Straw incineration smoke detection algorithm based on adaptive space perception

Abstract

The invention discloses a straw incineration smoke detection algorithm based on adaptive spatial perception, which comprises the following steps: acquiring an aerial image of a target area, making a smoke image data set, and dividing the data set into a training set, a verification set and a test set; constructing a smoke detection network comprising a feature extraction network, a target identification network and a post-processing network; an adaptive space sensing module is embedded in the feature extraction network; the target identification network comprises a weighted bidirectional feature pyramid network module, a multi-group collaborative attention module and a detection head, and is used for identifying smoke and straw in the image; the post-processing network calculates the relative position relation between the smoke and the straw; according to the method, the smoke recognition capability in a complex environment can be remarkably enhanced, the detection effect of a system on smoke in different scenes is improved, the representation capability of a model on semantic information is enhanced, so that the sensitivity to subtle differences is improved, and the recognition stability and reliability in a variable environment can be effectively improved in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of smoke detection, and in particular to a straw burning smoke detection algorithm based on adaptive spatial perception. Background Art

[0002] Burning straw has always been a traditional way for farmers to deal with agricultural waste. However, with the development of society, its harm to the environment has gradually been exposed. Harmful gases such as carbon dioxide, carbon monoxide, and nitrogen oxides released during straw burning every year not only directly aggravate global warming and promote the deterioration of the greenhouse effect, but also seriously pollute the atmospheric environment and pose a clear threat to residents' health.

[0003] Smoke is one of the main characteristics of straw burning. Smoke monitoring of straw burning can provide important scientific basis and decision-making support for straw burning management. In the existing technology, image processing technology is usually used to make up for the shortcomings of traditional monitoring methods that rely on environmental monitoring points distributed in specific areas and smoke sensors to monitor straw burning through visual monitoring means. However, the traditional image processing methods are as follows: 1. A smoke detection algorithm combining discrete wavelet transform and correlation analysis is proposed. This method extracts the target area through a Gaussian mixture model, uses wavelet transform to distinguish the low-frequency and high-frequency components of the image, and finally distinguishes smoke from non-smoke objects through correlation analysis of high-frequency information; 2. A smoke detection method based on RGB contrast image and shape constraint is proposed. This method performs wavelet transform based on RGB contrast image to distinguish smoke from other low-frequency signals, and analyzes the shape change and energy change of the region to determine whether smoke exists; 3. A smoke detection method based on image pyramid is proposed. First, local binary patterns and edge direction histograms are extracted, and then smoke feature vectors are generated through pooling and feature fusion. Finally, support vector machine is used for classification;

[0004] Although traditional smoke detection methods are effective in simple environments, they rely on artificially designed features, which often need to be adjusted according to specific scenarios and are difficult to cover all possible smoke features. Therefore, the above methods have shown great limitations when dealing with complex scenarios. With the innovation of artificial intelligence technology, smoke detection methods based on deep learning can adaptively extract more complex and diverse features by automatically learning smoke features in images, and have become an important development direction in the field of smoke detection: 1. In some technologies, the input layer structure is improved on the basis of YOLOv5s, and a straw burning smoke detection network suitable for multi-band remote sensing images is constructed, and the impact of spatial resolution on model detection is discussed; 2. In other technologies, in order to reduce the false alarm rate of smoke detection, the captured image is first blurred and filtered, and then the image clarity is evaluated using gradient information. Finally, FasterR-CNN is used to perform smoke detection on the deblurred image; 3. In some other technologies, a detection network combining deep separable convolution and channel attention mechanism is proposed, and a diverse dataset containing smoke from different geographical regions and crop types is constructed to achieve remote monitoring of straw burning smoke.

[0005] Existing detection methods based on image processing have made certain progress in the field of smoke monitoring, but it is worth noting that in real life not all smoke comes from straw burning. In addition, smoke images often contain complex background interference, which makes it difficult to distinguish between smoke and background, thus affecting the accuracy of smoke detectors. Although image processing technology has broad application prospects in straw burning monitoring, in practical applications, how to effectively deal with complex background interference is still one of the main challenges facing smoke detection technology. Summary of the invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a straw burning smoke detection algorithm based on adaptive spatial perception.

[0007] The technical solution adopted by an embodiment of the present invention to solve the technical problem is: a straw burning smoke detection algorithm based on adaptive spatial perception, comprising:

[0008] S1, obtain aerial images of the target area, create a smoke image dataset, and divide the dataset into a training set, a validation set, and a test set;

[0009] S2, constructing a smoke detection network including a feature extraction network, a target recognition network and a post-processing network, wherein step S2 includes the following steps:

[0010] S2.1, the feature extraction network embeds an adaptive spatial perception module to extract smoke features;

[0011] S2.2, the target recognition network includes a weighted bidirectional feature pyramid network module, a multi-group collaborative attention module and a detection head to identify smoke and straw in the image;

[0012] S2.3, the post-processing network calculates the relative position relationship between the smoke and the straw;

[0013] S3, training the smoke detection network multiple times using the training set and the validation set to obtain a smoke detection model;

[0014] S4: Use the test set to test the smoke detection model and verify the performance of the smoke detection model.

[0015] As one of the preferred embodiments of the present invention, in step S2.1, the calculation process of the adaptive space perception module is represented by the following formulas (1)-(5):

[0016]

[0017] w=Softmax(FC(ReLU(FC(AvgPool(x))))) (2)

[0018]

[0019] B=ReLU(BN(Conv(M))) (4)

[0020]

[0021] Among them, y i is the output of the i-th convolution kernel; w i is the weight of the i-th convolution kernel. The weight generation process consists of average pooling, two fully connected layers, ReLU layer and Softmax layer. AvgPool represents average pooling and FC represents fully connected layer. M is the aggregate feature. The spatial data enhancement process consists of average pooling, maximum pooling, channel concatenation, convolution and Sigmoid layer. MaxPool represents maximum pooling, Concat represents channel concatenation, Conv represents convolution and σ represents Sigmoid. ReLU and Sigmoid activation functions are expressed by formulas (6) and (7) respectively:

[0022]

[0023] As one of the preferred embodiments of the present invention, in step S2.2, the calculation process of the multi-group collaborative attention module is expressed by the following formulas (8)-(10):

[0024] S1=SE(Conv(X1)) (8)

[0025] S i =SE(Conv(Add(S i-1 ,X i )) (9)

[0026] G=Concat(S1,S2,...,S K ) (10)

[0027] Among them, for a given feature X∈R C×H×W , C, H and W represent the number of channels, height and width of the feature respectively. The feature X is divided into K groups along the channel direction, X = [X1, X2, ..., X K ]; SE means compression and stimulation of attention; Add means feature addition; G means semantically enhanced features.

[0028] As one of the preferred embodiments of the present invention, the division ratio of the training set, the validation set and the test set is 8:1:1.

[0029] Beneficial effects of the present invention:

[0030] 1. The acquisition of aerial images through the combination of drones and image processing technology not only greatly improves the automation level of the monitoring system, but also enhances the timeliness and spatial coverage of monitoring; drones can quickly cover a large area, especially some difficult-to-reach geographical locations or sudden combustion events in a short period of time, thus effectively making up for the shortcomings of traditional monitoring methods.

[0031] 2. In practical applications, adaptive spatial perception technology can significantly enhance the ability to identify smoke in complex environments and improve the system's smoke detection effect in different scenarios.

[0032] 3. The multi-group attention mechanism can enhance the model's ability to represent semantic information, thereby increasing sensitivity to subtle differences. In practical applications, it can effectively improve recognition stability and reliability in changing environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0034] Figure 1 It is the main flow chart of a straw burning smoke detection algorithm based on adaptive spatial perception;

[0035] Figure 2 is a flow chart of the adaptive spatial perception module;

[0036] Figure 3 Flowchart of the multi-group collaborative attention module;

[0037] Figure 4 This is a specific implementation flow chart of a straw burning smoke detection algorithm based on adaptive spatial perception. DETAILED DESCRIPTION

[0038] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it cannot be understood as a limitation on the scope of protection of the present invention.

[0039] In the description of the present invention, the meaning of "above", "below", "exceed", etc. is not inclusive of the number itself, and the meaning of "above", "below", "within", etc. is inclusive of the number itself. If there is a description of "first" or "second", it is only used to distinguish the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0040] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0041] In the present invention, unless otherwise clearly defined, the words "set", "install", "connect" and the like should be understood in a broad sense, for example, they can be directly connected or indirectly connected through an intermediate medium; they can be fixedly connected or detachably connected or integrally formed; they can be mechanically connected; they can be the internal connection of two elements or the interaction relationship between two elements. Those skilled in the art can reasonably determine the specific meanings of the above words in the present invention in combination with the specific content of the technical solution.

[0042] Reference Figures 1 to 4 , a straw burning smoke detection algorithm based on adaptive spatial perception, including:

[0043] S1, obtain aerial images of the target area, create a smoke image dataset, and divide the dataset into a training set, a validation set, and a test set;

[0044] S2, constructing a smoke detection network including a feature extraction network, a target recognition network and a post-processing network, wherein step S2 includes the following steps:

[0045] S2.1, the feature extraction network embeds an adaptive spatial perception module to extract smoke features;

[0046] S2.2, the target recognition network includes a weighted bidirectional feature pyramid network module, a multi-group collaborative attention module and a detection head to identify smoke and straw in the image;

[0047] S2.3, the post-processing network calculates the relative position relationship between the smoke and the straw;

[0048] S3, training the smoke detection network multiple times using the training set and the validation set to obtain a smoke detection model;

[0049] S4: Use the test set to test the smoke detection model and verify the performance of the smoke detection model.

[0050] 1) In the present invention, a drone equipped with a high-resolution sensor is used to collect smoke image data in real time when flying to a straw burning area; the drone can accurately capture the distribution of smoke from different angles through flexible flight paths and altitude adjustment; the collected image data is processed using the LabelImg tool, the smoke area in each image is accurately labeled, the specific location of the smoke is calibrated using multiple rectangular boxes, and corresponding label information is generated; the labeled images are divided into a training set, a validation set, and a test set according to an 8:1:1 ratio, wherein the training set is used to train a smoke detection model so that it can learn to identify and distinguish smoke areas from images; the validation set is used to evaluate and tune the performance of the model during the training process, help select the most appropriate model parameters and prevent overfitting; the test set is used to finally evaluate the generalization ability of the model and test its performance on unseen data to ensure that the model can effectively identify straw burning smoke in practical applications.

[0051] 2) This paper proposes a straw burning smoke detection algorithm (Straw Burning Smoke Detection, SBSDet) based on adaptive spatial perception. By analyzing the high-resolution images taken by drones, the morphological characteristics of smoke are accurately captured, smoke and straw are automatically identified, and then post-processing methods are used to determine whether the smoke is related to straw burning activities; the specific structure of the SBSDet model is as follows Figure 1 shown.

[0052] The SBSDet model first embeds an adaptive spatial perception module in the feature extraction network to effectively capture the fine-grained features of smoke, thereby enhancing the network's sensitivity to smoke features and feature representation capabilities; then introduces a weighted bidirectional feature pyramid network to optimize multi-scale feature fusion and efficiently integrate feature information at different levels, and then combines it with a multi-group attention module to further enhance the model's representation of semantic information, and uses a detection head to identify smoke and straw in the image; finally, by calculating the relative position relationship between smoke and straw and combining the physical characteristics of smoke generation (smoke is usually located above or near the straw), the source of the smoke can be accurately determined.

[0053] 3)Reference Figure 2 , in step S2.1, the calculation process of the adaptive spatial perception module is expressed by the following formulas (1)-(5):

[0054]

[0055] w=Softmax(FC(ReLU(FC(AvgPool(x))))) (2)

[0056]

[0057] B=ReLU(BN(Conv(M))) (4)

[0058]

[0059] Among them, y i is the output of the i-th convolution kernel; w i is the weight of the i-th convolution kernel. The weight generation process consists of average pooling, two fully connected layers, ReLU layer and Softmax layer. AvgPool represents average pooling and FC represents fully connected layer. M is the aggregate feature. The spatial data enhancement process consists of average pooling, maximum pooling, channel concatenation, convolution and Sigmoid layer. MaxPool represents maximum pooling, Concat represents channel concatenation, Conv represents convolution and σ represents Sigmoid. ReLU and Sigmoid activation functions are expressed by formulas (6) and (7) respectively:

[0060]

[0061] Adaptive spatial perception technology first applies multiple convolution kernels in parallel to process the input feature X. Each convolution kernel captures different scales and local features in the spatial domain. Based on the characteristics of the input features, the weight vector w = [w1, w2, ..., w n], the weights are adaptively adjusted according to the input, so that different weights are assigned to each convolution kernel; through this mechanism, the model can realize dynamic selection of convolution kernels and perform weighted aggregation of the extracted features to obtain the aggregated features M; the aggregated features M are normalized to avoid certain features dominating the learning process and accelerate the convergence of the optimization algorithm; the normalized features are then nonlinearly transformed through the ReLU activation function. By introducing nonlinear factors, the network has stronger expressive power and can fit more complex input-output mapping relationships; finally, spatial enhancement is used to further optimize the representation of smoke features, and by strengthening the effective learning of spatial information, the enhanced features E are finally output to improve the performance and robustness of the model in smoke detection tasks.

[0062] 4)Reference Figure 3 , in step S2.2, the calculation process of the multi-group collaborative attention module is expressed by the following formulas (8)-(10):

[0063] S1=SE(Conv(X1)) (8)

[0064] S i =SE(Conv(Add(S i-1 ,X i )) (9)

[0065] G=Concat(S1,S2,...,S K ) (10)

[0066] Among them, for a given feature X∈R C×H×W , C, H and W represent the number of channels, height and width of the feature respectively. The feature X is divided into K groups along the channel direction, X = [X1, X2, ..., X K ]; SE means compression and stimulation of attention; Add means feature addition; G means semantically enhanced features.

[0067] The SE attention mechanism dynamically adjusts the weights of each channel to enhance the network's attention to important features while suppressing irrelevant features. It compresses and excites feature maps to improve the model's ability to express features. The SE attention block first uses global average pooling to compress the spatial dimensions of each channel and generate a scalar to represent the global feature response of the channel. In the excitation phase, the SE attention block learns the weight of each channel through two fully connected layers and calculates the attention coefficient of each channel. Finally, the learned weights are weighted channel by channel with the input feature map to enhance attention to key semantic information while suppressing irrelevant features.

[0068] With the help of SE attention mechanism, corresponding importance weights are assigned to each feature group in order to more accurately strengthen the focus on key features; to enhance the synergy between feature groups, the enhanced features of the previous group are fused with the initial features of the next group respectively, so that the model can better utilize the information in the previous group of features and further improve the expression ability of the next group of features, so as to achieve effective collaborative interaction between feature groups and capture richer semantic information.

[0069] 5) The SBSDet model proposed in the present invention is implemented using the PyTorch framework and trained and tested on a single NVIDIA 4090D GPU. During the algorithm implementation process, multiple key libraries are used, including Numpy for numerical calculations, OpenCV for image processing, and Ultralytics for building a target detection model. During network training, the Adam optimizer with an initial learning rate of 0.001 is used to ensure efficient convergence speed and stability. The training was carried out for a total of 100 epochs, and each epoch was trained using data with a batch size of 32. The experimental parameter settings are shown in Table 1:

[0070] Table 1

[0071]

[0072] 6) The advantages of the present invention are: 1. The acquisition of aerial images is achieved through the combination of drones and image processing technology, which not only greatly improves the degree of automation of the monitoring system, but also enhances the timeliness and spatial coverage of monitoring; drones can quickly cover a large range of areas, especially some difficult-to-reach geographical locations or sudden combustion events in a short period of time, thereby effectively making up for the shortcomings of traditional monitoring methods; 2. In practical applications, adaptive spatial perception technology can significantly enhance the ability to identify smoke in complex environments and improve the system's detection effect on smoke in different scenarios; 3. The multi-group attention mechanism can enhance the model's ability to represent semantic information, thereby improving sensitivity to subtle differences. In practical applications, it can effectively improve recognition stability and reliability in changing environments.

[0073] Of course, the present invention is not limited to the above-mentioned embodiments, and those skilled in the art may make equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications and substitutions are all included in the scope defined by the claims of this application.

Claims

1. A straw burning smoke detection algorithm based on adaptive spatial perception, characterized in that: include: S1, obtain aerial images of the target area, create a smoke image dataset, and divide the dataset into a training set, a validation set, and a test set; S2, constructing a smoke detection network including a feature extraction network, a target recognition network and a post-processing network, wherein step S2 includes the following steps: S2.1, the feature extraction network embeds an adaptive spatial perception module to extract smoke features; S2.2, the target recognition network includes a weighted bidirectional feature pyramid network module, a multi-group collaborative attention module and a detection head to identify smoke and straw in the image; S2.3, the post-processing network calculates the relative position relationship between the smoke and the straw; S3, training the smoke detection network multiple times using the training set and the validation set to obtain a smoke detection model; S4: Use the test set to test the smoke detection model and verify the performance of the smoke detection model.

2. According to claim 1, a straw burning smoke detection algorithm based on adaptive spatial perception is characterized in that: In step S2.1, the calculation process of the adaptive spatial perception module is expressed by the following formulas (1)-(5): w=Softmax(FC(ReLU(FC(AvgPool(x))))) (2) B=ReLU(BN(Conv(M))) (4) Among them, y i is the output of the i-th convolution kernel; w i is the weight of the i-th convolution kernel. The weight generation process consists of average pooling, two fully connected layers, ReLU layer and Softmax layer. AvgPool represents average pooling and FC represents fully connected layer. M is the aggregate feature. The spatial data enhancement process consists of average pooling, maximum pooling, channel concatenation, convolution and Sigmoid layer. MaxPool represents maximum pooling, Concat represents channel concatenation, Conv represents convolution and σ represents Sigmoid. ReLU and Sigmoid activation functions are expressed by formulas (6) and (7) respectively:

3. According to claim 1, a straw burning smoke detection algorithm based on adaptive spatial perception is characterized in that ,In step S2.2, the calculation process of the multi-group collaborative attention module is expressed by the following formulas (8)-(10): S1=SE(Conv(X1)) (8) S i =SE(Conv(Add(S i-1 ,X i )) (9) G=Concat(S1,S2,...,S K ) (10) Among them, for a given feature X∈R C×H×W , C, H and W represent the number of channels, height and width of the feature respectively. The feature X is divided into K groups along the channel direction, X = [X1, X2, ..., X K ]; SE means compression and stimulation of attention; Add means feature addition; G means semantically enhanced features.

4. The straw burning smoke detection algorithm based on adaptive spatial perception according to claim 1 is characterized in that: The division ratio of the training set, validation set and test set is 8:1:1.

Images