Multi-scale expansion convolution attention temperature indicating paint interpretation method based on Gaussian window loss

Abstract

The invention discloses a multi-scale expansion convolution attention temperature indicating paint interpretation network based on Gaussian window loss, and aims to solve the problems of inconvenient operation, low precision and weak noise resistance of traditional temperature indicating paint interpretation and realize accurate measurement of surface temperature of aero-engine components. According to the method, RGB, HSV and Lab three-color space color gradation statistical features and mean value, median and mode class stability features are extracted and spliced into a 259 * 9 input feature tensor; a two-way multi-scale expansion convolution attention network (MDCA-Net) is constructed, channel association and feature long-range dependence are modeled in parallel, and key information is strengthened in combination with a learnable weight matrix; gaussian window loss (GW-Loss) is introduced, a discrete temperature label is converted into a Gaussian distribution soft label, temperature continuity is encoded, and overfitting is suppressed. Experiments show that on KN3A, KN6 and KN8 temperature indicating paint samples, the accuracy rates of the network respectively reach 97.58%, 94.25% and 98.25% when the error is equal to 0 DEG C, and the RMSE is 2.03 DEG C to 4.69 DEG C. The method has the advantages of automation, high precision and strong robustness, and provides a reliable engineering scheme for interpretation of the temperature indicating paint.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and intelligent temperature measurement, and particularly relates to a temperature indicating paint image automatic interpretation method based on deep learning. The method can be widely applied to high-precision measurement and automatic identification of surface temperature of devices such as aircraft engines, high-temperature structural parts and gas turbines. BACKGROUND

[0002] The accurate measurement of the surface temperature of hot end components such as aircraft engines during long-term operation is a key link to ensure safe operation and optimize thermal efficiency. As a high-temperature visual temperature measurement material, temperature indicating paint is widely used in the fields of gas turbines and aircraft engines due to its irreversible color change after heating. By observing the color change of the temperature indicating paint, the highest temperature experienced by the surface of the component can be intuitively inferred.

[0003] However, traditional temperature indicating paint interpretation methods mainly include manual observation and color mapping based on image processing. The manual method is greatly affected by subjective factors and has low resolution, making it difficult to meet the high-efficiency temperature measurement needs under complex working conditions. Although the automatic interpretation method based on images can improve the degree of automation to some extent, it still has shortcomings in noise interference, uneven lighting and color differentiation in the gradual temperature zone. When the color gradient is weak or the environmental noise is strong, the traditional algorithm is prone to interpretation errors. In addition, although the existing temperature interpretation method based on spectrum has high precision, it relies on expensive spectral instruments and contact measurement, making it difficult to achieve large-area automatic interpretation.

[0004] With the development of deep learning, convolutional neural networks have been applied to color mapping and image recognition tasks. However, conventional networks ignore the cooperative relationship between color spaces and the long-range dependence of feature dimensions, and the traditional cross-entropy loss function used in training fails to reflect the continuity between temperature levels, resulting in unstable model output and poor noise resistance.

[0005] To solve the above problems, the present application proposes a temperature indicating paint interpretation network based on multi-scale dilated convolution attention mechanism and Gaussian window loss, aiming to achieve high-precision and noise-resistant temperature interpretation of temperature indicating paint images. SUMMARY

[0006] The purpose of the present application is to provide a multi-scale dilated convolution attention temperature indicating paint interpretation method based on Gaussian window loss, which fuses multi-channel color features and stability features, introduces multi-scale dilated convolution structure and a learnable weight matrix, realizes cooperative modeling of cross-channel and long-range dependent features, and optimizes the model training process using a Gaussian window loss function, thereby improving interpretation accuracy and robustness.

[0007] In order to achieve the above object, the present application provides a multi-scale dilated convolution attention temperature indicating paint interpretation method based on Gaussian window loss, mainly including the following seven parts: the first part is the collection of temperature indicating paint data set; the second part is to extract multi-channel color features and stability features; the third part is to design a double-branch feature extraction structure, the first branch models cross-color space correlation in the channel dimension, and the second branch captures long-range feature dependence in the feature dimension; the fourth part is to use a multi-scale dilated convolution structure in the two branches of channels and features; the fifth part is to introduce a learnable weight matrix on the output results of the two branches, highlight the channel and feature positions that are crucial to temperature interpretation, and realize explicit attention guidance; the sixth part is to use a three-layer fully connected network to fuse the high-dimensional feature vectors of cross-channel correlation and long-range dependence; the seventh part is to use Gaussian loss to convert the real label into a Gaussian distribution soft label centered on the temperature level, and assign decreasing probability weights to adjacent temperature levels within a limited window range, so as to realize the probability supervision of temperature continuity, specifically: The first part includes three steps: Step 1, build a temperature indicating paint image acquisition platform, use D65 standard light source (color temperature 6500K), high-resolution industrial camera and horizontal workbench to form a standardized acquisition system, ensure uniform illumination and stable sample fixation; Step 2, image preprocessing, de-warping, geometric correction and center cropping are performed on the collected original image, the image is divided into 11x11 sample blocks to ensure sample consistency and resolution; Step 3, data set construction, the samples of different temperature levels are divided into training set, validation set and test set according to the ratio of 8:1:1, and a standardized temperature indicating paint data set is formed; The second part includes five steps: Step 4, color space conversion, convert the collected temperature indicating paint image to RGB, HSV and Lab three color spaces to provide a basis for multi-dimensional color feature extraction; Step 5, color scale statistical feature extraction, calculate the 256-dimensional normalized histogram of each channel in each color space to describe the color distribution; Step 6, stability feature extraction, calculate the mean, median and mode of the pixel values of each channel to reflect the stability of the image under illumination disturbance and noise; Step 7, feature splicing, splice the 256-dimensional color scale features of each channel with the 3-dimensional stability features to form a 259-dimensional channel vector; Step 8, multi-space fusion, splice the 9-channel features of RGB, HSV and Lab three color spaces to get a 9x259-dimensional feature matrix, which not only strengthens the color gradient difference, but also enhances the noise resistance; The third part includes three steps: Step 9, Input Definition: Input the 9×259 dimensional feature matrix into the network, treating it as a sequence of two different dimensions: feature dimension (9×259) and channel dimension (259×9). Step 10: Feature branch establishment. Construct the first branch in the feature dimension direction and use 1D convolution to capture the local and long-range dependency variation patterns to achieve hierarchical representation of the feature sequence. Step 11: Channel branch establishment. Construct a second branch in the channel dimension and use 1D convolution to extract collaborative features across color spaces, and model the channel correlation between RGB, HSV, and Lab. Part Four consists of two steps: Step 12, Feature branch convolution design: In the feature branches, an alternating convolution stack structure with an inflation rate of [1,3,1,3] is used to capture the dependencies between nearest and far-reaching features; Step 13, Channel Branch Convolution Design: In the channel branch, a combination of two 1D convolution layers with an dilation rate of [1,3] is used to extract cross-spatial correlation information. Part Five consists of three steps: Step 14: Define the weight matrix by introducing learnable weight matrices into the outputs of the channel branch and the feature branch, respectively. ; Step 15: Point-by-point weighting mechanism, perform point-by-point multiplication and weighting on the features of each output position (i,j) to highlight key channels and features; Step 16, Adaptive learning: The weight matrix parameters are automatically updated through backpropagation during training to achieve adaptive reinforcement of key feature locations. Part VI consists of five steps: Step 17, Feature Fusion: The outputs of the feature branches and channel branches after attention weighting are concatenated to form a fused feature matrix; Step 18: Flattening process. The fused features are flattened to form a one-dimensional feature vector for input into the classifier. Step 19, Fully Connected Layer Design: Set up a three-layer fully connected structure: the first layer outputs 2048 elements, the second layer outputs 512 elements, and the third layer is the output layer, corresponding to the number of temperature levels N, and uses the Softmax function to output the probability distribution; Step 20, Regularization and Activation: BatchNorm and Dropout layers are added after the output layer, respectively. The activation function is ReLU to prevent overfitting and improve nonlinear expressive power. Step 21, Temperature prediction output: The classifier outputs the predicted probability score for each temperature level, and the maximum probability method is used to determine the temperature level. Part VII consists of four steps: Step 22, loss function definition: replace the traditional cross-entropy loss with Gaussian window loss to encode temperature continuity; Step 23: Normalization process. Normalize the weights within all windows to ensure that their sum is 1, forming Gaussian distributed soft labels. Step 24, Loss Calculation: Calculate the network prediction probability. With soft tags Cross-entropy loss between: ; Step 25, Training and Optimization: The Adam optimizer is used for training. The batch size is set to 8, the learning rate is fixed, and the training is conducted for 200 rounds to prevent learning rate fluctuations from affecting convergence.

[0008] This invention presents a multi-scale dilated convolutional attention-based method for interpreting temperature-indicating paint samples based on Gaussian window loss. First, standardized images of temperature-indicating paint samples are acquired, creating an 11×11 image dataset. Then, the acquired paint images are converted to RGB, HSV, and Lab color spaces, and multi-channel color and stability features are extracted to form a 9×259-dimensional input feature matrix. Next, a dual-branch feature extraction structure is designed, where the first branch models cross-color space collaborative relationships in the channel dimension, and the second branch captures long-range feature dependencies in the feature dimension. Subsequently, multi-scale dilated convolutional structures are introduced into both the channel and feature branches, expanding the receptive field through combinations of [1,3] and [1,3,1,3] dilation rates to achieve multi-scale fusion of local and global features. Learnable weight matrices are introduced into the outputs of both branches to weight channels and feature positions point-by-point, thereby highlighting information crucial for temperature interpretation and achieving explicit attention guidance. The attention-weighted results are spliced ​​and fused, and a three-layer fully connected network is used to output the predicted scores for each temperature level, achieving automatic temperature interpretation. Finally, a Gaussian window loss function is used to transform the true labels into Gaussian distributed soft labels centered on the temperature level. Decreasing probability weights are assigned to neighboring temperature levels within a limited window, thereby achieving continuity modeling and smoothing supervision between temperature levels. This invention employs a multi-scale dilated convolution structure to better capture local details and global dependencies in temperature-indicating paint images. Through a learnable weight matrix and attention mechanism, this invention effectively enhances the expressive power of feature and channel dimensions. This invention uses a Gaussian loss function to replace the traditional discrete cross-entropy, explicitly expressing the physical continuity of temperature levels, thus significantly improving the model's interpretation accuracy and noise robustness. This invention can achieve high-precision temperature interpretation of temperature-indicating paint under complex lighting and noise conditions, possessing high engineering feasibility and application promotion value. Attached Figure Description

[0009] Figure 1 This is a schematic diagram showing the collection of standard samples for temperature-indicating paint. Figure 2 This is a diagram of the overall network framework of the present invention; Figure 3 The image shows the temperature interpretation accuracy and root mean square error distribution of MDCA-Net on the KN3A dataset. Figure 4 The training loss curves for Gaussian loss and cross-entropy loss are shown. Figure 5 Box plots of the root mean square error distribution of MDCA-Net temperature interpretation on the KN3A, KN6, and KN8 datasets; Figure 6 The figure shows the accuracy distribution of MDCA-Net on samples with added Gaussian noise when the errors are 0°C, 10°C, 20°C, and 30°C. Detailed Implementation

[0010] To better understand this invention, a multi-scale dilated convolutional attention-based method for interpreting temperature-sensitive paint based on Gaussian window loss is described in more detail below with reference to specific embodiments. In the following description, detailed descriptions of existing technologies may obscure the subject matter of this invention, and these descriptions will be omitted here.

[0011] Step 1, setting up the data acquisition device and photographing the temperature-indicating paint sample, is implemented as follows: Step 1.1: Set up a standard data acquisition platform in the laboratory or on-site. Figure 1 ): Includes a D65 standard light source (color temperature 6500K), a high-resolution camera and a horizontal worktable. The light source is arranged symmetrically on the left and right, and the camera is fixed perpendicular to the sample surface to ensure front lighting and no shadow shooting. Step 1.2: Heat the temperature-indicating paint sample according to the preset temperature sequence and take a picture to record the temperature when it stabilizes; collect the corresponding original image for each temperature point (used to build training / validation / test samples); record the metadata such as the labeled temperature, shooting parameters, and sample number of each image for traceability; Step 1.3, Saving Format and Quality Control: Images are saved in a lossless manner, and shooting parameters (exposure, white balance, gain) are fixed to ensure consistency in subsequent color analysis; Step 2, color space conversion and multi-channel feature fusion, is implemented as follows: Step 2.1, Color Space Conversion: Using the original image output from Step 1 as input, convert each 11×11 patch from RGB to HSV and Lab color spaces to obtain a total of 9 channels of data in three color spaces (R, G, B, H, S, V, L, a, b). Step 2.2, Color Scale Statistics: Calculate the normalized histogram for each channel to obtain a 256-dimensional color scale statistical vector; Step 2.3, Stability Statistics: Calculate the mean, median, and mode for each channel to obtain three-dimensional stability features; Step 2.4, Channel Vector Construction: The 256-dimensional histogram of each channel is concatenated with the 3-dimensional stability statistics to form a 259-dimensional channel vector; Step 2.5, Multi-spatial Fusion: Stack the 259-dimensional vectors of the 9 channels according to the channels to obtain the input feature matrix. Standardize X to facilitate network training; Step 3, input the definition and the two-branch parallel structure, the specific implementation is as follows: Step 3.1, Input definition: Take the standardized input feature matrix X output in Step 2 as input, and regard X as two sequences that can be processed in parallel. The channel sequence length is 9 and the feature sequence length is 259. Step 3.2, Parallel construction of two branches: The feature branch applies a 1D dilated convolution chain along the feature dimension (length 259) to capture local and long-range feature dependencies, and the channel branch applies a 1D convolution along the channel dimension (length 9) to model the collaborative relationship between color spaces; Step 4, the specific design and implementation of multi-scale dilated convolution are as follows: Step 4.1, Feature Branch Dilated Convolution: 1D convolutions with alternating dilation rate sequences [1,3,1,3] are used, which stack to form a larger equivalent receptive field to capture long-range dependencies; Step 4.2, Channel Branch Dilated Convolution: A two-layer 1D convolution with a dilation rate sequence [1,3] is used to simultaneously capture neighboring channel interactions and long-range relationships across channels; Step 5: Introduce a learnable point-by-point attention weight matrix onto the output of Step 4. The specific implementation is as follows: Step 5.1, Weight Matrix Definition: Define learnable matrices on the outputs of both branches respectively. Used for positional weighting; Step 5.2, point-by-point weighting mechanism: Perform point-by-point multiplication on each bit (channel position i, feature position j). The multiplication is a scalar broadcast scaling of the vector, highlighting the response of key channels and feature positions; Step 5.3, Weight Learning and Constraints: The weight matrix W and other network parameters are updated simultaneously through backpropagation; Step 6, feature fusion, flattening, and a three-layer fully connected classifier, are implemented as follows: Step 6.1, Feature concatenation: Using the weighted output features of the two branches from Step 5 as input, concatenate them along the specified dimension to obtain the fused feature matrix; Step 6.2, Flattening: Expand the fused matrix into a one-dimensional feature vector in a row-major or column-major manner so that it can be used as the input of the fully connected network; Step 6.3, Construction of a three-layer fully connected structure: A three-layer fully connected network is used to predict the temperature level. The first layer outputs 2048 neurons, the second layer outputs 512 neurons, and the third layer is the output layer. The number of output units corresponds to the total number of temperature levels. The probability of each temperature level is output through the Softmax activation function. Step 6.4, Regularization and Activation: Batch normalization and random deactivation are added between fully connected layers to reduce the risk of overfitting and improve the model's generalization ability. The activation function is ReLU to enhance the non-linear expressive power of the network. Step 6.5: Obtain the probability distribution of each temperature level through the output layer, and use the temperature level with the highest probability as the final temperature interpretation result of the model. Step 7, the construction and training strategy of Gaussian window loss, is implemented as follows: Step 7.1, Gaussian soft label construction: The discrete real temperature level labels are transformed into a soft label centered on the real temperature level. The soft label is assigned a higher probability near the real temperature level and a lower probability at more distant levels to establish a continuous relationship between temperature levels. The soft label is only effective within a certain window range, such as several levels near the real temperature level, and the probability of other levels is set to zero or ignored. Step 7.2, Soft label normalization: Rescale all probabilities of the above labels proportionally so that their sum equals 1, thereby forming a reasonable probability distribution, which serves as a supervision signal during the training process; Step 7.3, Loss Calculation: The cross-entropy loss between the Gaussian soft labels and the model output probabilities is used as the training objective, and compared with... Figure 4 The ordinary cross-entropy loss curve, GW-Loss, does not have the phenomenon of "approaching zero infinitely", the decline is smooth and there is no generalization problem caused by excessive convergence, and its advantages are significant. It distributes weights according to the temperature level distance, encodes physical continuity, and has both classification and regression smoothness tolerance, which fits the actual temperature changes and helps the model learn the temperature level related characteristics. Step 7.4, Training and Optimization Strategy: Use the Adam optimizer to train the model. The recommended batch size is 8, and the number of training rounds is about 200. During training, keep the learning rate constant or decay it according to the preset strategy. All model parameters, including the learnable weight matrix, are updated by backpropagation. Record the accuracy and error metrics during the training process to select the model with the best performance. Step 8, Model Temperature Test: Verify the temperature interpretation performance of the proposed multi-scale dilated convolutional attention-based temperature paint interpretation network based on Gaussian window loss under different datasets, different noise conditions, and different temperature ranges. The specific implementation is as follows: Step 8.1, Test Data Preparation: After training, input the temperature-indicating paint samples from the test set into the model for temperature prediction. The test data is obtained from a standard acquisition platform, and its light source, camera parameters, and sample positions are consistent with... Figure 1 The platform shown is consistent, ensuring that the distribution of the testing data is consistent with that of the training data.

[0012] Step 8.2, Temperature Prediction Execution: The test image is processed according to the same procedure as in the training phase, including color space conversion, multi-channel feature construction, and feature matrix formation. Subsequently, the feature matrix is ​​input into the network model of this invention. Through a dual-branch feature extraction module, a multi-scale dilated convolutional structure, and an attention weighting mechanism with learnable weight matrices, a fused high-dimensional feature vector is obtained. This vector is then processed through a three-layer fully connected network to output the predicted probabilities for all temperature levels. The final temperature prediction result is the temperature level corresponding to the highest probability. The testing process corresponds to... Figure 2 The network structure diagram is shown below; Step 8.3 involves a quantitative evaluation of the model's temperature prediction performance, primarily comprising three indicators: exact hit accuracy (the proportion of samples with a statistical prediction error of 0℃), tolerance accuracy (the proportion of samples with statistical prediction errors within 10℃, 20℃, and 30℃), reflecting the model's fault tolerance capability. Figure 6 The results demonstrate the accuracy changes of the model under different allowable error ranges after adding Gaussian noise of varying intensities, showcasing its strong robustness. The root mean square error (RMSE) reflects the overall error level by statistically analyzing the temperature deviations of the model across various datasets, including KN3A, KN6, and KN8. Figure 3 The accuracy and error distribution on the KN3A dataset are presented. Figure 5 The error box plots for different datasets clearly show the interval distribution characteristics of the interpretation error. This step aims to demonstrate the transferability and generalization ability of the present invention in different temperature-indicating paint models and different temperature ranges. This invention addresses the problems of large subjective errors and poor noise resistance in traditional temperature-indicating paint interpretation, as well as the neglect of cross-color space collaboration and long-range feature dependencies by conventional deep learning models. It proposes a multi-scale dilated convolutional attention-based temperature-indicating paint interpretation method based on Gaussian window loss. The method first acquires temperature-indicating paint images through a standardized platform, converts them to RGB, HSV, and Lab color spaces, and constructs a multi-dimensional input matrix by fusing color level statistics and stability features. Then, a dual-branch parallel structure is designed, coupled with multi-scale dilated convolution (channel branch dilation rate [1,3], feature branch dilation rate [1,3,1,3]), to efficiently capture cross-space collaboration relationships and long-range dependencies. Subsequently, a learnable point-by-point attention weight matrix is ​​introduced to highlight key features. Finally, a Gaussian window loss function is used to model temperature continuity, and a three-layer fully connected classifier is used for optimized training. This invention balances computational efficiency and interpretation performance, significantly improving interpretation accuracy, noise resistance, and generalization ability under complex working conditions, and is suitable for high-precision non-contact temperature measurement of high-temperature equipment such as aero-engines.

[0013] Although specific embodiments of the present invention have been described above, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes will be obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

Claims

1. A multi-scale dilated convolutional attention-based method for interpreting paint temperature readings based on Gaussian window loss, characterized in that... By standardizing the acquisition of temperature-indicating paint images and constructing a dataset, channel features and stability features from multiple color spaces are fused together. This method employs a dual-branch, multi-scale dilated convolutional structure to capture cross-spatial collaborative relationships and long-range feature dependencies. A learnable attention weight matrix is ​​introduced to enhance key information. After feature fusion via a fully connected network, a Gaussian window loss function is used to model the continuity of temperature-indicating paint color changes to optimize training. This achieves high-precision, noise-resistant temperature interpretation of temperature-indicating paint images under complex conditions. The method comprises seven parts: temperature-indicating paint dataset acquisition, multi-channel color and stability feature extraction, dual-branch feature extraction structure design, application of a multi-scale dilated convolutional structure, introduction of a learnable weight matrix, fusion via a three-layer fully connected network, and optimization of the Gaussian window loss function. The data collection for the temperature-indicating paint dataset includes three steps: Step 1: Set up a standardized acquisition system consisting of a D65 standard light source, a high-resolution camera, and a horizontal worktable to ensure uniform illumination and stable sample mounting. Step 2: Perform distortion correction, geometric correction, and center cropping on the acquired original image to divide the image into sample blocks of size 11×11; Step 3: Divide the samples of different temperature levels into training set, validation set and test set in a ratio of 8:1:1 to form a standardized temperature-indicating paint dataset; The multi-channel color feature and stability feature extraction part includes three steps: Step 4: Convert the acquired temperature-indicating paint image to three color spaces: RGB, HSV, and Lab. Step 5: Calculate a 256-dimensional normalized histogram for each channel of each color space to depict the color distribution; Step 6: Calculate the mean, median, and mode of pixel values ​​for each channel to extract stability features; Step 7: The 256-dimensional color level features of each channel are concatenated with the 3-dimensional stability features to form a 259-dimensional channel vector; Step 8: The nine channel features of the three color spaces are spliced ​​together to obtain a 9×259-dimensional feature matrix, realizing multi-space fusion; The design of the dual-branch feature extraction structure includes three steps: Step 9: Input the 9×259 dimensional feature matrix into the network, treating it as two different sequences: the feature dimension (9×259) and the channel dimension (259×9). Step 10: Construct the first branch in the feature dimension direction, and use 1D convolution to capture the local and long-range dependency variation patterns to achieve hierarchical representation of the feature sequence; Step 11: Construct a second branch in the channel dimension, use 1D convolution to extract cross-color space collaborative features, and model the channel relationship of the three color spaces; The application of multi-scale dilated convolutional structures includes three steps: Step 12: In the feature branch, an alternating convolution stack structure with an inflation rate of [1,3,1,3] is used to capture the dependencies between nearest and far-distance features; Step 13: In the channel branch, a combination of two 1D convolutional layers with an dilation rate of [1,3] is used to extract cross-spatial correlation information; The introduction of the learnable weight matrix includes three steps: Step 14: Introduce learnable weight matrices to the outputs of the channel branch and feature branch respectively. ; Step 15: Perform point-by-point multiplication and weighting on the features of each output position (i,j) to highlight key channels and features; Step 16: The weight matrix parameters are automatically updated through backpropagation during training to achieve adaptive enhancement of key feature positions; The convergence of the three-layer fully connected network includes three steps: Step 17: Concatenate the outputs of the feature branch and the channel branch after attention weighting to form a fused feature matrix; Step 18: Flatten the fused features to form a one-dimensional feature vector; Step 19: Construct a three-layer fully connected structure. The first layer outputs 2048 elements, the second layer outputs 512 elements, and the third layer is the output layer corresponding to the temperature level N. The Softmax function is used to output the probability distribution. Step 20: The predicted probability scores of each temperature level are output by the classifier, and the temperature level is determined by the maximum probability method. The optimization of the Gaussian window loss function includes three steps: Step 21: Replace the traditional cross-entropy loss with Gaussian window loss to encode temperature continuity; Step 22: Normalize the weights within all windows to ensure that the sum is 1, thus forming a Gaussian distributed soft label; Step 23: Calculate the cross-entropy loss between the network prediction probability ρ and the soft label w(k). ; Step 24: Train using the Adam optimizer with a batch size of 8, a fixed learning rate, and 200 training rounds.

2. The method for interpreting paint temperature indicators based on Gaussian window loss using multi-scale dilated convolutional attention as described in claim 1, characterized in that, In step 12, the feature branch adopts an alternating convolution stack structure with an inflation rate of [1,3,1,3]. In step 13, the channel branch adopts a combination of two 1D convolution layers with an inflation rate of [1,3] to capture long-range feature dependencies and cross-spatial correlation information, respectively.

3. The method for interpreting temperature-indicating paint based on multi-scale dilated convolutional attention using Gaussian window loss as described in claim 1, characterized in that... The learnable weight matrix W introduced in step 14 on the outputs of the channel branch and the feature branch x W y Dimension R 9 ˣ² 59 The key channels and features are highlighted by the point-by-point multiplication weighting in step 15.

4. The multi-scale dilated convolution attention-based method for interpreting temperature-indicating paint based on Gaussian window loss according to claim 1, characterized in that, In step 19, the first layer of the three-layer fully connected structure outputs 2048 units, the second layer outputs 512 units, and the third layer outputs the corresponding temperature level N and uses the Softmax function to output the probability distribution.

5. The method for interpreting paint temperature indicators based on Gaussian window loss using multi-scale dilated convolutional attention, as described in claim 1, is characterized in that... Step 22 normalizes the weights within the Gaussian window to form Gaussian distributed soft labels. Step 23 calculates the cross-entropy loss based on the soft labels w(k) and the network prediction probability ρ. Encoding temperature continuity.

Images