Infrared image gas intelligent detection method based on lightweight YOLO11 improvement

Abstract

The invention belongs to the field of gas leakage detection, and discloses an infrared image gas intelligent detection method based on lightweight YOLO11 improvement, which comprises the following steps: firstly, acquiring infrared gas leakage data required to be detected on site by using an infrared thermal imager, and obtaining an infrared gas image through pseudo-color conversion and high-sensitivity processing; then, manually marking a gas detection frame, and making an infrared gas leakage detection data set; the method comprises the following steps: acquiring a data set, training a deep learning network improved based on lightweight YOLO11 by using the data set, introducing a gas shape sensing module into the network, fully extracting gas features with complex shapes, and finally realizing intelligent detection of gas leakage in an infrared image by using a trained deep learning model and predicting the position of gas leakage. According to the invention, the infrared gas area with changeable and complex shapes is adaptively and efficiently extracted through the gas shape sensing module, and the prediction result is beneficial to timely finding and positioning the gas leakage phenomenon and area and rapidly taking treatment measures.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas leakage detection, and in particular relates to an infrared image gas intelligent detection method based on an improved lightweight YOLO11. Background Art

[0002] Early detection of gas leaks is crucial for industrial safety, especially in the petrochemical and natural gas transportation sectors. Prompt detection of gas leaks can effectively prevent accidents. Traditional gas leak detection methods rely on hardware sensors. While these sensors can provide relatively accurate data, they are expensive to deploy, have slow response times, and struggle to adapt to complex environments. With the development of infrared optical gas imaging (IR OGI), infrared image-based gas detection methods have emerged as an efficient, real-time alternative.

[0003] While existing infrared image gas detection methods based on convolutional neural networks (CNNs) have improved detection accuracy to a certain extent, the fixed sampling method of traditional convolution operations is difficult to adapt to the diversity and complexity of gas leak shapes, resulting in a significant decrease in the model's detection accuracy when the gas shape changes. Furthermore, while existing deformable convolution methods can improve the model's adaptability to gas shapes, their high computational complexity makes it difficult to meet real-time requirements. Therefore, improving the detection accuracy of gases in infrared images while maintaining detection efficiency remains a major challenge facing current gas detection technology. Summary of the Invention

[0004] The purpose of the present invention is to provide an improved infrared image gas intelligent detection method based on lightweight YOLO11 to solve the above technical problems.

[0005] To solve the above technical problems, the specific technical solution of the present invention is an infrared image gas intelligent detection method based on the improved lightweight YOLO11 as follows:

[0006] A method for intelligent gas detection in infrared images based on lightweight deformable convolution includes the following steps:

[0007] Step 1: Obtain infrared data: Use infrared thermal imaging equipment to obtain infrared data of the target area to be detected;

[0008] Step 2: Pseudo-color image conversion: grayscale pseudo-color mapping is performed on the collected infrared data to convert temperature information into color information to obtain an infrared pseudo-color image;

[0009] Step 3: High sensitivity processing: High sensitivity processing is performed on the infrared pseudo-color image after pseudo-color mapping. First, the adaptive histogram equalization method is applied, and then the image is sharpened.

[0010] Step 4: Construct an infrared gas image dataset: Artificially mark the gas area positions of the images after high-sensitivity processing and divide them into training and validation sets to construct an infrared gas image dataset;

[0011] Step 5: Apply a deep learning network based on the improved lightweight YOLO11 to train the divided infrared gas image dataset. The improved model introduces a gas shape perception module and achieves sufficient feature extraction of gases with different shapes by dynamically adjusting the sampling position of the feature extraction convolution kernel.

[0012] Step 6: Apply the trained model, input the infrared image to be detected, obtain the infrared image gas intelligent detection result, and output the prediction results of whether gas exists and the location of the gas area.

[0013] Furthermore, the step 1 includes the following steps:

[0014] The size of each frame of the image collected is 320×240 pixels, the color depth of the image is 16 bits, and each pixel value represents the temperature of the corresponding position.

[0015] Furthermore, the step 2 includes the following steps:

[0016] The 16-bit pixel value representing the temperature of the acquired image is mapped to the grayscale range of 0 to 255 through normalization. The normalization process uses the following formula:

[0017]

[0018] Among them, original_value represents the pixel value of the original image, min_value and max_value are the minimum and maximum values ​​in the image respectively. After normalization, the pixel values ​​of the image are converted to the range between 0 and 255. Subsequently, the normalized grayscale image is converted into an RGB three-channel image through linear mapping. For each normalized grayscale pixel value, a corresponding RGB image is generated directly by using it as the value of the RGB three channels.

[0019] Furthermore, the step 3 includes the following steps:

[0020] The adaptive histogram equalization method used divides the image into 8×8 blocks, performs histogram equalization on each block, and adjusts the contrast of the local area; then the image is sharpened and convolved with a 3×3 Laplacian operator to obtain the edge information of the image. Then, the sharpened image is weightedly superimposed with the original image to enhance the edge clarity of the gas leakage area in the image.

[0021] Furthermore, step 4 includes the following steps:

[0022] The manually labeled rectangular detection box of the gas area contains the center coordinate point (X c ,Y c ), as well as width and height (W, H), the training set and test set are randomly divided in a ratio of 4:1.

[0023] Furthermore, the step 5 includes the following steps:

[0024] The main body of the deep learning network improved based on lightweight YOLO11 is a lightweight model of the single-stage target detection YOLO11 framework. Its network depth value is 0.5, width value is 0.25, and the maximum number of channels is 1024. It consists of three parts: feature extraction network, feature aggregation network and lightweight detection head. The main body of the feature extraction network consists of a 3×3 convolutional layer and a gas shape perception module. There are four layers of feature extraction modules, and the last three layers of extracted features are output to the feature aggregation network; the feature aggregation network realizes the aggregation of multi-scale features through feature upsampling and downsampling, and further extracts gas shape perception features, and finally outputs the features to the lightweight convolution head; the lightweight convolution head predicts the results based on multi-scale features, and the prediction results remove redundant predictions through non-maximum suppression to obtain the final gas detection prediction results.

[0025] Furthermore, the main body of the gas shape perception module is composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. The input features will be extracted through these two parallel convolutions respectively, and then the output features are obtained through feature aggregation. The lightweight linear deformable convolution is a convolution with a shape of 1×5 or 5×1 and has linear dynamic constraints. The displacement of the upper and lower or left and right positions of the convolution kernel is limited to a unit displacement relative to the position of the adjacent convolution kernel. Assuming that the center position of the 5×1 linear convolution kernel is K3, its coordinates are (X3, Y3), and the remaining positions are K1, K2, K4, and K5 from top to bottom, then the coordinates of K2 and K4 are expressed as:

[0026] (X2, Y2) = (X3 ± Δd, Y3 - 1)

[0027] (X4, Y4) = (X3 ± Δd, Y3 + 1)

[0028] Among them, Δd = 1, which represents the unit displacement. The linear convolution kernel can adaptively adjust the linear shape of the convolution kernel according to the target shape. Similarly, the coordinates of K1 and K5 are expressed as:

[0029] (X1, Y1) = (X2 ± Δd, Y2 - 1)

[0030] (X5, Y5) = (X4 ± Δd, Y4 + 1)

[0031] Similarly, for a 1×5 linear convolution kernel, its convolution kernel coordinates can be expressed as K1, K2, K4, and K5 from left to right based on the center position K3:

[0032] (X1, Y1) = (X2-1, Y2±Δd)

[0033] (X2, Y2) = (X3-1, Y3±Δd)

[0034] (X4, Y4) = (X3 + 1, Y3 ± Δd)

[0035] (X5, Y5) = (X4 + 1, Y4 ± Δd)

[0036] Each lightweight linear deformable convolution also comes with a combined convolution of the same shape, which is used to learn the deformation displacement of its convolution kernel. The gas shape perception module also has a multi-layer feature extraction bottleneck design, which can extract multi-scale features and fuse them to fully extract gas features of complex shapes.

[0037] Furthermore, the input image size of the deep learning network improved based on lightweight YOLO11 in step 5 is 256×256 pixels, the training iteration batch is 200 times, the optimizer uses AdamW optimizer, the learning rate adaptive adjustment class is StepLR, the learning rate is 0.001, and the training batch size is 16.

[0038] Furthermore, the loss function used in the training of the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the positioning box loss function and the classification loss function. The positioning box loss function uses the CIoU loss function, and the calculation formula is as follows:

[0039]

[0040] Among them, B, w, h, b are the detection box predicted by the network, the width of the predicted detection box, the height of the predicted detection box and the coordinates of the center point of the predicted detection box, respectively. gt ,w gt ,h gt ,b gt is the true detection box, the width of the true detection box, the height of the true detection box and the coordinates of the center point of the true detection box. IoU represents the intersection-over-union ratio of two boxes. ρ(.) is the Euclidean distance between two points. c represents the diagonal distance of the minimum closed box that is compatible with both the predicted and true detection boxes. α and v measure the similarity between the predicted and true detection boxes. The classification loss function uses the Slide loss function, which is calculated as follows:

[0041]

[0042] Among them, μ is the IoU threshold for judging whether a sample is a positive sample, which is set to 0.5, y and y ′ Represent the true category and the category predicted by the network respectively. The final loss function calculation formula is as follows:

[0043]

[0044] The infrared image gas intelligent detection method based on the improved lightweight YOLO11 of the present invention has the following advantages:

[0045] The present invention first uses an infrared thermal imager to collect infrared gas leak data to be detected on-site. High-quality infrared gas images of the target gas are obtained through pseudo-color conversion and high-sensitivity processing. Subsequently, gas detection frames are manually annotated to create an infrared gas image dataset. This dataset is then used to train a deep learning network based on a modified lightweight YOLO11 algorithm. This network introduces a gas shape perception module to improve gas detection accuracy. Finally, the trained deep learning model is used to intelligently detect gas leaks in infrared images and predict their locations. This invention offers the advantages of low computational cost and high precision. The gas shape perception module adaptively and efficiently extracts infrared gas regions with complex and variable shapes. The prediction results help promptly detect gas leaks, locate the leak area, and quickly implement remedial measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of a method in an embodiment of the present invention.

[0047] Figure 2 1 is a model framework diagram of the improved YOLO11 in an embodiment of the present invention.

[0048] Figure 3 4 is an execution flow chart of the lightweight linear deformable convolution module in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of the infrared image gas intelligent detection method based on the improved lightweight YOLO11 in conjunction with the accompanying drawings.

[0050] The present invention provides an improved infrared image gas intelligent detection method based on lightweight YOLO11, comprising the following steps:

[0051] Step 1: Obtain infrared data: Use infrared thermal imaging equipment to obtain infrared data of the target area to be detected;

[0052] The size of each frame of the image collected is 320×240 pixels, the color depth of the image is 16 bits, and each pixel value represents the temperature of the corresponding position.

[0053] Step 2: Pseudo-color image conversion: grayscale pseudo-color mapping is performed on the collected infrared data to convert temperature information into color information to obtain an infrared pseudo-color image;

[0054] The 16-bit pixel value representing the temperature of the acquired image is mapped to the grayscale range of 0 to 255 through normalization. The normalization process uses the following formula:

[0055]

[0056] Where original_value represents the pixel value of the original image, min_value and max_value represent the minimum and maximum values ​​in the image, respectively. After normalization, the pixel values ​​of the image are converted to a range between 0 and 255. Subsequently, the present invention converts the normalized grayscale image into an RGB three-channel image through linear mapping. Specifically, for each normalized grayscale pixel value, a corresponding RGB image is generated by directly using it as the value of the RGB three channels.

[0057] Step 3: High sensitivity processing: High sensitivity processing is performed on the infrared pseudo-color image after pseudo-color mapping. First, the adaptive histogram equalization method is applied, and then the image is sharpened.

[0058] The adaptive histogram equalization method used divides the image into 8×8 blocks, performs histogram equalization on each block, and adjusts the contrast of the local area; then the image is sharpened and convolved with a 3×3 Laplacian operator to obtain the edge information of the image. Then, the sharpened image is weightedly superimposed with the original image to enhance the edge clarity of the gas leakage area in the image.

[0059] Step 4: Construct an infrared gas image dataset: Artificially mark the gas area positions of the images after high-sensitivity processing and divide them into training and validation sets to construct an infrared gas image dataset;

[0060] The manually labeled rectangular detection box of the gas area contains the center coordinate point (X c ,Y c ), as well as width and height (W, H), the training set and test set are randomly divided in a ratio of 4:1.

[0061] Step 5: Apply a deep learning network based on the improved lightweight YOLO11 to train the divided infrared gas image dataset. This improved model introduces a gas shape perception module and achieves sufficient feature extraction of gases with different shapes by dynamically adjusting the sampling position of the feature extraction convolution kernel.

[0062] The main body of the deep learning network improved based on lightweight YOLO11 is a lightweight model of the single-stage target detection YOLO11 framework. Its network depth value is 0.5, width value is 0.25, and the maximum number of channels is 1024. It consists of three parts: feature extraction network, feature aggregation network and lightweight detection head. Specifically, its feature extraction network body consists of 3×3 convolutional layer and gas shape perception module. There are four layers of feature extraction modules, and the last three layers of extracted features are output to the feature aggregation network; the feature aggregation network realizes the aggregation of multi-scale features through feature upsampling and downsampling, and further extracts gas shape perception features, and finally outputs the features to the lightweight convolution head; the lightweight convolution head predicts the results based on multi-scale features, and the prediction results remove redundant predictions through non-maximum suppression to obtain the final gas detection prediction results.

[0063] The main body of the gas shape perception module is composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. The input features will be extracted through these two parallel convolutions, and then the output features will be obtained through feature aggregation. Specifically, the lightweight linear deformable convolution is a convolution with a shape of 1×5 or 5×1 and has linear dynamic constraints. The displacement of the upper and lower (or left and right) positions of the convolution kernel is limited to a unit displacement relative to the position of the adjacent convolution kernel. Assuming that the center position of the 5×1 linear convolution kernel is K3, its coordinates are (X3, Y3), and the remaining positions are K1, K2, K4, and K5 from top to bottom, then the coordinates of K2 and K4 can be expressed as:

[0064] (X2, Y2) = (X3 ± Δd, Y3 - 1)

[0065] (X4, Y4) = (X3 ± Δd, Y3 + 1)

[0066] Among them, Δd = 1, which represents the unit displacement. The linear convolution kernel can adaptively adjust the linear shape of the convolution kernel according to the target shape. Similarly, the coordinates of K1 and K5 can be expressed as:

[0067] (X1, Y1) = (X2 ± Δd, Y2 - 1)

[0068] (X5, Y5) = (X4 ± Δd, Y4 + 1)

[0069] Similarly, for a 1×5 linear convolution kernel, its convolution kernel coordinates can be expressed as K1, K2, K4, and K5 from left to right based on the center position K3:

[0070] (X1, Y1) = (X2-1, Y2±Δd)

[0071] (X2, Y2) = (X3-1, Y3±Δd)

[0072] (X4, Y4) = (X3 + 1, Y3 ± Δd)

[0073] (X5, Y5) = (X4 + 1, Y4 ± Δd)

[0074] Each lightweight linear deformable convolution also comes with a combined convolution of the same shape, which is used to learn the deformation displacement of its convolution kernel. The gas shape perception module also has a multi-layer feature extraction bottleneck design, which can extract multi-scale features and fuse them to fully extract gas features of complex shapes.

[0075] The input image size of the deep learning network improved based on lightweight YOLO11 is 256×256 pixels, the training iteration batch is 200 times, the optimizer uses AdamW optimizer, the learning rate adaptive adjustment class is StepLR, the learning rate is 0.001, and the training batch size is 16.

[0076] The loss function used in the training of the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the positioning box loss function and the classification loss function. The positioning box loss function uses the CIoU loss function, and the calculation formula is as follows:

[0077]

[0078] Among them, B, w, h, b are the detection box predicted by the network, the width of the predicted detection box, the height of the predicted detection box and the coordinates of the center point of the predicted detection box, respectively. gt ,w gt ,h gt ,b gt is the true detection box, the width of the true detection box, the height of the true detection box, and the coordinates of the center point of the true detection box. IoU represents the intersection over union ratio of two boxes. ρ(.) is the Euclidean distance between two points. c represents the diagonal distance of the smallest closed box that is compatible with both the predicted and true detection boxes. α and v measure the similarity between the predicted and true detection boxes. The classification loss function uses the Slide loss function, which is calculated as follows:

[0079]

[0080] Among them, μ is the IoU threshold for judging whether a sample is a positive sample, which is set to 0.5, y and y ′ Represent the true category and the category predicted by the network respectively. The final loss function calculation formula is as follows:

[0081]

[0082] Step 6: Apply the trained model, input the infrared image to be detected, obtain the infrared image gas intelligent detection result, and output the prediction results of whether gas exists and the location of the gas area.

[0083] Example:

[0084] like Figure 1 As shown in the figure, the infrared image gas intelligent detection method based on the improved lightweight YOLO11 includes the following steps:

[0085] Step 1: Obtain infrared data; use infrared thermal imaging equipment to collect the infrared data required for detection on site;

[0086] Step 2: Pseudo-color image conversion: Perform pseudo-color conversion on the collected data to obtain an infrared pseudo-color image reflecting temperature information;

[0087] Step 3: High-sensitivity processing: Apply adaptive histogram equalization and sharpening processing to the infrared pseudo-color image to further highlight the gas area in the infrared pseudo-color image;

[0088] Step 4: Construct an infrared gas image dataset; manually annotate the gas area in the infrared image in the form of a rectangular box containing the center coordinate point (X c ,Y c ), as well as width and height (W, H), the training set and test set are randomly divided in a ratio of 4:1;

[0089] Step 5: Train an infrared image gas intelligent detection model based on the improved lightweight YOLO11. The image detection model uses a lightweight improved YOLO11 model with a network depth of 0.5, a width of 0.25, and a maximum number of channels of 1024. The improved YOLO11 model is trained using a self-built infrared gas detection dataset. The improved model introduces a gas shape perception module to enhance the learning ability of irregular gas areas. The training iteration batch is 200 times, the optimizer uses the AdamW optimizer, the learning rate adaptive adjustment class is StepLR, the learning rate is 0.001, and the training batch size is 16. The training loss function of this model is a hybrid loss function of the positioning box loss function and the classification loss function, using the CIoU loss function and the Slide loss function respectively.

[0090] like Figure 2As shown in the figure, the improved YOLO11 model consists of three modules: feature extraction module, feature fusion module and lightweight multi-scale detection head. The feature extraction module uses 3×3 convolution layer and gas shape perception operation to perform multi-scale feature extraction on the input image to be detected to obtain a multi-scale feature map. The feature fusion module obtains the multi-scale feature map and performs fusion feature extraction. The fused features are input into the lightweight multi-scale detection head. The multi-scale detection head outputs the predicted detection box at each scale and obtains the final infrared gas predicted detection box through non-maximum suppression.

[0091] like Figure 3 As shown in the figure, the improved YOLO11 model introduces a gas shape perception module to replace the traditional feature extraction layer. The module is predefined as two parallel 1×5 and 5×1 linear convolution kernels. The offset of the linear convolution kernel is learned by setting an additional convolution layer to adapt to the shape of the infrared gas, enhance the feature perception capability of complex and changeable gases, and improve the accuracy of gas recognition and positioning.

[0092] Step 6: Use the trained infrared gas image detection model to detect the gas area in the image to be detected and obtain the detection frame;

[0093] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A method for intelligent gas detection in infrared images based on lightweight deformable convolution, characterized in that: The following steps are involved: Step 1: Obtain infrared data: Use infrared thermal imaging equipment to obtain infrared data of the target area to be detected; Step 2: Pseudo-color image conversion: grayscale pseudo-color mapping is performed on the collected infrared data to convert temperature information into color information to obtain an infrared pseudo-color image; Step 3: High-sensitivity processing: High-sensitivity processing is performed on the infrared pseudo-color image after pseudo-color mapping. First, the adaptive histogram equalization method is applied, and then the image is sharpened. Step 4: Construct an infrared gas image dataset: Artificially mark the gas area positions of the images after high-sensitivity processing and divide them into training and validation sets to construct an infrared gas image dataset; Step 5: Apply a deep learning network based on the improved lightweight YOLO11 to train the divided infrared gas image dataset. The improved model introduces a gas shape perception module and achieves sufficient feature extraction of gases with different shapes by dynamically adjusting the sampling position of the feature extraction convolution kernel. Step 6: Apply the trained model, input the infrared image to be detected, obtain the infrared image gas intelligent detection result, and output the prediction results of whether gas exists and the location of the gas area.

2. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1 is characterized in that: The step 1 comprises the following steps: The size of each frame of the image collected is 320×240 pixels, the color depth of the image is 16 bits, and each pixel value represents the temperature of the corresponding position.

3. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1 is characterized in that: The step 2 comprises the following steps: The 16-bit pixel value representing the temperature of the acquired image is mapped to the grayscale range of 0 to 255 through normalization. The normalization process uses the following formula: Among them, original_value represents the pixel value of the original image, min_value and max_value are the minimum and maximum values ​​in the image respectively. After normalization, the pixel values ​​of the image are converted to the range between 0 and 255. Subsequently, the normalized grayscale image is converted into an RGB three-channel image through linear mapping. For each normalized grayscale pixel value, a corresponding RGB image is generated directly by using it as the value of the RGB three channels.

4. The method for intelligent gas detection in infrared images based on lightweight deformable convolution according to claim 1 is characterized in that: The step 3 comprises the following steps: The adaptive histogram equalization method used divides the image into 8×8 blocks, performs histogram equalization on each block, and adjusts the contrast of the local area; then the image is sharpened and convolved with a 3×3 Laplacian operator to obtain the edge information of the image. Then, the sharpened image is weightedly superimposed with the original image to enhance the edge clarity of the gas leakage area in the image.

5. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1 is characterized in that: The step 4 comprises the following steps: The manually labeled rectangular detection box of the gas area contains the center coordinate point (X c ,Y c ), as well as width and height (W, H), the training set and test set are randomly divided in a ratio of 4:

1.

6. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1 is characterized in that: The step 5 comprises the following steps: The main body of the deep learning network improved based on the lightweight YOLO11 is a lightweight model of the single-stage target detection YOLO11 framework. Its network depth value is 0.5, width value is 0.25, and the maximum number of channels is 1024. It consists of three parts: feature extraction network, feature aggregation network and lightweight detection head. The main body of the feature extraction network consists of a 3×3 convolutional layer and a gas shape perception module. There are four layers of feature extraction modules, and the extracted features of the last three layers are output to the feature aggregation network. The feature aggregation network aggregates multi-scale features through feature upsampling and downsampling, further extracts gas shape perception features, and finally outputs the features to the lightweight convolution head; the lightweight convolution head predicts the results based on the multi-scale features, and the prediction results remove redundant predictions through non-maximum suppression to obtain the final gas detection prediction results.

7. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 6 is characterized in that: The main body of the gas shape perception module is composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. The input features will be extracted through these two parallel convolutions respectively, and then the output features are obtained through feature aggregation. The lightweight linear deformable convolution is a convolution with a shape of 1×5 or 5×1 and has linear dynamic constraints. The displacement of the upper and lower or left and right positions of the convolution kernel is limited to a unit displacement relative to the position of the adjacent convolution kernel. Assuming that the center position of the 5×1 linear convolution kernel is K3, its coordinates are (X3, Y3), and the remaining positions are K1, K2, K4, and K5 from top to bottom, then the coordinates of K2 and K4 are expressed as: (X2, Y2) = (X3 ± Δd, Y3 - 1) (X4, Y4) = (X3 ± Δd, Y3 + 1) Among them, Δd = 1, which represents the unit displacement. The linear convolution kernel can adaptively adjust the linear shape of the convolution kernel according to the target shape. Similarly, the coordinates of K1 and K5 are expressed as: (X1, Y1) = (X2 ± Δd, Y2 - 1) (X5, Y5) = (X4 ± Δd, Y4 + 1) Similarly, for the 1×5 linear convolution kernel, the convolution kernel coordinates can be K1, K2 from left to right according to the center position K3. K4, K5 are represented as follows: (X1, Y1) = (X2-1, Y2±Δd) (X2, Y2) = (X3-1, Y3±Δd) (X4, Y4) = (X3 + 1, Y3 ± Δd) (X5, Y5) = (X4 + 1, Y4 ± Δd) Each lightweight linear deformable convolution also comes with a combined convolution of the same shape, which is used to learn the deformation displacement of its convolution kernel. The gas shape perception module also has a multi-layer feature extraction bottleneck design, which can extract multi-scale features and fuse them to fully extract gas features of complex shapes.

8. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 6 is characterized in that: The input image size of the deep learning network improved based on lightweight YOLO11 in step 5 is 256×256 pixels, the training iteration batch is 200 times, the optimizer uses AdamW optimizer, the learning rate adaptive adjustment class is StepLR, the learning rate is 0.001, and the training batch size is 16.

9. The method for intelligent gas detection in infrared images based on lightweight deformable convolution according to claim 6, characterized in that: The loss function used in the training of the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the positioning box loss function and the classification loss function. The positioning box loss function uses the CIoU loss function, and the calculation formula is as follows: Among them, B, w, h, b are the detection box predicted by the network, the width of the predicted detection box, the height of the predicted detection box and the coordinates of the center point of the predicted detection box, respectively. gt ,w gt ,h gt ,b gt is the true detection box, the width of the true detection box, the height of the true detection box and the coordinates of the center point of the true detection box. IoU represents the intersection-over-union ratio of two boxes. ρ(.) is the Euclidean distance between two points. c represents the diagonal distance of the minimum closed box that is compatible with both the predicted and true detection boxes. α and v measure the similarity between the predicted and true detection boxes. The classification loss function uses the Slide loss function, which is calculated as follows: Among them, μ is the IoU threshold for judging whether a sample is a positive sample, which is set to 0.5, y and y ′ Represent the true category and the category predicted by the network respectively. The final loss function calculation formula is as follows:

Images