Wind disturbance resistant infrared gas leakage time sequence segmentation detection method and system
Patent Information
- Application Number
- CN202611442718.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-17
- Publication Date
- 2026-10-27
AI Technical Summary
[0007]本发明要解决的技术问题是针对上述现有技术的不足，提供一种抗风扰的红外气体泄漏时序分割检测方法及系统，针对存在风场扰动及复杂背景干扰的工业场景进行红外气体泄漏检测，以解决现有技术中在复杂风扰和背景漂移条件下气体检测精度低、误报率高的问题
[0040]采用上述技术方案所产生的有益效果在于：本发明提供的一种抗风扰的红外气体泄漏时序分割检测方法及系统，采用基于四通道时序背景输入，结合风扰红外分割网络和鲁棒细化处理的检测方案，能够显著提高复杂工业场景下气体泄漏区域识别的稳定性，降低误报和漏检概率，提高检测的准确度和完整性。
Smart Images
Figure CN122989722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and industrial safety monitoring technology, and in particular to a wind-resistant infrared gas leak time-series segmentation detection method and system. Background Technology
[0002] Gases with infrared absorption properties, such as sulfur hexafluoride and methane, are widely used in industrial production. Leaks of these gases are difficult to detect directly, posing a significant risk of safety accidents and air pollution. Infrared thermal imaging technology is an important detection method, but traditional methods have low levels of intelligence. In recent years, with the development of deep learning, intelligent detection algorithms have been widely used, but existing intelligent detection technologies still have significant shortcomings under extreme wind disturbance conditions.
[0003] 1. Poor resistance to wind disturbance: Under strong wind conditions, gas plumes may become thinner, more fragmented, more dispersed, or even break in some areas, leading to missed detections by traditional intelligent detection models.
[0004] 2. Severe background interference: The industrial background of infrared images may darken slowly over time or undergo thermal distribution drift, causing the intelligent detection model to misjudge non-gas changes as gas, resulting in a decrease in the detection rate.
[0005] 3. Limitations of feature extraction: Existing intelligent detection models often struggle to take into account both the overall morphological features of the gas and minute edge details.
[0006] Therefore, there is an urgent need for an infrared gas leak detection method that can improve the stability and robustness of gas detection and effectively cope with wind disturbance and background drift. Summary of the Invention
[0007] The technical problem to be solved by the present invention is to provide a wind-resistant infrared gas leakage time-series segmentation detection method and system to address the shortcomings of the prior art. This method is designed for infrared gas leakage detection in industrial scenarios with wind field disturbances and complex background interference, thereby solving the problems of low gas detection accuracy and high false alarm rate in the prior art under complex wind disturbance and background drift conditions.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0009] On one hand, the present invention provides a wind-resistant infrared gas leakage time-series segmentation detection method, comprising:
[0010] Acquire infrared gas leak video data, select the previous frame, current frame, next frame and static background image to perform channel stitching, and construct four-channel time-series background input data;
[0011] Input the four-channel timing background data Input the wind-disturbed infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map;
[0012] Based on four-channel temporal background input data, temporal-background residual cue features are calculated; the temporal-background residual cue is used to characterize the difference between the current frame and the static background, the temporal changes between adjacent frames, and local fragmentation changes;
[0013] The fused feature maps are input into multi-scale dilated convolution branches, and the outputs of each branch are concatenated and fused to obtain wind disturbance gas recovery features. The fused feature maps are then updated with residuals using temporal-background residual cue features to obtain the recovered feature maps. Finally, the recovered wind disturbance segmentation score map is obtained by combining the initial segmentation score map with the original segmentation score map.
[0014] Based on the feature map after wind disturbance recovery, the temporal-background residual cue features, and the multilayer perceptron, the probability of gas presence is calculated. When the probability of gas presence is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection.
[0015] The segmentation score map after empty frame protection is corrected to obtain the final gas segmentation score map; the gas leakage area in the infrared image is determined based on the final gas segmentation score map.
[0016] Furthermore, the wind-disturbed infrared gas segmentation network is constructed based on the SegFormer semantic segmentation structure. The original input channels of the SegFormer semantic segmentation structure are expanded to four channels. In the first layer of the image block embedding layer, a parallel input structure is set, including the main path and the detail path. The main path uses a 7×7 convolutional kernel with a stride of 4 to extract the overall morphological features of the four-channel temporal background input data. The detail path uses two levels of 3×3 convolutional kernels to extract the weak details of the four-channel temporal background input data. The features are then superimposed through residual supplementation to obtain the initial features after structural fusion.
[0017] The initial features extracted from the first layer's image patch embedding layer and then fused are input into the SegFormer encoder and decoder head, and the fused feature map is obtained after processing. and initial segmentation score map .
[0018] Furthermore, the time-series-background residual cue includes background difference intensity, low-frequency background drift, high-frequency local variation, bidirectional time-series motion intensity, continuous variation trend, and low-frequency background drift ratio. These six cuees are spliced together along the channel dimension to obtain the time-series-background residual cue feature C.
[0019] Furthermore, the multi-scale dilated convolution branches include at least four dilated convolution branches with different expansion rates. The outputs of each branch are concatenated and fused to obtain the wind disturbance gas recovery features. Background suppression gating is generated based on temporal-background residual cue features. Based on gas recovery characteristics And time-series-background residual cue features C generated gas recovery gating ; for fused feature maps Perform residual update to obtain the restored feature map. ,in, The recovery strength coefficient;
[0020] The recovered feature map is then subjected to a convolution operation to generate a wind disturbance gas recovery and enhancement segmentation score map. The result is then overlaid onto the initial segmentation score map to obtain the segmentation score map after wind disturbance recovery. ,in, To restore the score enhancement coefficient.
[0021] Furthermore, the specific method for calculating the probability of gas presence based on the feature map after wind disturbance recovery, temporal-background residual cue features, and multilayer perceptron is as follows:
[0022] Feature map after wind disturbance recovery Perform global average pooling and global max pooling to obtain global average pooling features. and global max pooling features ;
[0023] Temporal-background residual cue features Perform global average pooling to obtain clue statistical features. The global average pooling feature, global max pooling feature, and cue statistics feature are concatenated and input into the multilayer perceptron to obtain the gas presence probability. ;
[0024] When the probability of gas presence is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection. ,in, This is the null frame protection strength coefficient. This is the threshold for determining the presence of gas.
[0025] Furthermore, the specific method for correcting the segmentation score map after empty frame protection to obtain the final gas segmentation score map is as follows:
[0026] The segmentation score image after protecting the empty frames Input the activation function to obtain the probability map after null frame protection. Then, soft expansion processing is performed to obtain a soft expansion probability map. Then, a soft corrosion treatment is performed to obtain a continuous gas probability map. ;
[0027] According to the probability diagram of continuous gases Probability map after empty frame protection The difference between them yields the candidate hole map: Where H represents a candidate void region located inside the gas mass; and a gas mass gating system is designed. ,in, The threshold for determining the gas matrix. For temperature coefficient, Indicates gas bulk gating;
[0028] Based on the gas matrix gating, the cavity candidate map, and the gas existence probability, the segmentation score map after empty frame protection is corrected to obtain the final gas segmentation score map. , The strength coefficient for filling voids.
[0029] Furthermore, the specific method for determining the gas leakage area in the infrared image based on the final gas segmentation score map is as follows:
[0030] Final gas segmentation score diagram The final gas probability map is obtained by inputting the Sigmoid activation function, and then the final gas leakage segmentation mask is obtained by thresholding. The gas leakage area in the infrared image is determined based on the final gas leakage segmentation mask.
[0031] On the other hand, the present invention also provides an infrared gas leak time-series segmentation detection system resistant to wind interference, including an input data construction module, an initial segmentation module, a cue feature construction module, a gas recovery module, a blank frame protection module, and a gas leak detection module;
[0032] The input data construction module acquires infrared gas leak video data, selects the previous frame, current frame, next frame and static background image to perform channel stitching, and constructs four-channel time-series background input data.
[0033] The initial segmentation module inputs the four-channel temporal background data. Input the wind-disturbed infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map;
[0034] The cue feature construction module calculates the temporal-background residual cue features based on four-channel temporal background input data; the temporal-background residual cue is used to characterize the difference between the current frame and the static background, the temporal changes between adjacent frames, and local fragmentation changes;
[0035] The gas recovery module inputs the fused feature map into the multi-scale dilated convolution branch, and splices and convolves the outputs of each branch to obtain the wind disturbance gas recovery feature; and combines the temporal-background residual cue features to perform residual update on the fused feature map to obtain the recovered feature map, and further obtains the segmentation score map after wind disturbance recovery;
[0036] The empty frame protection module calculates the gas presence probability based on the feature map after wind disturbance recovery, the temporal-background residual cue features, and the multilayer perceptron. When the gas presence probability is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection.
[0037] The gas leak detection module corrects the segmentation score map after empty frame protection to obtain the final gas segmentation score map; and determines the gas leak area in the infrared image based on the final gas segmentation score map.
[0038] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the wind-resistant infrared gas leak timing segmentation detection method.
[0039] Fourthly, this application proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned wind-resistant infrared gas leakage time-series segmentation detection method.
[0040] The beneficial effects of adopting the above technical solution are as follows: The wind-disruption-resistant infrared gas leak time-series segmentation detection method and system provided by the present invention adopts a detection scheme based on four-channel time-series background input, combined with wind-disruption infrared segmentation network and robust refinement processing, which can significantly improve the stability of gas leak area identification in complex industrial scenarios, reduce the probability of false alarms and missed detections, and improve the accuracy and completeness of detection.
[0041] Specifically, this invention utilizes the previous frame, current frame, next frame, and static background image to construct four-channel temporal background input data, enabling the detection process to simultaneously possess gas motion change information and leak-free background reference information. This allows for a more accurate distinction between real gas diffusion and background grayscale drift, reducing false detections caused by slow changes in the infrared background. Simultaneously, a parallel detail supplementation branch is constructed at the input end of the wind-disrupted infrared gas segmentation network, which can supplement weak gas edges and local textures that are easily lost in the main path features, improving the recognition ability of weak and fragmented gases.
[0042] By generating background suppression gating and gas recovery gating through time-background residual cues, the pseudo-response in the low-frequency background drift region is suppressed, while the weak gas region that becomes faint, fragmented and dispersed under wind disturbance conditions is restored and enhanced, thereby reducing the probability of missed detection in strong wind disturbance scenarios.
[0043] By protecting empty frames or low-confidence frames with the probability of gas presence, false gas responses in images without gas leaks can be suppressed. At the same time, by using soft morphological closure processing constrained by gas body gating, voids are filled in the interior or adjacent areas of the broken gas body, improving the continuity and integrity of the final gas leak segmentation mask.
[0044] In summary, this invention can improve the accuracy, robustness, and practicality of infrared gas leak segmentation detection under complex conditions such as wind disturbance, background drift, weak gas, empty frames, and gas bulk fragmentation. It is suitable for automatic identification, range estimation, and safety warning of gas leak areas in industrial sites. Attached Figure Description
[0045] Figure 1 This is a flowchart of a wind-disruption-resistant infrared gas leakage time-series segmentation detection method provided in Embodiment 1 of the present invention. Detailed Implementation
[0046] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0047] Example 1:
[0048] This embodiment provides a wind-resistant infrared gas leak segmentation detection method, such as... Figure 1 As shown, the steps include the following.
[0049] Step S1: Acquire infrared gas leak video data, select the previous frame, current frame, next frame, and static background image, and stitch them together to construct four-channel time-series background input data, represented as follows: ,in , , These represent the previous frame, current frame, and next frame of the infrared gas leak video, respectively. This is a static background image.
[0050] Step S2: Input the four-channel timing background data Input the wind-disrupted infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map.
[0051] In one specific implementation, the wind-disruption infrared gas segmentation network is constructed based on the SegFormer semantic segmentation structure. The original input channels of the SegFormer semantic segmentation structure are expanded to four channels. In the first layer, the Patch Embedding layer, a parallel input structure is set, including a main path and a detail path. The main path uses a 7×7 convolutional kernel with a stride of 4 to extract the overall morphological features of the four-channel temporal background input data. The detail path uses two levels of 3×3 convolutional kernels to extract the weak detail features of the four-channel temporal background input data. These features are then superimposed using residual compensation to obtain the initial features after structural fusion, expressed by the formula:
[0052] ;
[0053] in, These are the initial features after structural fusion. Input data for the four-channel timing background. To control the coefficient of supplementary intensity of the main road to the detailed road, () represents a 7×7 convolution kernel. This represents a two-level 3×3 convolution kernel.
[0054] The initial features extracted from the first layer of Patch Embedding are then input into the SegFormer encoder and decoder head, and the resulting fused feature map is obtained after processing. and initial segmentation score map .
[0055] Step S3: Input background data based on four-channel timing The temporal-background residual cue features are calculated. These features characterize the differences between the current frame and the static background, temporal variations between adjacent frames, and local fragmentation changes. Specifically, they include the following cue features:
[0056] Background difference intensity: ,in, This indicates the intensity of change of the current frame relative to the static background image.
[0057] Low-frequency background drift: ,in, This indicates the average pooling operation. Used to characterize background drift that is large-area, smooth, and slowly changing.
[0058] High-frequency local variations: ,in, Used to characterize local abrupt changes, edge variations, and fragmented gas variations in the current frame.
[0059] Two-way temporal motion intensity: ,in, Used to characterize the intensity of temporal changes in the forward and backward directions of the current frame.
[0060] Continuous trend of change: ,in, Represents the linear rectified function. Used to characterize regions where the direction of change is continuous and consistent between adjacent frames;
[0061] Low-frequency background drift percentage: ,in, To prevent constants with a denominator of zero, Used to characterize the proportion of low-frequency background drift in the overall change.
[0062] By concatenating the above six clues along the channel dimension, we obtain the temporal-background residual clue features: .
[0063] Step S4: Input the fused feature map F into the multi-scale dilated convolution branches respectively, and concatenate and fuse the outputs of each branch to obtain the wind disturbance gas recovery features; and combine the temporal-background residual cue features to analyze the fused feature map. Perform residual update to obtain the restored feature map. And combine the initial segmentation score map to obtain the segmentation score map after wind disturbance recovery;
[0064] In this embodiment, the multi-scale dilated convolution branch includes at least four dilated convolution branches with different dilation rates. For example, the dilation rate can be: The dilated convolution branches with different dilation rates output the following: The outputs of each branch are concatenated and fused using convolution to obtain the wind disturbance gas recovery features: Smaller expansion ratios are used to extract local weak gas and edge details, while larger expansion ratios are used to extract large-scale diffuse gas morphology.
[0065] Generate background suppression gating based on temporal-background residual cue features: ,in, This indicates background suppression gating, used to reduce the response in background drift regions.
[0066] Based on gas recovery characteristics And time-series-background residual cue feature C generated gas recovery gating:
[0067] in, This indicates gas recovery gating, used to enhance weak gases, broken gases, and gas regions with a tendency to move continuously.
[0068] For fused feature maps Perform residual update to obtain the restored feature map. : ,in, This is the recovery strength coefficient.
[0069] Furthermore, the recovered feature map is subjected to a convolution operation to generate a wind disturbance gas recovery and enhancement segmentation score map. : The result is then overlaid onto the initial segmentation score map to obtain the segmentation score map after wind disturbance recovery. : ,in, To restore the score enhancement coefficient.
[0070] Through the above processing, the model can enhance the weak gas, broken gas, and wind-disturbed diffusion gas regions while suppressing the background drift region.
[0071] Step S5: Recover the feature map after wind disturbance Perform global average pooling and global max pooling to obtain global average pooling features. and global max pooling features Where GAP represents global average pooling and GMP represents global max pooling.
[0072] Simultaneously, the temporal-background residual cue features Perform global average pooling to obtain clue statistical features: The global average pooling feature, global max pooling feature, and cue statistics feature are concatenated and input into the multilayer perceptron to obtain the probability of gas presence: .
[0073] When the probability of gas presence is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection. : ,in, This is the null frame protection strength coefficient. This is the threshold for determining the presence of gas. This step reduces the false gas response in empty frames, background-drifted frames, or frames with weak texture interference.
[0074] Step S6: Extract the segmentation score image after null frame protection. Input the activation function to obtain the probability map after null frame protection. Then, soft expansion processing is performed to obtain a soft expansion probability map. : Then, a soft corrosion treatment is performed to obtain a continuous gas probability diagram: ,in, This represents a probability diagram for continuous gases.
[0075] According to the probability diagram of continuous gases Probability map after empty frame protection The difference between them yields the candidate hole map: Where H represents a candidate void region located inside the gas mass, and a gas mass gating system is designed. ,in, The threshold for determining the gas matrix. For temperature coefficient, This indicates gas bulk gating.
[0076] Based on the gas matrix gating, the cavity candidate map, and the gas existence probability, the segmentation score map after empty frame protection is corrected to obtain the final gas segmentation score map: ,in, This is the final gas segmentation score map. The strength coefficient for filling voids.
[0077] With this processing method, hole completion mainly occurs inside or near the gas body and is constrained by the confidence probability of gas presence, thus avoiding the expansion of erroneous responses in empty frames or background false alarm areas.
[0078] Step S7: Calculate the final gas segmentation score. By inputting the Sigmoid activation function, the final gas probability map is obtained: Then, the final gas leak segmentation mask is obtained by thresholding, and the gas leak area in the infrared image is determined based on the final gas leak segmentation mask.
[0079] In one implementation, either the basic segmentation mode or the robust refinement mode can be selected according to actual needs.
[0080] When using the basic segmentation pattern, the initial segmentation score map will be... By directly inputting the Sigmoid activation function, we obtain the initial probability map: Then, an initial gas segmentation mask is obtained through thresholding, and the gas leakage area in the infrared image is determined based on the initial gas leakage segmentation mask, i.e., steps 1-2 are executed. This mode has a shorter calculation process and is suitable for scenarios with weak wind disturbance and relatively stable background.
[0081] When the robust thinning mode is used, steps 1-7 are executed to obtain the robust thinning segmentation and detection results.
[0082] Example 2:
[0083] This embodiment provides a wind-resistant infrared gas leak time-series segmentation detection system, including an input data construction module, an initial segmentation module, a cue feature construction module, a gas recovery module, a blank frame protection module, and a gas leak detection module;
[0084] The input data construction module acquires infrared gas leak video data, selects the previous frame, current frame, next frame and static background image to perform channel stitching, and constructs four-channel time-series background input data.
[0085] The initial segmentation module inputs the four-channel temporal background data. Input the wind-disturbed infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map;
[0086] The cue feature construction module calculates the temporal-background residual cue features based on four-channel temporal background input data; the temporal-background residual cue is used to characterize the difference between the current frame and the static background, the temporal changes between adjacent frames, and local fragmentation changes;
[0087] The gas recovery module inputs the fused feature map into the multi-scale dilated convolution branch, and splices and convolves the outputs of each branch to obtain the wind disturbance gas recovery feature; and combines the temporal-background residual cue features to perform residual update on the fused feature map to obtain the recovered feature map, and further obtains the segmentation score map after wind disturbance recovery;
[0088] The empty frame protection module calculates the gas presence probability based on the feature map after wind disturbance recovery, the temporal-background residual cue features, and the multilayer perceptron. When the gas presence probability is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection.
[0089] The gas leak detection module corrects the segmentation score map after empty frame protection to obtain the final gas segmentation score map; and determines the gas leak area in the infrared image based on the final gas segmentation score map.
[0090] Example 3:
[0091] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
[0092] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the wind-resistant infrared gas leak segmentation detection method described in the various embodiments of this application.
[0093] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned wind-resistant infrared gas leak segmentation detection method.
[0094] Example 4:
[0095] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned wind-resistant infrared gas leak segmentation detection method.
[0096] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.
[0097] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.
Claims
1. A wind-resistant infrared gas leak time-series segmentation detection method, characterized in that, include: Acquire infrared gas leak video data, select the previous frame, current frame, next frame and static background image to perform channel stitching, and construct four-channel time-series background input data; Input the four-channel timing background data Input the wind-disturbed infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map; Based on four-channel temporal background input data, temporal-background residual cue features are calculated; the temporal-background residual cue is used to characterize the difference between the current frame and the static background, the temporal changes between adjacent frames, and local fragmentation changes; The fused feature maps are input into multi-scale dilated convolution branches, and the outputs of each branch are concatenated and fused to obtain wind disturbance gas recovery features. The fused feature maps are then updated with residuals using temporal-background residual cue features to obtain the recovered feature maps. Finally, the recovered wind disturbance segmentation score map is obtained by combining the initial segmentation score map with the original segmentation score map. Based on the feature map after wind disturbance recovery, the temporal-background residual cue features, and the multilayer perceptron, the probability of gas presence is calculated. When the probability of gas presence is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection. The segmentation score map after empty frame protection is corrected to obtain the final gas segmentation score map; the gas leakage area in the infrared image is determined based on the final gas segmentation score map.
2. The wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to claim 1, characterized in that, The wind-disturbed infrared gas segmentation network is constructed based on the SegFormer semantic segmentation structure. The original input channels of the SegFormer semantic segmentation structure are expanded to four channels. In the first layer of the image block embedding layer, a parallel input structure is set, including the main path and the detail path. The main path uses a 7×7 convolutional kernel with a stride of 4 to extract the overall morphological features of the four-channel temporal background input data. The detail path uses two levels of 3×3 convolutional kernels to extract the weak details of the four-channel temporal background input data. The features are then superimposed through residual supplementation to obtain the initial features after structural fusion. The initial features extracted from the first layer's image patch embedding layer and then fused are input into the SegFormer encoder and decoder head, and the fused feature map is obtained after processing. and initial segmentation score map .
3. The wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to claim 2, characterized in that, The temporal-background residual cue includes background difference intensity, low-frequency background drift, high-frequency local variation, bidirectional temporal motion intensity, continuous variation trend, and low-frequency background drift ratio. These six cuees are spliced together along the channel dimension to obtain the temporal-background residual cue feature C.
4. The wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to claim 3, characterized in that, The multi-scale dilated convolution branch includes at least four dilated convolution branches with different expansion rates. The outputs of each branch are concatenated and fused to obtain the wind disturbance gas recovery feature. Background suppression gating is generated based on temporal-background residual cue features. Based on gas recovery characteristics And time-series-background residual cue features C generated gas recovery gating ; for fused feature maps Perform residual update to obtain the restored feature map. ,in, The recovery strength coefficient; The recovered feature map is then subjected to a convolution operation to generate a wind disturbance gas recovery and enhancement segmentation score map. The result is then overlaid onto the initial segmentation score map to obtain the segmentation score map after wind disturbance recovery. ,in, To restore the score enhancement coefficient.
5. The wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to claim 4, characterized in that, The specific method for calculating the probability of gas presence based on the feature map after wind disturbance recovery, time-series-background residual cue features, and multilayer perceptron is as follows: Feature map after wind disturbance recovery Perform global average pooling and global max pooling to obtain global average pooling features. and global max pooling features ; Temporal-background residual cue features Perform global average pooling to obtain clue statistical features. The global average pooling feature, global max pooling feature, and cue statistics feature are concatenated and input into the multilayer perceptron to obtain the gas presence probability. ; When the probability of gas presence is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection. ,in, This is the null frame protection strength coefficient. This is the threshold for determining the presence of gas.
6. The wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to claim 5, characterized in that, The specific method for correcting the segmentation score map after empty frame protection to obtain the final gas segmentation score map is as follows: The segmentation score image after protecting the empty frames Input the activation function to obtain the probability map after null frame protection. Then, soft expansion processing is performed to obtain a soft expansion probability map. Then, a soft corrosion treatment is performed to obtain a continuous gas probability map. ; According to the probability diagram of continuous gases Probability map after empty frame protection The difference between them yields the candidate hole map: Where H represents a candidate void region located inside the gas mass; and a gas mass gating system is designed. ,in, The threshold for determining the gas matrix. For temperature coefficient, Indicates gas bulk gating; Based on the gas matrix gating, the cavity candidate map, and the gas existence probability, the segmentation score map after empty frame protection is corrected to obtain the final gas segmentation score map. , The strength coefficient for filling voids.
7. The wind-disruption-resistant infrared gas leak time-series segmentation detection method according to claim 6, characterized in that, The specific method for determining the gas leakage area in the infrared image based on the final gas segmentation score map is as follows: Final gas segmentation score diagram The final gas probability map is obtained by inputting the Sigmoid activation function, and then the final gas leakage segmentation mask is obtained by thresholding. The gas leakage area in the infrared image is determined based on the final gas leakage segmentation mask.
8. A wind-disruption-resistant infrared gas leak time-series segmentation detection system, implemented based on the wind-disruption-resistant infrared gas leak time-series segmentation detection method of claim 1, characterized in that, It includes an input data construction module, an initial segmentation module, a clue feature construction module, a gas recovery module, a null frame protection module, and a gas leak detection module; The input data construction module acquires infrared gas leak video data, selects the previous frame, current frame, next frame and static background image to perform channel stitching, and constructs four-channel time-series background input data. The initial segmentation module inputs the four-channel temporal background data. Input the wind-disturbed infrared gas segmentation network to obtain the fused feature map and the initial segmentation score map; The cue feature construction module calculates the temporal-background residual cue features based on four-channel temporal background input data; the temporal-background residual cue is used to characterize the difference between the current frame and the static background, the temporal changes between adjacent frames, and local fragmentation changes; The gas recovery module inputs the fused feature map into the multi-scale dilated convolution branch, and splices and convolves the outputs of each branch to obtain the wind disturbance gas recovery feature; and combines the temporal-background residual cue features to perform residual update on the fused feature map to obtain the recovered feature map, and further obtains the segmentation score map after wind disturbance recovery; The empty frame protection module calculates the gas presence probability based on the feature map after wind disturbance recovery, the temporal-background residual cue features, and the multilayer perceptron. When the gas presence probability is lower than a preset threshold, a negative empty frame protection bias is generated and superimposed on the segmentation score map after wind disturbance recovery to obtain the segmentation score map after empty frame protection. The gas leak detection module corrects the segmentation score map after empty frame protection to obtain the final gas segmentation score map; and determines the gas leak area in the infrared image based on the final gas segmentation score map.
9. A computer-readable storage medium storing executable instructions for performing the wind-disruption-resistant infrared gas leak time-series segmentation detection method according to any one of claims 1-7, characterized in that: When the instruction is executed, it causes the processor to perform the wind-resistant infrared gas leak timing segmentation detection method.
10. A computer program product for executing the wind-disruption-resistant infrared gas leakage time-series segmentation detection method according to any one of claims 1-7, characterized in that: It includes a computer program or instructions that, when executed by a processor, implement the wind-resistant infrared gas leak timing segmentation detection method.