A non-contact visual temperature detection method and system in foggy environments

Abstract

This invention discloses a non-contact visual temperature detection method and system in foggy environments, belonging to the field of information technology. The method involves setting target points for the high-temperature target under test according to a preset temperature range and interval, thus determining the temperature point to be measured. Conditional water mist temperature images of the temperature point under test are acquired in the foggy environment and input into a trained adversarial defogging model for defogging processing to obtain a defogging temperature image. High-temperature pixels are extracted using an image segmentation algorithm to obtain high-temperature region pixel blocks, which are then input into a trained light image-temperature field mapping model for feature mapping to obtain the temperature data of the high-temperature target. This invention achieves accurate temperature measurement in foggy environments, solving the problems of traditional industrial temperature measurement methods being susceptible to fog interference, easily losing key high-temperature features, having weak detail perception capabilities, and low measurement accuracy. It can effectively adapt to non-contact temperature measurement scenarios in various complex industrial environments.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a non-contact visual temperature detection method and system in foggy environments. Background Technology

[0002] Precise temperature measurement plays a crucial role in ensuring safe production, improving efficiency, and reducing energy consumption in industries such as power, petrochemicals, and aerospace. Currently, infrared thermography is a common temperature measurement method. However, considering the complex industrial production environment, the target is subject to interference from multiple sources of infrared radiation. This external radiation can easily overlap with the target's own radiation signal, leading to confusion in radiation source identification. Furthermore, infrared images have low resolution, making it difficult to meet the requirements for precise temperature measurement in complex environments. With the development of industrial visible light cameras, the visible light images they capture contain rich temperature information, and their high resolution and other superior performance have led to their increasingly widespread application in industrial production.

[0003] Currently, common methods for temperature measurement based on visible light images generally employ physical models or data-driven approaches. Physical models establish empirical functions between visible light images and temperature using colorimetry, trichromatic methods, etc. (For details, see Jiang, Zhi-Wei, Zi-Xue Luo, and Huai-Chun Zhou. "A simple measurement method of temperature and emissivity of coal-fired flames from visible radiation image and its application in a CFB boiler furnace." Fuel88.6(2009):980-987.). Data-driven methods establish nonlinear mappings between image features and temperature (Tan X, Yin C, Liu J, et al. A Fast Measurement Method of Temperature Field on High Temperature SurfaceBased on Color CCD[C] / / 2024 7th International Conference on Pattern Recognition and Artificial Intelligence (PRAI). IEEE, 2024: 1107-1112.). By analyzing the sources of error in high-temperature measurement scenarios, a fast temperature field measurement method for multilayer sensors in high-temperature environments is designed. In addition, a high-temperature soft measurement model based on backpropagation neural network (BPNN) can be established based on the molten salt temperature image features (Chen Xinyu, Wu Xinyu, Liu Feifei, et al. Research on high-temperature soft measurement of rare earth molten salt based on color CCD [J]. Rare Earth, 2025, 46(01):99-108.) for temperature measurement.

[0004] However, these methods largely assume image clarity and fail to account for fog interference in real-world scenarios. Fog interference causes reduced brightness, decreased color saturation, and hue shift in visible light images, interfering with the thermal radiation characteristics of high-temperature regions and severely degrading the accuracy of subsequent temperature measurements. Therefore, proposing a temperature measurement method that removes fog interference while accurately preserving key features of high-temperature regions, and providing an end-to-end temperature measurement solution that covers fog noise filtering, retention of effective high-temperature features, temperature error correction, accurate analysis of temperature data, and stable output of temperature results, is a key research direction for technical personnel. Summary of the Invention

[0005] In view of this, the present invention provides a non-contact visual temperature detection method and system in foggy environments. It builds an adversarial defogging network and combines it with a temperature mapping model based on ResNet18 to achieve accurate output of temperature values. This solves the problems of traditional industrial temperature measurement methods being susceptible to fog interference, easy loss of key high-temperature features, weak detail perception, and low temperature measurement accuracy. The overall temperature measurement results are of excellent accuracy and can be effectively adapted to non-contact temperature measurement scenarios in various complex industrial environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A non-contact visual temperature detection method in a foggy environment includes the following steps: S1 Preparation Steps: Set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and determine the temperature point to be measured. S2 acquisition steps: Acquire the conditional water mist temperature image of the temperature point to be measured under fog conditions; S3 Dehazing Step: Input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing to generate a dehazed temperature image; S4 segmentation step: Use image segmentation algorithm to extract high-temperature pixels from the dehazed temperature image to obtain high-temperature region pixel blocks; S5 mapping step: Input the pixel blocks of the high-temperature region into the trained light image-temperature field mapping model for feature mapping to obtain the temperature data of the high-temperature target to be measured.

[0007] Optionally, in the above method, the high-temperature target to be tested in the S1 preparation step is an M335 blackbody calibration source; correspondingly, the preset temperature range is 650℃ to 750℃, the preset temperature interval is 1℃, and the target points are set to 101.

[0008] Optionally, in the above method, the adversarial dehazing model in S3 includes a generator and a discriminator connected in sequence; the generator adopts an encoder-decoder architecture, and the discriminator adopts a local region discrimination structure. The generator consists of a color perception module, a downsampling path, a residual enhancement module, an upsampling path, and an output layer connected in sequence. The color perception module structure includes a channel attention branch and a spatial attention branch, which are connected in sequence to the downsampling path.

[0009] Optionally, the training process of the adversarial dehazing model in S3 includes the following: Image acquisition steps: Acquire temperature images of the target test point, and acquire conditional water mist temperature images of the target test point according to the preset water mist concentration; Image dehazing step: Use a generator to map the conditional water mist temperature image to obtain the dehazed temperature image of the target test point; Image discrimination steps: Based on the local region discrimination mechanism of the discriminator, the realism of the dehazed temperature image is judged according to the temperature image to obtain the dehazing discrimination result; Model optimization steps: Using the dehazing discrimination results, compare the difference between the distribution characteristics of the temperature image and the distribution characteristics of the dehazed temperature image with the preset feature difference; if it is greater than the preset feature difference, return to the image dehazing step to optimize the network parameters of the generator; if it is less than the preset feature difference, determine the current adversarial dehazing model as the trained adversarial dehazing model.

[0010] Optionally, the adversarial dehazing model in S3 operates based on region adaptive regularization using a brightness threshold. The mask generation strategy based on the brightness threshold is as follows: ; in, For indicator functions; For pixels The mean value across all color channels; Brightness threshold; mask The value is 1 in high-temperature areas and 0 in non-high-temperature areas.

[0011] Optionally, the above method, based on mask design for color fidelity loss, applies constraints only to the high-temperature region. The specific constraints are as follows: ; in, To preserve color fidelity, To generate an image; A true, fog-free image; This is element-wise multiplication; for Norm, For masking.

[0012] The above method, optionally, allows the total loss function of the adversarial dehazing model to be a weighted average of the reconstruction loss, color fidelity loss, and adversarial loss; the specific expression for the total loss function is as follows: ; in, To rebuild the losses, To combat the losses.

[0013] Optionally, the above method uses ResNet18 to construct a light image-temperature field mapping model. The training process of the light image-temperature field mapping model in S5 specifically includes: Pixel block acquisition steps: Obtain the temperature image of the high-temperature target for testing, and use the image segmentation algorithm to extract high-temperature pixels to obtain high-temperature region pixel blocks; Feature capture steps: Use the current ResNet18 to capture the image features of pixel blocks in high-temperature regions to obtain the corresponding temperature data; The calculation steps are as follows: Calculate the mean absolute error and root mean square error of the temperature data, and use the mean absolute error and root mean square error as the evaluation metrics of the current ResNet18. Model iteration steps: Determine whether the evaluation index meets the preset evaluation requirements. If it meets the preset evaluation requirements, then determine the current ResNet18 as the trained light image-temperature field mapping model. If it does not meet the preset evaluation requirements, then optimize the network parameters of ResNet18, determine the optimized ResNet18 as the current ResNet18, and return to the feature capture step.

[0014] A non-contact visual temperature detection system for foggy environments, used to implement the non-contact visual temperature detection method for foggy environments as described in any of the above claims, comprising a preparation module, an acquisition module, a defogging module, a segmentation module, and a mapping module connected in sequence; The preparation module is used to set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and to determine the temperature point to be measured. The acquisition module is used to acquire conditional water mist temperature images of the temperature points to be measured based on preset water mist concentrations. The dehazing module is used to input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing and generate a dehazed temperature image. The segmentation module is used to extract high-temperature pixels from the dehazed temperature image using an image segmentation algorithm, resulting in high-temperature region pixel blocks. The mapping module is used to input the pixel blocks of high-temperature regions into the trained light image-temperature field mapping model for mapping processing, so as to obtain the temperature data of the high-temperature target to be measured.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention provides a non-contact visual temperature detection method and system in foggy environments, which has the following beneficial effects: (1) Targeted dehazing and interference resistance: Construct an adversarial dehazing network in which the generator and discriminator work together, optimize the image dehazing effect by relying on the encoder-decoder architecture and the channel-space collaborative attention module, and combine it with a local region discrimination mechanism to effectively overcome the fog interference problem in industrial scenarios; (2) High-fidelity feature preservation: Based on the brightness threshold, a high-temperature region mask is generated, an adaptive multi-objective loss function for the region is built, and a high-temperature color fidelity constraint is added to accurately preserve the core features of the high-temperature region and avoid the loss of key temperature measurement information during the defogging process; (3) Strong detail perception capability: The gradient constraint mechanism of high temperature region is introduced to enhance the discriminator’s ability to capture and identify local details in high temperature region, optimize the detail quality of dehazed image, and provide a high-quality image foundation for subsequent temperature calculation; (4) High-precision temperature measurement adaptation: Based on ResNet18, a dedicated temperature mapping model is built to perform temperature analysis on the optimized fog-free high-definition image, so as to achieve accurate temperature value output. The temperature measurement accuracy is excellent and can be stably adapted to non-contact temperature measurement scenarios in various complex industrial environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of a non-contact visual temperature detection method in a foggy environment disclosed in this invention. Figure 2 This is a flowchart illustrating a non-contact visual temperature detection method in a foggy environment, as disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram of the generator for the adversarial dehazing model disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the discriminator structure of the adversarial dehazing model disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the color perception module in the generator of the adversarial dehazing model disclosed in an embodiment of the present invention; Figure 6 This is a comparison chart of temperature measurement errors of the target under test before and after defogging in a foggy environment, as disclosed in an embodiment of the present invention. Figure 7 This is a comparison chart of temperature measurement errors of the target under test before and after defogging in a dense fog environment, as disclosed in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0020] See Figure 1 As shown, to adapt to complex industrial production environments and meet different temperature measurement needs, this invention discloses a non-contact visual temperature detection method in foggy environments, comprising the following steps: S1 Preparation Steps: Set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and determine the temperature point to be measured. S2 acquisition steps: Acquire the conditional water mist temperature image of the temperature point to be measured under fog conditions; S3 Dehazing Step: Input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing to generate a dehazed temperature image; S4 segmentation step: Use image segmentation algorithm to extract high-temperature pixels from the dehazed temperature image to obtain high-temperature region pixel blocks; S5 mapping step: Input the pixel blocks of the high-temperature region into the trained light image-temperature field mapping model for feature mapping to obtain the temperature data of the high-temperature target to be measured.

[0021] Optionally, in the S1 preparation step, the high-temperature target to be tested is an M335 blackbody calibration source.

[0022] The present invention discloses a non-contact visual temperature detection method in a foggy environment, which needs to be able to measure the temperature of different high-temperature objects. The M335 blackbody calibration source adopts a precision graphite cavity with an emissivity of 0.995~1.0, which is close to that of an ideal blackbody. It can effectively reduce the temperature measurement radiation error and meet the high-precision calibration requirements of infrared equipment. Therefore, the present invention takes the M335 blackbody calibration source as an example to provide a detailed description of the temperature measurement of high-temperature targets in a foggy environment.

[0023] The M335 blackbody calibration source has a temperature measurement range of 300~1500℃, covering high-temperature industrial testing scenarios in industries such as metallurgy, power, and chemicals, and is compatible with the calibration of various medium- and high-temperature equipment. The device is equipped with a digital PID temperature control system, achieving long-term temperature stability of ±1℃ / 8h, with rapid heating, effectively improving on-site calibration efficiency and adapting to routine batch calibration operations. Simultaneously, the M335 blackbody calibration source supports NIST metrological traceability, comes with a formal calibration certificate, and ensures compliant and valid test data, meeting the needs of industrial acceptance, third-party testing, and scientific research testing. Furthermore, the M335 blackbody calibration source has a compact and lightweight structure, with built-in temperature control alarm and heat dissipation structure, strong anti-interference capabilities, and high stability, making it suitable for long-term calibration work in complex industrial scenarios such as workshops and production lines. Based on the working characteristics of the M335 blackbody calibration source, this invention also specifically designed image acquisition conditions for acquiring temperature images in foggy environments to obtain water mist temperature images with significant characteristics that enable the temperature detection method of this invention.

[0024] Correspondingly, the preset temperature range is 650℃ to 750℃, the preset temperature interval is 1℃, and the target points are set to 101.

[0025] When this invention is used to detect the temperature of other types of high-temperature targets in a foggy environment, the basic parameters of the high-temperature target can be obtained, such as from the procurement data of the high-temperature target, the operating instructions of the operator, and the operating range of the normal working process. Then, the corresponding preset temperature range and preset temperature interval are selected, and the target point is set to collect temperature images under foggy conditions.

[0026] Optionally, the adversarial dehazing model in S3 includes a generator and a discriminator connected in sequence; the generator adopts an encoder-decoder architecture, and the discriminator adopts a local region discrimination structure. See Figure 3 As shown, the generator includes a color perception module, a downsampling path, a residual enhancement module, an upsampling path, and an output layer connected in sequence; see also Figure 5 As shown, the color perception module structure includes a channel attention branch and a spatial attention branch, which are sequentially connected to the downsampling path. (See also...) Figure 4 As shown, the discriminator adopts a local region discrimination structure and outputs a local region scoring matrix to enhance the adversarial dehazing model's judgment of the authenticity of local details. The discriminator determines the authenticity of the input image by verifying whether the quality of the current dehazed image meets the standard of a real and clear image.

[0027] Specifically, the mathematical description of the color perception module is as follows: ; ; in, This is the original image; For channel attention; Spatial attention; This is an element-wise multiplication method that applies color adjustment parameters to the original input to achieve residual color correction.

[0028] As can be seen from the above, when performing temperature detection in a foggy environment, this invention mainly employs an adversarial defogging model to defog images of the high-temperature target obtained from different water fog concentrations. Although the learning model can perform image processing, without training, the quality of the processed images is inconsistent, which can easily affect the accuracy of subsequent temperature detection. Therefore, this invention selects an adversarial defogging model and trains and debugs it before applying it to the temperature detection of high-temperature targets. Optionally, the training process of the adversarial defogging model in S3 specifically includes: Image acquisition steps: Acquire temperature images of the target test point, and acquire conditional water mist temperature images of the target test point according to the preset water mist concentration; Image dehazing step: Use a generator to map the conditional water mist temperature image to obtain the dehazed temperature image of the target test point; Image discrimination steps: Based on the local region discrimination mechanism of the discriminator, the realism of the dehazed temperature image is judged according to the temperature image to obtain the dehazing discrimination result; Model optimization steps: Using the dehazing discrimination results, compare the difference between the distribution characteristics of the temperature image and the distribution characteristics of the dehazed temperature image with the preset feature difference; if it is greater than the preset feature difference, return to the image dehazing step to optimize the network parameters of the generator; if it is less than the preset feature difference, determine the current adversarial dehazing model as the trained adversarial dehazing model.

[0029] See Figure 2 As shown, taking the M335 blackbody calibration source as an example, an adversarial dehazing network is trained by acquiring and utilizing images under different conditions.

[0030] Due to the reliability, stability, and high precision of the M335 blackbody calibration source, this invention employs a calibration source of a different specification and model than those in the above embodiments as a high-temperature target for image dehazing learning in the adversarial dehazing model. Specifically, the temperature range of the M335 blackbody calibration source in this embodiment is set to 650℃ to 750℃, with a temperature interval of 1℃, for a total of 101 temperature points.

[0031] This invention uses a fog generator to set water mist concentrations to simulate varying degrees of fog interference in an industrial environment. Specifically, a water mist atomization rate of 360 ml / h is defined as high concentration (dense fog), and 180 ml / h as low concentration (light fog). First, 20 fog-free images are acquired at each temperature point as a learning control group for the adversarial defogging model. Then, an artificial fog environment is created using the fog generator, placing the M335 blackbody calibration source sequentially at both low concentration (light fog) and high concentration (dense fog) levels (180 ml / h). Simultaneously, 20 images are acquired at each of the 101 temperature points of the M335 blackbody calibration source at each water mist concentration. Finally, these images are combined with the 20 fog-free images at each temperature point, resulting in a total of 6060 images collected for training the adversarial defogging model.

[0032] Building upon the description of the adversarial dehazing model structure in the previous embodiment, a region-adaptive regularization based on a brightness threshold is proposed in the model's learning network. To distinguish between high-temperature and non-high-temperature regions, a mask generation strategy based on a brightness threshold is designed, utilizing the high brightness characteristics of high-temperature regions.

[0033] Optionally, the adversarial dehazing model in S3 operates based on region adaptive regularization using a brightness threshold. The mask generation strategy based on the brightness threshold is as follows: ; in, For indicator functions; For pixels The mean value across all color channels; Brightness threshold; pixels mask at the location The value is set to 1 in high-temperature regions and 0 in non-high-temperature regions, providing spatial guidance for subsequent regional adaptive constraints.

[0034] Traditional global loss functions tend to overcorrect the thermal radiation characteristics of high-temperature regions. Mask-based methods, however, can be more effective. M The design aims to preserve color fidelity, applying constraints only to high-temperature areas. L 1. The specific constraints are as follows: ; in, To preserve color fidelity, To generate an image; This is a true, fog-free image (i.e., a captured temperature image); This is element-wise multiplication; for Norm, For masking.

[0035] Color fidelity loss selectively constrains high-temperature areas through masking, preventing the destruction of their inherent color and brightness characteristics during dehazing, while also avoiding interference with fog removal in non-high-temperature areas. Furthermore, the total loss function of the adversarial dehazing model includes not only color fidelity loss, but also reconstruction loss, color fidelity loss, and adversarial loss. These three are weighted to form the total loss function of the adversarial dehazing model. Through synergistic optimization of global constraints and local protection, the overall realism of the dehazed image is ensured while preserving the color features of high-temperature areas. The specific expression for the total loss function is as follows: ; in, To rebuild the losses, To combat the losses.

[0036] After determining the overall loss function of the adversarial dehazing model, to enhance the discriminator's ability to perceive local details in high-temperature regions and avoid redundant constraints from global gradient penalties, a gradient norm constraint is applied to the high-temperature regions using a high-temperature mask. This ensures that the discriminator provides stable training signals, guiding the generator to generate richly detailed local features. Specifically, interpolated samples are first constructed: ; in, For interpolation samples, The weights are uniformly distributed random weights; this construction method only uses real samples in high-temperature regions. J and generate samples G ( z Interpolation is used to directly retain the true sample in non-high-temperature regions. J This feature reduces computational redundancy and improves constraint specificity. Based on interpolation samples... The gradient penalty loss in the high-temperature region is defined as: ; in, Gradient penalty loss; The mathematical expectation of the interpolated sample distribution in the high-temperature region; For discriminator For interpolated samples The gradient; for L The 2-norm constrains the gradient norm to be close to 1, ensuring stable feedback from the discriminator on details in high-temperature regions and guiding the generator to generate detailed high-temperature region features.

[0037] A visible light image-temperature field mapping model is constructed. Based on the temperature image of the high-temperature target used for testing, i.e. the temperature image without water mist, the high-temperature region of the defogging temperature image or the temperature image without water mist is extracted using an image segmentation algorithm. Pixel blocks are cropped at the center point and input into a ResNet18 deep residual network to output temperature values. The network is trained using MAE and RMSE as evaluation metrics.

[0038] Optionally, a light image-temperature field mapping model is constructed using ResNet18. The training process of the light image-temperature field mapping model in S5 specifically includes: Pixel block acquisition steps: Obtain the temperature image of the high-temperature target for testing, and use the image segmentation algorithm to extract high-temperature pixels to obtain high-temperature region pixel blocks; Feature capture steps: Use the current ResNet18 to capture the image features of pixel blocks in high-temperature regions to obtain the corresponding temperature data; The calculation steps are as follows: Calculate the mean absolute error and root mean square error of the temperature data, and use the mean absolute error and root mean square error as the evaluation metrics of the current ResNet18. Model iteration steps: Determine whether the evaluation index meets the preset evaluation requirements. If it meets the preset evaluation requirements, then determine the current ResNet18 as the trained light image-temperature field mapping model. If it does not meet the preset evaluation requirements, then optimize the network parameters of ResNet18, determine the optimized ResNet18 as the current ResNet18, and return to the feature capture step.

[0039] Reference Figure 6 and Figure 7Based on the trained adversarial defogging model and the trained light image-temperature field mapping model in the above embodiments of the present invention, it can be seen that, regardless of whether it is light fog or dense fog, the temperature value obtained by temperature detection through these two trained models after defogging is closer to the true temperature value than the temperature value before defogging; and the error after defogging is also closer to the error zero line than the error before defogging. The present invention, through the synergistic coupling design of the adversarial defogging network and the temperature mapping model, effectively solves the technical pain points of traditional industrial infrared temperature measurement methods in foggy interference scenarios, such as large temperature measurement deviation, loss of high-temperature features, and poor robustness. It can accurately filter out fog noise without destroying the key temperature features of the high-temperature region, significantly improving the accuracy and stability of temperature measurement under complex foggy conditions. At the same time, the model system disclosed in this invention takes into account the adaptability to extreme scenarios of slight interference in light fog and strong interference in dense fog, exhibiting excellent generalization performance and effectively avoiding the shortcomings of traditional single temperature measurement algorithms, such as narrow fog adaptability and drastic error fluctuations. Compared to conventional defogging temperature measurement solutions, this invention enables end-to-end linkage between image defogging optimization and precise temperature mapping, greatly reducing the deviation between the measured temperature and the true value, effectively reducing the system measurement error of industrial high-temperature detection, and providing reliable technical support for non-contact high-precision temperature monitoring in complex industrial scenarios such as power and metallurgy. It has strong engineering practical value and scenario adaptability.

[0040] Furthermore, taking the M335 blackbody calibration source as an example, different learning models were used to detect its temperature. The detection results are shown in Table 1, where thin represents light fog and thick represents dense fog. The results show that this invention achieves a significant accuracy improvement compared to the feature fusion attention network FFANet, the integrated defogging network AODNet, the cascaded dynamic filter defogging network CasDyFNet, and the detail enhancement attention network DEANet.

[0041] Table 1 Comparison of temperature measurement results from five methods

[0042] The calculation methods for each evaluation indicator are as follows: The formula for Mean Absolute Error (MAE) is as follows: ; in, The number of samples; To predict temperature; This is the actual temperature.

[0043] The root mean square error (RMSE) is shown below: ; in, The number of samples; To predict temperature; This is the actual temperature.

[0044] The comparison of various existing mainstream defogging temperature measurement network models in the table above shows that traditional models generally suffer from the problem of difficulty in simultaneously achieving defogging and noise reduction while preserving high-temperature features. Under foggy interference conditions, they are prone to issues such as loss of details in high-temperature areas, color shift, and temperature mapping distortion, resulting in large final temperature measurement errors and insufficient detection stability. This invention, relying on a channel-space collaborative attention mechanism and an adaptive constraint strategy for high-temperature areas, effectively removes fog noise from images and restores degraded image details while significantly enhancing the fidelity of key features of high-temperature targets, avoiding the drawbacks of excessive defogging and weakening of effective features in conventional models. Simultaneously, the optimized ResNet18 temperature mapping model achieves precise coupling and matching between image features and the temperature field, significantly reducing system temperature measurement errors in complex industrial foggy environments. This invention demonstrates superior temperature measurement accuracy and numerical stability under various fog interference conditions. Its model generalization and anti-interference capabilities are significantly better than existing similar algorithms. It effectively compensates for the shortcomings of traditional defogging temperature measurement models, such as weak adaptability to industrial scenarios and low accuracy limits. It can better meet the high-precision and high-reliability non-contact temperature detection needs in complex industrial sites and has outstanding technical advantages and engineering application value.

[0045] This invention also discloses a non-contact visual temperature detection system for use in foggy environments, for achieving, etc. Figure 1 The method for non-contact visual temperature detection in foggy environments includes a preparation module, an acquisition module, a defogging module, a segmentation module, and a mapping module connected in sequence. The preparation module is used to set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and to determine the temperature point to be measured. The acquisition module is used to acquire conditional water mist temperature images of the temperature points to be measured based on preset water mist concentrations. The dehazing module is used to input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing and generate a dehazed temperature image. The segmentation module is used to extract high-temperature pixels from the dehazed temperature image using an image segmentation algorithm, resulting in high-temperature region pixel blocks. The mapping module is used to input the pixel blocks of high-temperature regions into the trained light image-temperature field mapping model for mapping processing, so as to obtain the temperature data of the high-temperature target to be measured.

[0046] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-contact visual temperature detection method in a foggy environment, characterized in that, Includes the following steps: S1 Preparation Steps: Set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and determine the temperature point to be measured. S2 acquisition steps: Acquire the conditional water mist temperature image of the temperature point to be measured under fog conditions; S3 Dehazing Step: Input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing to generate a dehazed temperature image; S4 segmentation step: Use image segmentation algorithm to extract high-temperature pixels from the dehazed temperature image to obtain high-temperature region pixel blocks; S5 mapping step: Input the pixel blocks of the high-temperature region into the trained light image-temperature field mapping model for feature mapping to obtain the temperature data of the high-temperature target to be measured.

2. The non-contact visual temperature detection method in a foggy environment according to claim 1, characterized in that, In the S1 preparation step, the high-temperature target to be tested is the M335 blackbody calibration source; correspondingly, the preset temperature range is 650℃ to 750℃, the preset temperature interval is 1℃, and the target points are set to 101.

3. The non-contact visual temperature detection method in a foggy environment according to claim 1, characterized in that, The adversarial dehazing model in S3 consists of a generator and a discriminator connected in sequence; the generator adopts an encoder-decoder architecture, and the discriminator adopts a local region discrimination structure. The generator consists of a color perception module, a downsampling path, a residual enhancement module, an upsampling path, and an output layer connected in sequence. The color perception module structure includes a channel attention branch and a spatial attention branch, which are connected in sequence to the downsampling path.

4. The non-contact visual temperature detection method in a foggy environment according to claim 3, characterized in that, The training process of the adversarial dehazing model in S3 specifically includes: Image acquisition steps: Acquire temperature images of the target test point, and acquire conditional water mist temperature images of the target test point according to the preset water mist concentration; Image dehazing step: Use a generator to map the conditional water mist temperature image to obtain the dehazed temperature image of the target test point; Image discrimination steps: Based on the local region discrimination mechanism of the discriminator, the realism of the dehazed temperature image is judged according to the temperature image to obtain the dehazing discrimination result; Model optimization steps: Using the dehazing discrimination results, compare the difference between the distribution characteristics of the temperature image and the distribution characteristics of the dehazed temperature image with the preset feature difference; if it is greater than the preset feature difference, return to the image dehazing step to optimize the network parameters of the generator; if it is less than the preset feature difference, determine the current adversarial dehazing model as the trained adversarial dehazing model.

5. The non-contact visual temperature detection method in a foggy environment according to claim 1, characterized in that, In S3, the adversarial dehazing model operates based on region adaptive regularization using a brightness threshold. The mask generation strategy based on the brightness threshold is as follows: ； in, For indicator functions; For pixels The mean value across all color channels; Brightness threshold; mask The value is 1 in high-temperature areas and 0 in non-high-temperature areas.

6. The non-contact visual temperature detection method in a foggy environment according to claim 5, characterized in that, Based on the color fidelity loss of the mask design, constraints are applied only to the high-temperature region. The specific constraints are as follows: ； in, To preserve color fidelity, To generate an image; A true, fog-free image; This is element-wise multiplication; for Norm, For masking.

7. The non-contact visual temperature detection method in a foggy environment according to claim 6, characterized in that, The total loss function of the adversarial dehazing model is a weighted average of the reconstruction loss, color fidelity loss, and adversarial loss; the specific expression for the total loss function is as follows: ； in, To rebuild the losses, To combat the losses.

8. The non-contact visual temperature detection method in a foggy environment according to claim 1, characterized in that, A light image-temperature field mapping model is constructed using ResNet18. The training process of the light image-temperature field mapping model in S5 specifically includes: Pixel block acquisition steps: Obtain the temperature image of the high-temperature target for testing, and use the image segmentation algorithm to extract high-temperature pixels to obtain high-temperature region pixel blocks; Feature capture steps: Use the current ResNet18 to capture the image features of pixel blocks in high-temperature regions to obtain the corresponding temperature data; The calculation steps are as follows: Calculate the mean absolute error and root mean square error of the temperature data, and use the mean absolute error and root mean square error as the evaluation metrics of the current ResNet18. Model iteration steps: Determine whether the evaluation index meets the preset evaluation requirements. If it meets the preset evaluation requirements, then determine the current ResNet18 as the trained light image-temperature field mapping model. If it does not meet the preset evaluation requirements, then optimize the network parameters of ResNet18, determine the optimized ResNet18 as the current ResNet18, and return to the feature capture step.

9. A non-contact visual temperature detection system for foggy environments, characterized in that, A non-contact visual temperature detection method for foggy environments as described in any one of claims 1-8 includes a preparation module, an acquisition module, a defogging module, a segmentation module, and a mapping module connected in sequence. The preparation module is used to set the target point for the high-temperature target to be measured according to the preset temperature range and preset temperature interval, and to determine the temperature point to be measured. The acquisition module is used to acquire conditional water mist temperature images of the temperature points to be measured based on preset water mist concentrations. The dehazing module is used to input the conditional water mist temperature image into the trained adversarial dehazing model for dehazing processing and generate a dehazed temperature image. The segmentation module is used to extract high-temperature pixels from the dehazed temperature image using an image segmentation algorithm, resulting in high-temperature region pixel blocks. The mapping module is used to input the pixel blocks of high-temperature regions into the trained light image-temperature field mapping model for mapping processing, so as to obtain the temperature data of the high-temperature target to be measured.

Images