Safety production infrared image intelligent sensing detection method based on noise self-supervision

Abstract

The invention relates to the technical field of computer vision and intelligent perception, in particular to a safety production infrared image intelligent perception detection method based on noise self-supervision. According to the method, a training set is constructed by integrating a multi-source infrared image data set so as to simulate data distribution differences in a real environment, a deep learning model based on a U-Net architecture is constructed, and the model integrates an attention feature fusion module and a noise guide representation learning module, so that the learning efficiency is improved. According to the method, adaptive fusion of multi-scale features and learning of noise insensitive features are respectively realized, system evaluation is performed on the performance of the model by adopting indexes such as intersection-to-parallel ratio, accuracy and false alarm rate, and the optimized model is deployed in a safety production monitoring system. And meanwhile, the sensing precision of the edge contour of the small target is improved, direct adaptation of a single model to multiple scenes is realized, and the generalization ability and the detection precision of the model in a complex noise environment are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and intelligent perception, in particular to a safety production infrared image intelligent perception detection method based on noise self-supervision. BACKGROUND

[0002] Infrared small target detection technology is a key research direction in the field of infrared imaging and computer vision, and has important value in safety production monitoring, environmental monitoring and other applications. Due to the low contrast, complex background, and poor signal-to-noise ratio of infrared images, combined with the fact that small targets usually occupy only a few pixels and lack obvious shape and texture features, the detection task faces severe challenges.

[0003] At present, infrared small target detection methods can be mainly divided into two categories: traditional image processing methods and deep learning-based methods. Traditional image processing methods are mostly based on manually designed feature extraction and background modeling strategies. For example, methods based on the local contrast mechanism of the human visual system enhance the difference between the target and the background by simulating the perception characteristics of the human eye for salient targets; methods based on filtering and background suppression attempt to separate the target by constructing a background estimation model. However, such methods have poor adaptability to complex scenes and diverse noise, and are difficult to cope with fluctuations in image quality caused by sensor differences and environmental changes in actual applications, resulting in insufficient generalization ability and high false alarm rate when deployed across scenes.

[0004] Deep learning-based methods have made significant progress in recent years. The encoder-decoder architecture represented by U-Net has become the mainstream framework for infrared small target detection due to its ability to effectively fuse multi-scale features. To further improve performance, researchers have introduced attention mechanisms, feature pyramids, Transformers, and other modules. For example, the local contrast attention network improves small target detection by enhancing the response strength of the target region; the densely nested attention network enhances the model's ability to capture subtle features through iterative feature fusion; and another study introduces a spatial-channel cross-attention mechanism into the Transformer module to establish long-range dependencies and improve global semantic understanding.

[0005] Although these methods have achieved good results on specific datasets, there are still the following outstanding problems in actual cross-scene and multi-device deployment: (1) Poor domain adaptation ability: Due to differences in infrared imaging device models, working environments, and climate conditions, infrared images from different sources have significant differences in data distribution. Traditional deep learning models trained on a single dataset perform poorly when faced with distributionally different test data, making it difficult to meet the requirements for model generalization in actual applications.

[0006] (2) Insufficient noise robustness: The inherent thermal noise of infrared images, sensor noise, and atmospheric disturbances will seriously affect the detection accuracy of small targets. Most existing methods only optimize the detection performance on clear samples, lack of specialized modeling of real noise patterns, leading to high false alarm rate and poor detection stability in complex noise environments.

[0007] (3) Limitations of existing improvement schemes: Although some current improvement methods such as data augmentation and multi-scale fusion can improve model performance to some extent, most of them still remain at the surface feature enhancement level and fail to fundamentally solve the problem of inconsistent cross-domain distribution and noise interference. In particular, there is a lack of noise modeling capability for the physical characteristics of infrared images and effective constraint mechanism for cross-domain feature consistency.

[0008] Therefore, there is an urgent need for an infrared small target detection method that can effectively overcome domain bias, improve noise robustness, and have strong generalization ability to meet the high standard requirements for detection model stability and adaptability in safety production and other practical application scenarios. SUMMARY

[0009] In view of the problem that infrared images are easily affected by factors such as thermal noise and background interference during acquisition, the present application provides a safety production infrared image intelligent perception detection method based on noise self-supervision.

[0010] To achieve the above purpose, the technical scheme adopted by the present application is as follows: The present application discloses a safety production infrared image intelligent perception detection method based on noise self-supervision, comprising the following steps: Data preparation step: obtain an infrared image training set, which includes data sets from multiple sources to simulate different data distribution environments; Model training step: construct a deep learning model based on U-Net architecture, which includes an attention feature fusion module and a noise-guided representation learning module; wherein the attention feature fusion module is embedded in the decoding stage of the U-Net model, used for adaptive fusion of multi-scale features to enhance target feature expression and suppress background interference; the noise-guided representation learning module adds noise to the training data and uses feature consistency constraints to make the model learn noise-insensitive features; Model testing step: use the test set to evaluate the model performance, use the intersection over union IoU, accuracy Pd and false alarm rate Fa as evaluation indicators to verify the superiority of the model in target shape description and positioning accuracy; Perception detection step: deploy the trained and verified model in the safety production monitoring system, use the model to detect small targets in real-time collected or existing safety production scene infrared images, output the segmentation or positioning results of the target, and realize safety monitoring and early warning.

[0011] Further, the attention feature fusion module realizes multi-scale feature fusion in the following way: extracting multi-level feature maps from the U-Net, wherein the low-level features contain target edge and texture details, and the high-level features contain target global contour semantic information; performing channel unification and layer normalization processing on the multi-level feature maps; using an attention mechanism to weight and fuse the normalized features, and superimposing the fused features and the initial input features through a residual connection to output the final prediction features.

[0012] Further, the multi-level feature maps are represented as ; wherein , H, W, and C represent the height, width, and channel number of the input picture respectively; i represents the layer number; represents the real space.

[0013] Further, the multi-level feature maps are patch feature embedded to obtain a group of embedded layer features ; wherein , and ; is the channel number of the corresponding layer; The channel unification unifies the channel number of the input feature map through 1x1 convolution and 3x3 convolution, and the specific calculation is as follows: , wherein represents layer normalization; is the spliced output; represents the i-th embedded layer feature after layer normalization processing.

[0014] Further, the specific calculation of the attention mechanism fusion includes: , , , wherein represents the query vector; and represent and convolution respectively; represents a complementary feedforward network; represents the output feature after feature attention mechanism fusion; represents the key vector; represents the transpose of the value vector; This represents the weights of the 1×1 convolution kernel used to generate the key and value vectors; Represents the weights of the 3×3 convolution kernel used to generate the key and value vectors; Represents the activation function; Representative spliced ​​feature map Total number of channels; This represents the transpose of the key vector.

[0015] Furthermore, the noise-guided representation learning module includes: to training samples Add Gaussian noise to generate noise samples ;in, It has a mean of 0 and a variance of Gaussian thermal noise, For hyperparameters; training samples With noise samples Inputting a U-Net network with shared weights, we obtain the predicted features of clean samples. and noise sample prediction features ; Calculate the feature consistency loss: ; The feature consistency loss is combined with the main loss function to form the total loss function, which is: ;in, It is a binary cross-entropy loss; These are sample labels; It is the weighting coefficient.

[0016] Furthermore, the parameters of the deep learning model are optimized by minimizing the total loss function, specifically: or,

[0017] in, Indicates the final model parameters; This indicates the number of training samples.

[0018] Furthermore, the U-Net model uses ResNet50 as the backbone network and sets the number of downsampling layers to 4.

[0019] Furthermore, during model training, the segmentation threshold was set to 0.5, the initial learning rate was set to 0.001, the batch size was set to 8, and the training period was set to 1000.

[0020] Furthermore, the model training process is optimized using a binary cross-entropy loss function, accelerated by the Adam optimizer, and the model parameters are initialized using a weight initialization method.

[0021] This invention addresses the technical deficiencies in the prior art and offers the following advantages: Through noise-guided self-supervised learning and attention feature fusion mechanisms, this invention enables the model to achieve strong anti-interference capabilities and cross-domain generalization. It maintains stable detection performance in complex environments such as dust and temperature / humidity variations, while improving the perception accuracy of small target edge contours, achieving direct adaptation of a single model to multiple scenarios. This not only reduces reliance on high-end infrared sensors and saves hardware costs, but also significantly reduces data annotation requirements through the self-supervised learning paradigm, lowering deployment and maintenance costs. This facilitates the widespread adoption of unmanned intelligent monitoring in high-risk areas, effectively ensuring personnel safety, promoting the intelligent transformation of high-risk industries, and supporting energy conservation and emission reduction for enterprises through precise equipment status monitoring, thus contributing to green industrial development. Attached Figure Description

[0022] 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 some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0023] Figure 1 Flowchart of a noise-self-supervised infrared image intelligent sensing and detection method for safe production; Figure 2 A framework diagram of a noise-self-supervised infrared image intelligent sensing and detection method for safe production; Figure 3 A framework diagram for attention feature fusion; Figure 4 A diagram illustrating a noise-guided high-resolution self-supervised learning process; Figure 5 A comparison chart showing the detection results of noise prediction and no-noise prediction; Figure 6 A comparison chart of qualitative recall metrics for different methods on the RealScene-ISTD and IRSTD-1K datasets. Detailed Implementation

[0024] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] like Figure 1 The present invention discloses a noise-self-supervised intelligent sensing and detection method for infrared images in safe production, comprising the following steps: Data preparation steps: Obtain an infrared image training set, which includes datasets from multiple sources to simulate different data distribution environments; It should be noted that, to simulate the complex and varied data distribution environment in the real world, a cross-dataset setup was adopted for model training and validation. Specifically, public datasets of infrared small targets from different sources with different characteristics were selected, such as IRSTD-1K and RealScene-ISTD. These datasets were mixed and then randomly divided into training, validation, and test sets according to a certain ratio (e.g., 8:1:1). The purpose of this step is to construct a training environment with diverse data distribution, thereby prompting the model to learn more generalizable features from the source and laying the foundation for subsequent evaluation of the model's cross-domain performance.

[0027] Model training steps: Construct a deep learning model based on the U-Net architecture. The deep learning model includes an attention feature fusion module and a noise-guided representation learning module. The attention feature fusion module is embedded in the decoding stage of the U-Net model and is used to adaptively fuse multi-scale features to enhance the target feature representation and suppress background interference. The noise-guided representation learning module adds noise to the training data and uses feature consistency constraints to enable the model to learn noise-insensitive features. It should be noted that the U-Net model with multi-layer attention feature fusion was used as the baseline during training, and ResNets50 was used as its backbone network. The number of downsampling layers was set to 4. The model was optimized using the binary cross-entropy loss function, and training was accelerated using the Adam optimizer. The Kaiming initialization method was used to initialize the model's weights and biases. Furthermore, the segmentation threshold was set to 0.5, the initial learning rate to 0.001, the batch size to 8, and the training epochs to 1000. All experiments were performed on a computer equipped with an NVIDIA TitanXp graphics card.

[0028] Model testing steps: Use the test set to evaluate the model performance, and use the intersection-over-union ratio (IoU), accuracy (Pd), and false alarm rate (Fa) as evaluation metrics to verify the model's superiority in target shape description and positioning accuracy; It should be noted that, to comprehensively evaluate model performance, the Intersection over Union (IoU) is used to measure the model's ability to describe the shape of the target, while accuracy (Pd) and false alarm rate (Fa) are used to evaluate the model's localization accuracy. IoU effectively reflects the degree of overlap between the predicted bounding box and the ground truth bounding box, while Pd and Fa measure the proportion of the model that correctly detects the target and the proportion that incorrectly detects the background, respectively.

[0029] Perception and detection steps: Deploy the trained and validated model in the safety production monitoring system, use the model to perform small target detection on real-time acquired or existing infrared images of safety production scenarios, and output the segmentation or localization results of the targets to achieve safety monitoring and early warning.

[0030] It should be noted that the trained and validated model is integrated into the safety production monitoring system. The model is used to perform real-time small target detection (such as high-temperature hotspots, unauthorized intruders, and equipment malfunctions) on infrared video streams or images from real-time acquired or existing safety production scenarios (such as oil fields, chemical plants, and substations). The model outputs segmentation masks or bounding boxes for the targets, and the system triggers an early warning mechanism accordingly, thereby achieving intelligent and highly robust safety monitoring.

[0031] Furthermore, the attention feature fusion module achieves multi-scale feature fusion in the following way: Multi-level feature maps are extracted from U-Net, where low-level features contain target edges and texture details, and high-level features contain global contour semantic information of the target. The multi-level feature maps are subjected to channel unification and layer normalization processing; The normalized features are weighted and fused using an attention mechanism, and the fused features are superimposed on the initial input features through residual connections to output the final predicted features.

[0032] Furthermore, the multi-level feature map is represented as follows: ;in, H, W, and C represent the height, width, and number of channels of the input image, respectively; i represents the number of layers. It represents the space of real numbers.

[0033] Furthermore, the multi-level feature maps Perform patch feature embedding to obtain a set of embedding layer features. ;in ,and ; This represents the number of channels in the corresponding layer. The number of channels in the input feature map is uniformly determined by 1×1 convolution and 3×3 convolution, specifically calculated as follows: , in, Representation layer normalization; It is the output of splicing; This represents the i-th embedding layer feature after layer normalization.

[0034] Furthermore, the specific computation of the attention mechanism fusion includes: , , , in, Represents the query vector; and Represent and convolution; This represents a complementary feedforward network; This represents the output features after fusion using the feature attention mechanism; Represents the key vector; Represents the transpose of a value vector; This represents the weights of the 1×1 convolution kernel used to generate the key and value vectors; Represents the weights of the 3×3 convolution kernel used to generate the key and value vectors; Represents the activation function; Representative spliced ​​feature map Total number of channels; This represents the transpose of the key vector.

[0035] Furthermore, the noise-guided representation learning module includes: to training samples Add Gaussian noise to generate noise samples ;in, It has a mean of 0 and a variance of Gaussian thermal noise, For hyperparameters; training samples With noise samples Inputting a U-Net network with shared weights, we obtain the predicted features of clean samples. and noise sample prediction features ; Calculate the feature consistency loss: ; The feature consistency loss is combined with the main loss function to form the total loss function, which is: ;in, It is a binary cross-entropy loss; These are sample labels; It is the weighting coefficient.

[0036] Furthermore, the parameters of the deep learning model are optimized by minimizing the total loss function, specifically: or, , in, Indicates the final model parameters; This indicates the number of training samples.

[0037] Furthermore, the U-Net model uses ResNet50 as the backbone network and sets the number of downsampling layers to 4.

[0038] Furthermore, the model training process is optimized using a binary cross-entropy loss function, accelerated by the Adam optimizer, and the model parameters are initialized using a weight initialization method.

[0039] like Figures 2 to 6 As shown, to improve the generalization performance of infrared small target detection models in cross-noise domain environments, this invention proposes a noise-guided self-supervised learning method based on a fully convolutional network (U-Net) with Feature Attention Fusion. Feature Attention Fusion and noise-guided representation learning are the two core modules, which optimize model performance from the perspectives of "strengthening feature representation" and "improving noise-resistant generalization," respectively. Both are deeply adapted to the characteristics of infrared small targets—"weak features and susceptible to noise interference"—and are tightly integrated with the U-Net infrastructure.

[0040] The attention feature fusion module is embedded in the decoding stage of U-Net. Its core function is to adaptively fuse multi-scale features to enhance target recognition and suppress background and noise interference. It obtains high-level features through ResNet-10 and a downsampling module based on max pooling layers. , in This corresponds to low-level detail features including target edges and textures. (In this embodiment, the number of downsampling layers is set to 4, so here) for This corresponds to high-order semantic features that contain the global contour of the target; then the high-order features are... Patch feature embedding obtained First, use 1×1 and 3×3 convolutions to unify the number of channels in all input feature maps to: , Then, the embedding layer features are fused using an attention mechanism: , , , The feature attention mechanism is used to eliminate the dimensionality barrier in subsequent fusion; then, the feature maps after channel unification are element-wise summed to obtain the preliminary fused feature prototype. ; By using residual connections to superimpose the weighted fused features with the initial input features, feature loss during the fusion process is avoided, and the final output is the fused predicted features. This output feature combines low-order detail positioning accuracy with high-order semantic recognition capability, providing strong representational support for accurate infrared small target detection.

[0041] Then, the noise-guided representation learning module addresses the inherent thermal noise and atmospheric clutter issues in infrared images. It guides the model to learn noise-insensitive core features by "actively adding real-world sensor acquisition noise + feature consistency constraints," while indirectly mitigating the domain shift effect of multi-source noise. Its initial training set... In this context, N represents the number of training images. Gaussian-distributed thermal noise is added to simulate real-world noise interference under different sensors and environments. The formula for constructing the noise samples is: , Then clean training samples training samples with noise Inputting the shared-weights U-Net feature extraction network into each sample yields feature maps at each level for both classes. Clean samples are then used to predict features. and noise sample prediction features Through the analysis of and Global feature calculation of MSE loss: , This loss is used as a constraint to guide the model to learn a representation where "the core features of the same target are consistent regardless of noise," forcing the model to actively filter noise disturbances and focus on the essential information of the target; at the same time, the loss of this module will be combined with the main loss of the detection task (binary classification cross-entropy loss). Combined to form the total loss function: , in This is the weight coefficient, with a default value of 1. Experiments have verified that this weight can balance the main task and noise-resistant learning. The model minimizes... and The mapping relationship is learned by using the feature space distance between them. This feature consistency helps constrain the model to learn basic features that are insensitive to noise. or,

[0042] By updating network parameters through backpropagation, the model maintains high detection accuracy and a low false alarm rate even in complex noisy environments and cross-domain scenarios. This high resolution and noise resistance allow the model to preserve spatial details while enhancing its discriminative power, thus helping to more accurately capture subtle radiometric differences between the target and the background. Furthermore, this characteristic naturally prompts the model to focus on target boundary regions, significantly improving the accuracy of edge-aware segmentation.

[0043] In summary, the present invention provides a noise-self-supervised intelligent sensing and detection method for safe production using infrared images: At the technical level, this invention fundamentally improves the practicality and reliability of the model in real industrial environments. Through a noise-guided self-supervised learning mechanism, the model is forced to learn essential features that are insensitive to various complex noises, thereby achieving excellent anti-interference and cross-domain generalization capabilities. This allows it to maintain high detection accuracy and low false alarm rate even under harsh conditions that degrade image quality, such as dust and temperature and humidity changes. Simultaneously, the attention feature fusion module, through adaptive weighting of multi-scale features, enhances the model's ability to perceive small targets and their edge contours, achieving more accurate target segmentation and localization. This allows the model to be deployed directly and efficiently in various safety production scenarios, such as high-temperature equipment monitoring, personnel intrusion detection, and hazardous material leaks, without cumbersome tuning, solving the adaptation dilemma of traditional methods that require "one model per scenario."

[0044] Furthermore, the model's inherent robustness reduces reliance on high-end, expensive infrared sensing hardware, allowing companies to deploy using existing conventional equipment and significantly saving on hardware purchase and upgrade costs. On the other hand, the self-supervised learning paradigm reduces reliance on massive amounts of accurately labeled data, avoiding expensive and time-consuming manual annotation work when migrating the model to new scenarios or devices. This results in a significant cost advantage throughout the project's lifecycle, especially during large-scale deployments.

[0045] Finally, the reliable and low-cost technical solution provided by this invention enables the widespread adoption of 24 / 7 unmanned intelligent monitoring in high-risk areas such as high-temperature, high-pressure, and high-dust environments, significantly reducing personnel safety risks and effectively improving the level of enterprise safety production assurance. This technological paradigm provides a feasible path for the intelligent transformation of high-risk industries such as power and chemicals, accelerating the shift from "manual inspection" to "intelligent early warning." Furthermore, by accurately monitoring equipment status and issuing early warnings of abnormal overheating, this invention also helps enterprises optimize energy use, achieve energy conservation and consumption reduction, and thus serve the national strategy of green industrial development.

[0046] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0047] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0048] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0049] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0051] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A safety production infrared image intelligent perception detection method based on noise self-supervision, characterized in that, The method comprises the following steps: a data preparation step: obtaining an infrared image training set, the training set comprising data sets from multiple sources to simulate different data distribution environments; a model training step: constructing a deep learning model based on a U-Net architecture, the deep learning model comprising an attention feature fusion module and a noise-guided representation learning module; wherein the attention feature fusion module is embedded in the decoding stage of the U-Net model, for adaptively fusing multi-scale features to enhance target feature expression and suppress background interference; the noise-guided representation learning module learns noise-insensitive features by adding noise to the training data and using feature consistency constraints; a model testing step: using a test set to evaluate the performance of the model, using the intersection over union (IoU), accuracy (Pd), and false alarm rate (Fa) as evaluation indicators to verify the superiority of the model in target shape description and positioning accuracy; a perception detection step: deploying the trained and verified model in a safety production monitoring system, using the model to detect small targets in real-time collected or existing safety production scene infrared images, outputting target segmentation or positioning results, and realizing safety monitoring and early warning.

2. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 1, characterized in that, The attention feature fusion module realizes multi-scale feature fusion in the following way: extracting multi-level feature maps from the U-Net, wherein the low-level features contain target edges and texture details, and the high-level features contain target global contour semantic information; performing channel unification and layer normalization on the multi-level feature maps; using an attention mechanism to weight and fuse the normalized features, and through residual connection, the fused features are superimposed with the initial input features to output the final prediction features.

3. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 2, wherein: The multi-level feature map is represented as ; wherein, , H, W, C represent the height, width and channel number of the input picture respectively; i represents the layer number; represents the real number space.

4. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 3, wherein: the multi-level feature map is obtained performing plaque feature embedding to obtain a group of embedding layer features ; wherein , and ; is the number of channels of the corresponding layer The channel unification unifies the number of channels of the input feature map through 1x1 convolution and 3x3 convolution, and the specific calculation is: ， wherein, denotes layer normalization; is the output of the concatenation; denotes the i-th embedding layer feature after layer normalization.

5. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 4, characterized in that, The specific calculation of the attention mechanism fusion includes: ， ， ， wherein, represents a query vector; and respectively represent and convolutions; denotes a complementary feed-forward network; denotes the output feature after the fusion of the feature attention mechanism; represents a key vector; represents the transpose of a value vector; denotes the weights of a 1x1 convolution kernel used to generate the key and value vectors; represents the weights of a 3x3 convolution kernel used to generate the key and value vectors; represents an activation function; represents the total number of channels of the feature map after concatenation; represents the transpose of a key vector.

6. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 1, characterized in that, The noise-guided representation learning module includes: To the training sample Add Gaussian noise to generate a noisy sample ; wherein, is a Gaussian thermal noise with mean 0 and variance , is a hyperparameter; The training sample is input into a U-Net network sharing weights to obtain clean sample predicted features and noise sample predicted features The clean sample predicted features and the noise sample predicted features are input into a U-Net network sharing weights to obtain clean sample predicted features and noise sample predicted features ​ Computing feature consistency loss: ; The feature consistency loss is combined with a main loss function to form a total loss function, and the total loss function is: ; wherein, is a binary cross-entropy loss; is a sample label; is a weight coefficient.

7. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 6, characterized in that, The parameters of the deep learning model are optimized by minimizing the total loss function, specifically: or, wherein, denotes the final model parameters; denotes the number of training samples.

8. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 1, characterized in that: The U-Net model uses ResNet50 as the backbone network, and the number of downsampling layers is set to 4.

9. The safety production infrared image intelligent perception detection method based on noise self-supervision of claim 1, characterized in that, In model training, the segmentation threshold is set to 0.5, the initial learning rate is set to 0.001, the batch size is set to 8, and the training period is set to 1000.

10. The safety production infrared image intelligent perception detection method based on noise self-supervision according to claim 1, characterized in that, The model training process uses a binary cross-entropy loss function for optimization, and the Adam optimizer is used to accelerate training, and the model parameters are initialized through a weight initialization method.

Images