 3.3, texture attention map A Applied to the third smoke sub-feature map V With high-level local features Perform weighted fusion to obtain the intermediate global smoke attention feature; Step 3.4, repeat the above steps 3.1 to 3.3, and input the intermediate global smoke attention feature into the self-attention expression module again. After three steps of sub-feature map decomposition, texture attention map calculation and feature fusion, the global smoke attention feature is finally obtained. .

5. The texture-aware image smoke detection method according to claim 4, characterized in that: In step 3.2, the texture attention map A The calculation formula is expressed as: ； in, Represents the first smoke sub-feature map Q No. i The vector corresponding to the spatial position, Represents the second smoke sub-feature map K and the first smoke sub-characteristic map Q Corresponding to i The vector at the spatial position j A quantity, represents the transpose operation, Represents a normalization operation.

6. The texture-aware image smoke detection method according to claim 5, characterized in that: In step 3.3, the texture attention map A Applied to the third smoke sub-feature map V With high-level local features In the method of performing weighted fusion to obtain the intermediate global smoke attention feature, the calculation formula corresponding to the intermediate global smoke attention feature is expressed as: ； in, represents the intermediate global smoke attention feature, Indicates the n The channel range corresponding to the spatial coordinates is Indicates the n The spatial coordinates of m The attention value corresponding to the channel index, Indicates the n The spatial coordinates of m The sub-eigenvalue corresponding to the channel index.

7. The texture-aware image smoke detection method according to claim 1, characterized in that: In step 5, based on the feature decoding module, the low-level local features, the global smoke attention features, and the smoke object-level features are used as inputs, and feature fusion is performed and mapped into a smoke probability map. The smoke probability map is the result of smoke detection. The method includes the following steps: Step 5.1, the smoke object level features With global smoke attention feature Fusion is performed to obtain the first fusion feature ; Step 5.2, the first fusion feature and low-level local features Fusion is performed to obtain the channel feature map ; Step 5.3, channel feature map After a channel-by-channel nonlinear normalization operation, the smoke probability map is obtained.

8. The texture-aware image smoke detection method according to claim 7, characterized in that: In step 5.1, the first fusion feature The corresponding calculation formula is expressed as: ； in, represents the residual convolution operation, Indicates the splicing operation along the channel direction; In step 5.2, the channel feature map The corresponding calculation formula is expressed as: ； in, represents the normal convolution operation, Indicates an upsampling operation where the number of channels is divided by 4 and the spatial resolution is doubled; In step 5.3, the corresponding calculation formula is expressed as: ； in, Represents smoke probability map No. c channels, Represents channel feature map No. c channels.

Images