An infrared image gas intelligent detection method based on lightweight YOLO11 improvement

Abstract

The application belongs to the field of gas leakage detection, and discloses an infrared image gas intelligent detection method based on improved lightweight YOLO11, which comprises the following steps: first, using an infrared thermal imager to collect infrared gas leakage data required for detection on site, and obtaining an infrared gas image through pseudo-color conversion and high-sensitivity processing; then, marking a gas detection frame by hand to make an infrared gas leakage detection dataset; and then, training a deep learning network based on improved lightweight YOLO11 by using the dataset, wherein the network introduces a gas shape perception module, which can fully extract gas features of complex shapes, and finally, realizing intelligent detection of gas leakage in an infrared image by using the trained deep learning model and predicting the position of gas leakage. The gas shape perception module can adaptively and efficiently extract infrared gas regions with variable and complex shapes, and the prediction result is helpful for timely discovering and positioning gas leakage phenomena and regions and taking quick processing measures.

Description

Technical Field

[0001] This invention belongs to the field of gas leak detection technology, and in particular relates to an intelligent gas detection method based on a lightweight YOLO11 improved infrared image. Background Technology

[0002] Early detection of gas leaks is crucial for industrial safety, especially in the petrochemical and natural gas transportation sectors, where timely detection can effectively prevent accidents. Traditional gas leak detection methods rely heavily on hardware sensors, which, while providing relatively accurate data, are costly to deploy, have slow response times, and are difficult to adapt to complex environments. With the development of infrared optical gas imaging (IR OGI) technology, infrared image-based gas detection methods are gradually becoming a more efficient and real-time alternative.

[0003] Existing infrared image gas detection methods based on convolutional neural networks (CNNs) have improved detection accuracy to some extent. However, the fixed sampling method of traditional convolution operations is difficult to adapt to the diversity and complexity of gas leakage patterns, resulting in a significant drop in 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 them difficult to meet real-time requirements. Therefore, how to improve the detection accuracy of gases in infrared images while maintaining detection efficiency remains a significant challenge for current gas detection technology. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent gas detection method based on a lightweight YOLO11-based improved infrared image, in order to solve the aforementioned technical problems.

[0005] To address the aforementioned technical problems, the specific technical solution of the present invention, a method for intelligent gas detection in infrared images based on a lightweight YOLO11 improvement, is as follows:

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

[0007] Step 1: Acquire Infrared Data: Use an infrared thermal imaging device to acquire infrared data of the target area to be detected;

[0008] Step 2: Pseudocolor image conversion: Perform grayscale pseudocolor mapping on the acquired infrared data to convert temperature information into color information and obtain an infrared pseudocolor image;

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

[0010] Step 4: Construct an infrared gas image dataset: Mark the locations of artificial gas regions in the images after high-sensitivity processing, and divide them into training and validation sets to construct an infrared gas image dataset;

[0011] Step 5: Train the divided infrared gas image dataset using a lightweight YOLO11-based improved deep learning network. The improved model introduces a gas shape perception module, which dynamically adjusts the sampling position of the feature extraction convolution kernel to achieve full feature extraction of gases of different shapes.

[0012] Step 6: Apply the trained model by inputting the infrared image to be detected, obtaining the intelligent gas detection result of the infrared image, and outputting the prediction result of whether gas exists and the location of the gas region.

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

[0014] Each captured image frame is 320×240 pixels in size, with a color depth of 16 bits, and each pixel value represents the temperature at the corresponding location.

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

[0016] The 16-bit pixel values ​​representing temperature in the acquired image are normalized and mapped to a grayscale range of 0 to 255. The normalization process uses the following formula:

[0017]

[0018] Wherein, 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 standardization, the pixel values ​​of the image are converted to the range between 0 and 255. Subsequently, the standardized grayscale image is converted into an RGB three-channel image through linear mapping. For each standardized grayscale pixel value, a corresponding RGB image is generated directly by using it as the value of the three RGB channels.

[0019] Furthermore, 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 by convolving the image with a 3×3 Laplacian operator to obtain the edge information of the image. Finally, the sharpened image is weighted and superimposed with the original image to enhance the edge clarity of the gas leak area in the image.

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

[0022] The manually labeled rectangular detection frame for the gas area includes the center coordinates (X) of the rectangular detection frame. c ,Y c The training set and the test set are randomly divided in a 4:1 ratio, with width and height (W, H).

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

[0024] The lightweight YOLO11-based deep learning network is a lightweight model of the single-stage object detection YOLO11 framework. Its network depth is 0.5, width is 0.25, and maximum number of channels is 1024. It consists of three parts: a feature extraction network, a feature aggregation network, and a lightweight detection head. The feature extraction network consists of four layers: 3×3 convolutional layers and a gas shape perception module. The features extracted from the last three layers are fed into the feature aggregation network. The feature aggregation network aggregates multi-scale features through feature upsampling and downsampling, further extracting gas shape perception features, and finally outputting these features to the lightweight convolutional head. The lightweight convolutional head predicts the results based on the multi-scale features. The prediction results are then processed by non-maximum suppression to remove redundant predictions, yielding the final gas detection prediction result.

[0025] Furthermore, the gas shape sensing module is mainly composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. The input features are extracted through these two parallel convolutions, and then the output features are obtained through feature aggregation. The lightweight linear deformable convolutions are 1×5 or 5×1 convolutions with linear dynamic constraints. The displacement of the convolution kernel's vertical or horizontal position is restricted to a unit displacement relative to the position of the adjacent convolution kernel. Assuming the center position of the 5×1 linear convolution kernel is K3, with coordinates (X3, Y3), and the remaining positions from top to bottom are K1, K2, K4, and K5, then the coordinates of K2 and K4 are expressed as follows:

[0026] (X2,Y2)=(X3±Δd,Y3-1)

[0027] (X4,Y4)=(X3±Δd,Y3+1)

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

[0029] (X1,Y1)=(X2±Δd,Y2-1)

[0030] (X5,Y5)=(X4±Δd,Y4+1)

[0031] Similarly, for a 1×5 linear convolution kernel, its kernel coordinates can be represented from left to right as K1, K2, K4, and K5, 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 to learn the deformation displacement of its convolution kernel. The gas shape perception module also features a multi-layer feature extraction bottleneck design, which can extract and fuse multi-scale features to achieve full extraction of gas features with complex shapes.

[0037] Furthermore, in step 5, the input image size of the deep learning network based on the lightweight YOLO11 improvement is 256×256 pixels, the training batch size is 200, the optimizer is 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 for training the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the localization loss function and the classification loss function. The localization loss function uses the CIoU loss function, and the calculation formula is as follows:

[0039]

[0040] Where B, w, h, and b represent the predicted detection bounding box, the width of the predicted detection bounding box, the height of the predicted detection bounding box, and the coordinates of the center point of the predicted detection bounding box, respectively. gt ,w gt ,h gt ,b gt Here, represents the ground truth bounding box, its width, height, and center coordinates. IoU represents the intersection-over-union ratio of the two boxes. ρ(.) calculates the Euclidean distance between the two points. c represents the diagonal distance of the smallest closed box that is compatible with both the predicted and ground truth bounding boxes. α and v measure the similarity between the predicted and ground truth bounding boxes. The classification loss function uses the Slide loss function, calculated as follows:

[0041]

[0042] Where μ is the IoU threshold for determining whether a sample is a positive sample, set to 0.5, and y and y ′ Let represent the true category and the category predicted by the network, respectively. The final loss function is calculated as follows:

[0043]

[0044] The intelligent gas detection method based on lightweight YOLO11 for infrared imaging, as proposed in this invention, has the following advantages:

[0045] This invention first uses an infrared thermal imager to collect infrared gas leak data for detection in the field, and obtains high-quality infrared gas images through pseudo-color conversion and high-sensitivity processing. Then, gas detection boxes are manually annotated to create an infrared gas image dataset. This dataset is then used to train a lightweight YOLO11-based improved deep learning network, which incorporates 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 location. This invention has the advantages of low computational cost and high accuracy. The gas shape perception module adaptively and efficiently extracts infrared gas regions with varied and complex shapes, and the prediction results help to promptly detect gas leaks, locate the leak area, and quickly take appropriate measures. Attached Figure Description

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

[0047] Figure 2 This is a model framework diagram of the improved YOLO11 in this embodiment of the invention.

[0048] Figure 3 This is an execution flowchart of the lightweight linear deformable convolution module in an embodiment of the present invention. Detailed Implementation

[0049] To better understand the purpose, structure, and function of this invention, the following detailed description, in conjunction with the accompanying drawings, provides an intelligent gas detection method for infrared images based on a lightweight YOLO11 improvement.

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

[0051] Step 1: Acquire Infrared Data: Use an infrared thermal imaging device to acquire infrared data of the target area to be detected;

[0052] Each captured image frame is 320×240 pixels in size, with a color depth of 16 bits, and each pixel value represents the temperature at the corresponding location.

[0053] Step 2: Pseudocolor image conversion: Perform grayscale pseudocolor mapping on the acquired infrared data to convert temperature information into color information and obtain an infrared pseudocolor image;

[0054] The 16-bit pixel values ​​representing temperature from the acquired image are normalized and mapped to a grayscale range of 0 to 255. The normalization process uses the following formula:

[0055]

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

[0057] Step 3: High-sensitivity processing: The infrared pseudo-color image after pseudo-color mapping is processed with high sensitivity. First, an 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 by convolving the image with a 3×3 Laplacian operator to obtain the edge information of the image. Finally, the sharpened image is weighted and superimposed with the original image to enhance the edge clarity of the gas leak area in the image.

[0059] Step 4: Construct an infrared gas image dataset: Mark the locations of artificial gas regions in 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 frame for the gas area includes the center coordinates (X) of the rectangular detection frame. c ,Y c The training set and the test set are randomly divided in a 4:1 ratio, with width and height (W, H).

[0061] Step 5: Train the divided infrared gas image dataset using a lightweight YOLO11-based improved deep learning network. This improved model introduces a gas shape perception module, which dynamically adjusts the sampling position of the feature extraction convolution kernel to achieve full feature extraction of gases of different shapes.

[0062] The lightweight YOLO11-based deep learning network is a lightweight model of the single-stage object detection YOLO11 framework. Its network depth is 0.5, width is 0.25, and maximum number of channels is 1024. It consists of three parts: a feature extraction network, a feature aggregation network, and a lightweight detection head. Specifically, the feature extraction network consists of four layers: 3×3 convolutional layers and a gas shape perception module. The features extracted from the last three layers are fed into the feature aggregation network. The feature aggregation network aggregates multi-scale features through feature upsampling and downsampling, further extracting gas shape perception features, and finally outputting these features to the lightweight convolutional head. The lightweight convolutional head predicts the results based on the multi-scale features. The prediction results are then processed by non-maximum suppression to remove redundant predictions, yielding the final gas detection prediction result.

[0063] The gas shape perception module is mainly composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. Input features are extracted through these two parallel convolutions, and then output features are obtained through feature aggregation. Specifically, the lightweight linear deformable convolution is a 1×5 or 5×1 convolution with linear dynamic constraints. The vertical (or horizontal) displacement of its convolution kernel is restricted to a unit displacement relative to the adjacent convolution kernel position. Assuming the center position of the 5×1 linear convolution kernel is K3 with coordinates (X3, Y3), and the remaining positions from top to bottom are K1, K2, K4, and K5, 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] Where Δd = 1 represents a unit displacement, and the linear convolution kernel can adaptively adjust its linear shape 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 kernel coordinates can be represented from left to right as K1, K2, K4, and K5, 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 to learn the deformation displacement of its convolution kernel. The gas shape perception module also features a multi-layer feature extraction bottleneck design, which can extract and fuse multi-scale features to achieve full extraction of gas features with complex shapes.

[0075] The input image size of the deep learning network based on the lightweight YOLO11 improvement is 256×256 pixels, the training batch size is 200, the optimizer used is 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 for training the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the localization loss function and the classification loss function. The localization loss function uses the CIoU loss function, and the calculation formula is as follows:

[0077]

[0078] Where B, w, h, and b represent the predicted detection bounding box, the width of the predicted detection bounding box, the height of the predicted detection bounding box, and the coordinates of the center point of the predicted detection bounding box, respectively. gt ,w gt ,h gt ,b gt The values ​​represent the ground truth bounding boxes, their width, height, and center coordinates. IoU represents the intersection-over-union ratio of the two boxes. ρ(.) calculates the Euclidean distance between the two points. c represents the diagonal distance of the smallest closed box that is compatible with both the predicted and ground truth bounding boxes. α and v measure the similarity between the predicted and ground truth bounding boxes. The classification loss function uses the Slide loss function, calculated as follows:

[0079]

[0080] Where μ is the IoU threshold for determining whether a sample is a positive sample, set to 0.5, and y and y ′ These represent the true category and the category predicted by the network, respectively. The final loss function is calculated as follows:

[0081]

[0082] Step 6: Apply the trained model by inputting the infrared image to be detected, obtaining the intelligent gas detection result of the infrared image, and outputting the prediction result of whether gas exists and the location of the gas region.

[0083] Example:

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

[0085] Step 1: Acquire infrared data; Use infrared thermal imaging equipment to collect the infrared data to be detected on-site;

[0086] Step 2: Pseudocolor image conversion; Perform pseudocolor conversion on the acquired data to obtain an infrared pseudocolor image that reflects temperature information;

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

[0088] Step 4: Construct an infrared gas image dataset; manually label the gas regions in the infrared images using rectangular boxes, including the center coordinates (X, Y, X). c ,Y c The training set and the test set are randomly divided in a 4:1 ratio, including width and height (W, H).

[0089] Step 5: Train an infrared image gas intelligent detection model based on a lightweight YOLO11 improvement; The image detection model adopts 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. This improved model introduces a gas shape perception module to enhance the learning ability of irregular gas regions. The training batch size is 200 times, the optimizer is AdamW, 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 localization box loss function and the classification loss function, using the CIoU loss function and the Slide loss function respectively.

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

[0091] like Figure 3 As shown, the improved YOLO11 model introduces a gas shape perception module to replace the traditional feature extraction layer. This module is predefined as two parallel 1×5 and 5×1 linear convolution kernels, and learns the offset of the linear convolution kernel by setting an additional convolution layer to adapt to the shape of infrared gas, thereby enhancing the feature perception capability of complex and variable gases and improving the accuracy of gas identification and localization.

[0092] Step 6: Using the trained infrared gas image detection model, gas regions are detected in the image to be detected, and detection boxes are obtained;

[0093] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the 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 within the protection scope of the present invention.

Claims

1. A method for intelligent gas detection in infrared images based on lightweight deformable convolution, characterized in that, Includes the following steps: Step 1: Acquire Infrared Data: Use an infrared thermal imaging device to acquire infrared data of the target area to be detected; Step 2: Pseudocolor image conversion: Perform grayscale pseudocolor mapping on the acquired infrared data to convert temperature information into color information and obtain an infrared pseudocolor image; Step 3: High-sensitivity processing: The infrared pseudo-color image after pseudo-color mapping is processed with high sensitivity. First, an adaptive histogram equalization method is applied, and then the image is sharpened. Step 4: Construct an infrared gas image dataset: Mark the locations of artificial gas regions in the images after high-sensitivity processing, and divide them into training and validation sets to construct an infrared gas image dataset; Step 5: Train the divided infrared gas image dataset using a lightweight YOLO11-based improved deep learning network. The improved model introduces a gas shape perception module, which dynamically adjusts the sampling position of the feature extraction convolution kernel to achieve full feature extraction of gases of different shapes. The main body of the lightweight YOLO11-based deep learning network is a lightweight model of the single-stage object detection YOLO11 framework. Its network depth is 0.5, width is 0.25, and maximum number of channels is 1024. It consists of three parts: a feature extraction network, a feature aggregation network, and a lightweight detection head. The main body of the feature extraction network consists of 3×3 convolutional layers and a gas shape perception module, with a total of four feature extraction modules. The features extracted by the last three layers are then fed into the feature aggregation network. The feature aggregation network aggregates multi-scale features through feature upsampling and downsampling, and further extracts gas shape perception features. Finally, the features are output to a lightweight convolutional head. The lightweight convolutional head performs result prediction based on multi-scale features. The prediction results are filtered to remove redundant predictions through non-maximum suppression, resulting in the final gas detection prediction result. The gas shape sensing module is mainly composed of parallel 1×5 and 5×1 lightweight linear deformable convolutions. Input features are extracted through these two parallel convolutions, and then output features are obtained through feature aggregation. The lightweight linear deformable convolutions are 1×5 or 5×1 in shape and have linear dynamic constraints. The vertical or horizontal displacement of the convolution kernel is restricted to a unit displacement relative to the position of adjacent convolution kernels. Assuming the center position of the 5×1 linear convolution kernel is... Its coordinates are The remaining positions, from top to bottom, are: , , , ,but and The coordinates are represented as: in, This represents a unit displacement. The linear convolution kernel can adaptively adjust its linear shape according to the target shape. Similarly, and The coordinates are represented as: Similarly, for a 1×5 linear convolution kernel, its kernel coordinates can be determined based on its center position. From left to right , , , , respectively represented as: Each lightweight linear deformable convolution also comes with a combined convolution of the same shape to learn the deformation displacement of its convolution kernel. The gas shape perception module also features a multi-layer feature extraction bottleneck design, which can extract and fuse multi-scale features to achieve full extraction of gas features with complex shapes. Step 6: Apply the trained model by inputting the infrared image to be detected, obtaining the intelligent gas detection result of the infrared image, and outputting the prediction result of whether gas exists and the location of the gas region.

2. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1, characterized in that, Step 1 includes the following steps: Each captured image frame is 320×240 pixels in size, with a color depth of 16 bits, and each pixel value represents the temperature at the corresponding location.

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

4. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1, characterized in that, Step 3 includes 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 by convolving the image with a 3×3 Laplacian operator to obtain the edge information of the image. Finally, the sharpened image is weighted and superimposed with the original image to enhance the edge clarity of the gas leak area in the image.

5. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1, characterized in that, Step 4 includes the following steps: The manually labeled rectangular detection frame for the gas area includes the center coordinates of the rectangular detection frame. as well as width and height The training set and the test set are randomly divided in a 4:1 ratio.

6. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1, characterized in that, In step 5, the input image size of the deep learning network based on the lightweight YOLO11 improvement is 256×256 pixels, the training batch size is 200, the optimizer is AdamW, the learning rate adaptive adjustment class is StepLR, the learning rate is 0.001, and the training batch size is 16.

7. The intelligent gas detection method in infrared images based on lightweight deformable convolution according to claim 1, characterized in that, The loss function used for training the deep learning network model based on lightweight deformable convolution is a hybrid loss function of the localization loss function and the classification loss function. The localization loss function uses the CIoU loss function, and the calculation formula is as follows: in, These represent the predicted bounding box, its width, height, and center coordinates, respectively. This represents the actual bounding box, its width, height, and center coordinates. This represents the intersection-union ratio of the two boxes. To calculate the Euclidean distance between two points, This represents the diagonal distance of the smallest closed box that is compatible with both the predicted and actual detection boxes. and The measurement focuses on the similarity between the predicted and the actual bounding boxes. The classification loss function used is the Slide loss function, calculated as follows: in, The IoU threshold for determining whether a sample is a positive sample is set to 0.

5. and Let represent the true category and the category predicted by the network, respectively. The final loss function is calculated as follows: 。

Images