An image processing method, system, device and medium based on infrared thermal imaging

Abstract

The application discloses an image processing method, system, device and medium based on infrared thermal imaging, and belongs to the technical field of image processing. The method comprises the following steps: an expert experience PoE method is used to construct an expert field FoE model; the expert field FoE model is optimized and trained to obtain an optimized FoE model; a Gamma FoE denoising algorithm is constructed, a denoised image is defined based on an image posterior distribution, and an image processing model is constructed; based on the image processing model, an energy minimization problem is constructed by combining the optimized FoE model and a gamma distribution, and an optimal denoised image is solved; and the optimal denoised image is enhanced, and a processed infrared thermal imaging image is output. The Gamma FoE-GTHE algorithm constructed in the disclosure significantly improves the resolution of the infrared image, and compared with the traditional image processing technology, the accuracy of model recognition is improved.

Description

Technical Field

[0001] This application belongs to the field of image processing technology, and in particular relates to an image processing method, system, device and medium based on infrared thermal imaging. Background Technology

[0002] Due to the unique imaging principle of infrared thermal imaging, measurement results are easily affected by factors such as emissivity, environmental reflection, and atmospheric conditions (humidity, dust, distance), which may cause the image to appear somewhat blurry and make it difficult to distinguish very fine details. Interference from thermal noise and shot noise can manifest as graininess, spots, stripes, or unwanted pixel fluctuations in the image, leading to unclear images and inaccurate temperature measurements.

[0003] In the field of infrared thermal imaging image processing, existing technologies include common image processing methods such as image denoising and image enhancement. Since most existing image processing technologies are used to process optical images, the effect of using these algorithms to process infrared images is not very ideal.

[0004] Therefore, it is necessary to provide a new image processing method, system, device, and medium based on infrared thermal imaging to solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this disclosure is to provide an image processing method, system, device, and medium based on infrared thermal imaging in order to solve the above-mentioned problems.

[0006] This disclosure achieves the above objectives through the following technical solutions: An image processing method based on infrared thermal imaging includes the following steps: An expert domain FoE model is constructed using the expert experience PoE method. The expert domain FoE model is optimized and trained to obtain the optimized FoE model; Based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, a Gamma FoE denoising algorithm is constructed. The denoised image is defined based on the posterior distribution of the image, and an image processing model is constructed. Based on the image processing model, and combining the optimized FoE model and gamma distribution, an energy minimization problem is constructed to solve for the optimal denoised image. The optimal denoised image is enhanced to output the processed infrared thermal imaging image.

[0007] As a further optimization of this disclosure, an expert domain FoE model is constructed using the expert experience PoE method, including: Based on the expert experience PoE framework, the Student-t distribution is used as the expert function to fit the statistical characteristics of infrared images, thereby obtaining the image probability density function, which is then converted into the Gibbs distribution form according to the Hammersley-Clifford theorem. The PoE framework is extended to construct an expert domain FoE model. The neighborhood system is built with image pixels as graph nodes. The potential function of the Markov random field (MRF) is characterized by an expert function, thus completing the modeling of the global statistical properties of the image.

[0008] As a further optimization of this disclosure, the expert domain FoE model is optimized and trained to obtain an optimized FoE model, including: The optimization is achieved by updating the parameters of the expert domain FoE model through gradient updates, using hybrid Monte Carlo sampling to accelerate the sampling process, and using the EM algorithm to iteratively estimate the hidden variables to optimize the gamma distribution parameters, thus obtaining the optimized FoE model.

[0009] As a further optimization of this disclosure, a Gamma FoE denoising algorithm is constructed based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution. The denoised image is defined based on the posterior distribution of the image, and an image processing model is constructed, including: Using image pixels as observation points, a probability density model of a mixed gamma distribution is constructed; By introducing latent state variables to label the pixel source distribution, the joint conditional density function of the complete data is obtained; Using Bayesian maximum a posteriori probability (MAP) as the criterion, an objective function for infrared image denoising is established, resulting in an image processing model.

[0010] As a further optimization of this disclosure, the energy minimization problem integrates the prior energy of the expert domain FoE model with the gamma distribution likelihood energy, and achieves the optimal solution by iteratively updating the gradient.

[0011] As a further optimization of this disclosure, the optimal denoised image is enhanced by: The gradient value of each pixel in the image is calculated using the median difference method; Set a gradient threshold and retain pixels with gradients greater than the threshold for histogram statistics; Histogram equalization is performed on the selected pixels to complete image enhancement.

[0012] As a further optimization of this disclosure, the gradient threshold is adaptively adjusted according to the infrared image scene and noise level to remove background areas without information and retain the edge details of the target.

[0013] An image processing system based on infrared thermal imaging includes: The expert model building module is used to build expert domain FoE models using the expert experience PoE method. The model optimization module is used to optimize and train the expert domain FoE model to obtain an optimized FoE model. The image processing model construction module is used to construct a Gamma FoE denoising algorithm based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, define the denoised image based on the posterior distribution of the image, and construct the image processing model. The denoising solution module is used to construct an energy minimization problem based on the image processing model, combined with the optimized FoE model and gamma distribution, and solve for the optimal denoised image. An enhanced output module is used to enhance the optimal denoised image and output the processed infrared thermal imaging image.

[0014] An electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor is used to execute the program stored in the memory to implement the image processing method based on infrared thermal imaging.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the image processing method based on infrared thermal imaging.

[0016] The beneficial effects of this disclosure are as follows: The constructed Gamma FoE-GTHE algorithm significantly improves the resolution of infrared images and enhances the accuracy of model recognition compared to traditional image processing techniques. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method in an embodiment of this disclosure; Figure 2 This is a schematic diagram illustrating the implementation process of the expert domain model in an embodiment of this disclosure; Figure 3 This is a flowchart of the GTHE algorithm in an embodiment of this disclosure; Figure 4This is a system structure block diagram of an embodiment of this disclosure; Figure 5 This is a block diagram of the device structure in an embodiment of this disclosure. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 1 As shown, an image processing method based on infrared thermal imaging includes the following steps: S1. Construct an expert domain FoE model using the expert experience PoE method, including: Using the Product-of-Experts (PoE) approach, construct an expert domain model (Field of Experts (FoE): The expert domain model effectively simulates the potential function of a Markov Random Field (MRF) by integrating the combined effects of multiple expert distributions. This design allows the model to utilize a set of linear filters to describe the image response, thereby capturing the structural features within the image. Compared to traditional potential function models, the FoE model has the ability to handle a larger pixel range, making it particularly suitable for the analysis of high-dimensional image data. Furthermore, this model can extract higher-order prior knowledge from the training images, enabling the potential function to more accurately reflect the statistical properties and semantic information of the image.

[0021] The core idea of ​​the FoE model lies in using the joint action of multiple expert distributions to describe the probability distribution of local image features. These expert distributions are typically defined based on the responses of linear filters, with each expert capturing and emphasizing different characteristics in the image, such as edges, textures, or smooth regions. In this way, the model can more comprehensively express the local structural information of the image, and the FoE model achieves effective modeling of complex image patterns while preserving image details.

[0022] A significant advantage of the FoE model lies in its flexibility. Because the potential function is optimized through the training data, the FoE model can adapt to different image categories or task requirements. During training, the FoE model decomposes image patches using a pre-defined set of linear filters, improving the likelihood function of the training images to optimize each parameter of the potential value. This learning approach enhances the model's ability to represent complex aspects of images, enabling it to capture higher-level and broader-domain past data.

[0023] The specific implementation process of the expert domain model is as follows: Figure 2 As shown, to collect specific attributes, the provided image is first processed using block segmentation or filtering techniques. Then, an expert distribution is used to represent these local features to compute the feasible response of the function. Finally, a probabilistic model of the entire image is created by integrating all local responses. Because the FoE model can effectively interpret both global and local mathematical features of an image, it is frequently used in various fields of image denoising, reconstruction, and segmentation.

[0024] Based on the PoE framework, each expert processes the data in an independent linear subspace by mapping each local patch of the image onto a linear component: Considering that infrared images are usually accompanied by a long "tailwave" characteristic, the Student-t distribution with a heavier tailwave is chosen as the expert function to better model the statistical characteristics of infrared images, as shown below: (1); In the formula, It is the probability density function; The input image; This is the normalization function; This represents the set of parameters for all expert domain models; Indicates the first The output of a linear filter; where It is a linear filter. It is its transpose. Therefore It is a vector inner product, which is equivalent to the response output of the i-th linear filter applied to the image x. Indicates the shape of the distribution; For expert functions; Indicates the first A set of parameters for an expert domain model; Represents a linear filter; This represents the total number of expert domain models.

[0025] According to the Hammersley-Clifford theorem, formula (1) can be reformulated as a probability density function in Gibbs form, as shown below: (2); In the formula, express The energy function is as follows: (3).

[0026] The traditional PoE framework is extended to build a FoE model, which models the global statistical properties of images: By defining image pixels as nodes in a graph to construct a neighborhood system, the FoE model defines the neighborhood of each node as a maximal clique, and then describes the global structure of the image through a potential function; the global probability density is in the form shown below: (4); In the formula, The total number of pixels is represented; to represent the neighborhood of the i-th pixel, the FoE model uses an expert function to characterize the potential function of the Markov random field (MRF).

[0027] The role of expert functions is to help models more accurately capture the statistical regularities of images by responding to local image features (such as edges or textures). To more effectively describe the characteristics of radiative attenuation in a scene, expert functions are defined in the following form: (5); In the formula, Indicates the first The output of a linear filter; These are parameters of the expert function, used to control the response value. The weights. Therefore, formula (4) can be rewritten as: (6); In the formula, Let x represent the k-th image patch. The entire image x is divided into k overlapping or non-overlapping local patches. It refers to the k-th small block, which is usually represented as a vector.

[0028] S2. Optimize and train the expert domain FoE model to obtain an optimized FoE model, including: In infrared images, the FoE model can capture the correlation of dark values ​​in different image regions, making it suitable for complex geometries and noise distributions. The FoE model directly models the overlap of image patches. Considering that these patches are highly correlated, the training objective of the FoE model is to maximize the log-likelihood function of the training data, making the model distribution approximate the real data distribution.

[0029] Since the FoE model does not have a closed-form solution, its parameter optimization is achieved through the following methods: (7); In the formula, This represents the global probability density function of the FOE model. Indicates the learning rate; Represents the energy function For parameters The partial derivatives; the first term in parentheses. The second term in parentheses represents the average value of the training data. The gradient represents the expected value of the model distribution; P is the positive phase label, indicating that the gradient is calculated under the true data distribution; X is the negative phase label, indicating that the gradient is calculated under the model distribution.

[0030] Hybrid Monte Carlo (HMC) sampling is used to accelerate the sampling process, leveraging the gradient of the potential function to explore the sample space more efficiently. The EM algorithm is used to iteratively estimate the hidden state variables in the FoE model, thereby optimizing the parameters of the Gamma distribution.

[0031] S3. Based on the prior indices of the optimized FoE model and the likelihood function of the gamma distribution, construct a Gamma FoE denoising algorithm, define the denoised image based on the posterior distribution of the image, and construct an image processing model, including: When dealing with Gaussian noise, the noise estimate typically has a significant impact on Gaussian FoE denoising algorithms. If the noise variance cannot be accurately estimated and thousands of iterations are required over a long period, the denoising effect will be limited.

[0032] Therefore, a Gamma FoE denoising algorithm based on prior indices of expert domain models and likelihood functions of gamma distribution is proposed. This algorithm is based on the Gaussian FoE algorithm and incorporates the special characteristics of infrared images, using the posterior distribution as the definition basis for the denoised image to construct an accurate data model.

[0033] In the data model, observation points Represents the pixels in the image, where Iterate through all pixels. The probability density function of the Gamma distribution is shown below: (8); In the formula, Indicates that given model parameters Under the given conditions, the marginal probability of the j-th observation data Aj. Let be the conditional probability density function of the k-th component, representing that if the data point is known to come from the k-th component, then... The probability value of occurrence, These are the parameters of the component itself; Represents the observed value; It is a set of mixed variables. satisfy: (9); Is with parameters The relevant Gamma distribution function is shown below: (10); For the parameters of the k-th component At that time, the observed data The probability density. Let be the set of parameters for the k-th component. This explicitly includes two parameters: ={ , } For shape parameters, This is the scale parameter. The exponentially decaying term determines the long-tail behavior of the distribution; They are statistically independent, so The joint conditional density function is shown below: (11); Through matrices that cannot be visualized, i.e., hidden state variables , To increase the number of observation points .in This is called an indicator, as shown below: (12); By increasing the scale of the data, complete data can be obtained. The joint conditional density function is shown below: (13).

[0034] S4. Based on the image processing model, and combining the optimized FoE model and gamma distribution, construct an energy minimization problem to solve for the optimal denoised image, including: Mixed variables and hidden state variables The correlation coefficients of the gamma distributions are shown below: (14); (15); In the formula, Let represent the posterior probability that the j-th data point belongs to the k-th mixture component. This represents the parameter given the k-th component. Below, observed The gamma distribution probability density. The denominator is the sum of the weighted densities of all components, normalized to make... =1. The new mixture weight for the k-th component is equal to the average posterior probability of all data points for that component. This represents the current parameter set of the k-th component, which includes probability, mixing weights, shape parameters, and rate parameters. The update of the shape parameters is obtained using the likelihood equation. The updated formula uses the difference between the logarithm of the data and the digamma value as the numerator, and the total probability of that component as the denominator, used for weighted averaging. Similarly, This indicates an update to the rate parameter. It is the Gamma function; This represents the Digamma function (i.e., the logarithmic derivative of the Gamma function).

[0035] The posterior likelihood of an infrared image is defined entirely by the parameter likelihood estimates derived from the gamma distribution and the prior data of the FoE model. Given a noisy image, to remove the noise and recover the true image R, denoising is achieved by constructing a maximum a posteriori (MAP) estimate by maximizing the posterior probability.

[0036] First, the FoE model uses Bayesian rules to remove noise, as shown below: (16); In the formula, It represents the posterior probability of the true image R given the observed image S; It represents the likelihood function (i.e., the probability of observing S when the true image is R); Let S represent the prior probability of the true image R; S represent the observed image; R represent the true image; and the speckle noise infrared image follows a gamma distribution. The probability density of the observation process is as follows: (17); Then, the logarithmic estimation in the infrared image is as follows: (18); The first term in the formula, This represents the conditional probability density of observing the entire image given a reference image R. The logarithm is used to transform the product into a summation, facilitating numerical computation and optimization. Outer layer summation. Summing is performed on all N pixels in the image, assuming each pixel is conditionally independent. Inner layer summation. Summing the K mixture components indicates that each pixel may come from one of K different statistical models. Let the mixture weight (prior probability) of the k-th component satisfy the following condition: . This indicates that, given a reference pixel, under the k-th component... Observed pixels The conditional probability density.

[0037] The second term in the formula, This represents the intensity value of the j-th pixel in the currently observed image. This represents the intensity value of the j-th pixel in the reference image. This represents the shape parameter of the k-th gamma component. This represents the scaling parameter of the k-th gamma component. This represents the gamma function, used for normalization to ensure the probability density integral is 1. (Molecular) The term representing the core shape of the gamma distribution. Denominator This indicates that the intensity of the observed image is scaled proportionally to the reference image, which here refers to the reference image. As a scale modulation factor. Exponential term. This indicates the control of the decay rate, reflecting a decrease in the probability that the observed value deviates from the proportional value.

[0038] Finally, by simply manipulating the latent state variable Z, we solve the likelihood solution of the derivative equation that minimizes R, and obtain the energy minimization expression as follows: (19); In the formula, This indicates the search for the optimal reference image estimate. This minimizes the energy function E. For regularization terms, The number of filters or features. It is the weight coefficient of the i-th regularization term. Represents a penalty function used to apply... Apply smoothness or sparsity constraints. A linear operator, acting on Extract local features or gradients from an image. For the likelihood term, it is the negative log-likelihood from the mixed model. It is the Lagrange multiplier of the k-th mixture component, used to balance the effect of that component on energy. Let be the latent state variable vector of the k-th component. Let be the logarithm of the prior weights of the k-th mixture component. The logarithm of the reference image is used as the scale modulation factor. The negative term of the logarithm of the observed image, derived from the gamma distribution. The logarithm of a term is followed by a negative sign. It is the negative logarithmic form of the exponential term, derived from Take the logarithm. and Through Definite The penalty value.

[0039] The optimal solution is obtained based on the L-BFGS algorithm: Minimizing infrared images can be viewed as a series of optimization tasks, and the optimal solution can be obtained by applying the L-BFGS algorithm. The L-BFGS algorithm is a quasi-Newton technique that can efficiently solve optimization problems and has the advantages of numerical stability and fast convergence. By simply inputting a matching gradient function in each iteration, the L-BFGS algorithm can update the correlation matrix using the gradient vector.

[0040] therefore, The required gradient function is as follows: (20); In the formula, For the energy function E on the image The gradient is a property of... A vector or matrix of the same size represents the direction in which energy increases the fastest at each pixel. It is the number of filters. The weight coefficient of the i-th regularization term controls the contribution of this term to the gradient. A linear operator, acting on Extract local features, for The transpose of is used to map the derivative from the feature domain back to the image domain, ensuring dimension matching. (.) represents the penalty function. The first derivative, This represents the response after applying the i-th filter to the current estimated image.

[0041] S5. Enhance the optimal denoised image and output the processed infrared thermal imaging image, including: A histogram of a digital image is a discrete function used to describe the statistical relationship between gray values ​​and their probabilities in an image. The definition of a histogram is as follows: (twenty one); In the formula, This represents the total number of pixels in the image; Indicates the first The specific value of each gray level; Indicates the grayscale value in the image The number of pixels; P( ) represents grayscale value The relative frequency of occurrence.

[0042] When the contrast of a two-dimensional grayscale image is relatively low, the grayscale values ​​of its pixels are generally concentrated in a certain range; conversely, if the color pixel values ​​are widely distributed in the range of [0, 255], the image will have a clearer visual effect, richer details, and stronger contrast.

[0043] The main purpose of histogram equalization (HE) is to transform the grayscale histogram of the original image into a histogram with a uniform distribution of pixel values. This transformation can soften the grayscale levels of the image, making the image histogram appear flatter and more consistent, thereby enhancing contrast and detail.

[0044] General HE (High-Contrast Enhancement) algorithms are very effective at enhancing target information in optical images. However, due to the unique characteristics of infrared images, a portion of the background occupies a large area and is nearly solid-color. Since these areas lack any useful information, there is no need to enhance the contrast. Therefore, the HE algorithm process needs to remove these unwanted areas. Considering a digital image as a discrete two-dimensional function f(x,y), the gradient of the image is the direction in which the function f(x,y) changes most rapidly. The gradient values ​​are higher at target edges and lower in relatively smooth areas of the image.

[0045] This method uses the median difference method to determine the gradient of the image. The specific steps are as follows: (twenty two); in, and The definition is as follows: (twenty three); (twenty four); in, and Represents the differential values ​​in the x and y directions. It is the value of the image pixels; These are the coordinates of the pixel.

[0046] To address the unique characteristics of infrared images, this embodiment proposes a gradient threshold histogram equalization (GTHE) method for infrared images, such as... Figure 3 As shown. T is a preset threshold and must be adjusted according to the actual infrared image enhancement effect. If the T value is set too low, it may fail to eliminate some irrelevant background areas and is easily affected by isolated noise points and target false edges; if the T value is set too high, it may cause the loss of important details in the image. To enhance the effect of various infrared images, the selection of the T value needs to be tested repeatedly.

[0047] The infrared image preprocessing process based on the Gamma distribution-based GTHE expert domain algorithm is presented here. Transfer learning is used to conduct a comparative classification test on a self-made dataset before and after infrared image preprocessing. The experimental model is a self-developed CNN model. Test results show that the classification accuracy of infrared images after using the proposed method is significantly improved compared to before image processing, with the best improvement reaching 9.39 percentage points, as shown in Table 1.

[0048] Table 1. Classification and Comparison Test Results

[0049] The analysis of the table above leads to the conclusion that image denoising and enhancement techniques greatly improve the performance of the model, and this improvement is even more pronounced when the sample size is small.

[0050] Therefore, this disclosure adopts the above-mentioned image processing method based on infrared images, and the Gamma FoE-GTHE algorithm significantly improves the resolution of infrared images, thereby improving the accuracy of model recognition; compared with traditional image processing techniques, the improvement is 10 percentage points.

[0051] like Figure 4 As shown, embodiments of this disclosure provide an image processing system based on infrared thermal imaging, including: The expert model building module is used to build expert domain FoE models using the expert experience PoE method. The model optimization module is used to optimize and train the expert domain FoE model to obtain an optimized FoE model. The image processing model construction module is used to construct a Gamma FoE denoising algorithm based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, define the denoised image based on the posterior distribution of the image, and construct the image processing model. The denoising solution module is used to construct an energy minimization problem based on the image processing model, combined with the optimized FoE model and gamma distribution, and solve for the optimal denoised image. An enhanced output module is used to enhance the optimal denoised image and output the processed infrared thermal imaging image.

[0052] The implementation process of the functions and roles of each module in the above system is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0053] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0054] In the above embodiments, any number of modules can be combined into one module, or any one module can be split into multiple modules. Alternatively, at least some of the functionality of one or more of these modules can be combined with at least some of the functionality of other modules and implemented in one module. At least one of all modules can be at least partially implemented as hardware circuitry, such as a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System-on-Chip, a System-on-Substrate, a System-on-Package, an Application-Specific Integrated Circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of all modules can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0055] See Figure 5 The electronic device provided in the embodiments of this disclosure includes a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120 and the memory 1130 communicate with each other through the communication bus 1140. Memory 1130 is used to store computer programs; When processor 1110 executes the program stored in memory 1130, it implements an image processing method based on infrared thermal imaging, including the following steps: An expert domain FoE model is constructed using the expert experience PoE method. The expert domain FoE model is optimized and trained to obtain the optimized FoE model; Based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, a Gamma FoE denoising algorithm is constructed. The denoised image is defined based on the posterior distribution of the image, and an image processing model is constructed. Based on the image processing model, and combining the optimized FoE model and gamma distribution, an energy minimization problem is constructed to solve for the optimal denoised image. The optimal denoised image is enhanced to output the processed infrared thermal imaging image.

[0056] The aforementioned communication bus 1140 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0057] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0058] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1130 may also be at least one storage device located remotely from the aforementioned processor 1110.

[0059] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0060] Embodiments of this disclosure also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the image processing method based on infrared thermal imaging as described above.

[0061] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the image processing method based on infrared thermal imaging according to embodiments of this disclosure.

[0062] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0063] The Gamma FoE-GTHE algorithm proposed in this disclosure addresses the problems of infrared thermal imaging being susceptible to interference from emissivity, environmental reflection, atmospheric conditions, thermal noise, and speckle noise, as well as the poor adaptability of traditional optical image processing algorithms. Through integrated processing of expert domain modeling, Gamma distribution statistical fitting, Bayesian denoising, and gradient constraint enhancement, it achieves a comprehensive improvement in infrared image quality. It exhibits significant advantages in denoising performance, detail preservation, edge sharpness, environmental robustness, recognition accuracy, and engineering feasibility. Specific beneficial effects are as follows: I. Effectively suppresses multiple types of noise, significantly improving image signal-to-noise ratio. Infrared images are susceptible to thermal noise, speckle noise, stripe noise, and environmental particle noise. Traditional methods such as median filtering, Gaussian filtering, and Wiener filtering are designed for optical images and have limited ability to suppress infrared heavy-tailed noise, easily resulting in noise residue or blurred details. This disclosure constructs a dedicated denoising mechanism based on an expert domain FoE model and Gamma distribution, matches the statistical characteristics of infrared images with the Student-t distribution, captures local structure through a multi-expert linear filter, and optimizes model parameters using hybrid Monte Carlo sampling and the EM algorithm, enabling accurate separation of noise from the real signal. While preserving the effective information of the image, it can significantly eliminate interference such as speckles, stripes, and pixel fluctuations, making the image smoother and cleaner, and greatly improving the signal-to-noise ratio, thus solving the problems of blurred infrared images and inaccurate temperature measurement at the source.

[0064] II. Precisely preserve edges and details, resolving defects such as blurred edges and overlapping targets. Infrared images of road scenes often suffer from problems such as blurred edges of pedestrians, vehicles, and road conditions, overlapping targets, and loss of detail, directly affecting the reliability of perception. This disclosure uses a FoE model to characterize the global structure and local neighborhood relationships of the image, and utilizes a Markov random potential function to preserve high-frequency information such as edges, corners, and textures, avoiding excessive smoothing during denoising. Subsequently, GTHE gradient threshold histogram equalization is used, and pixel gradients are calculated using the median difference method. Enhancement is applied only to effective regions with high gradients, avoiding backgrounds with no information, thus strengthening target contours and details without producing artifacts or distortion. After processing, the edges of targets such as pedestrians, vehicles, lane lines, and obstacles are clear and distinct, with high separation between targets and background, effectively improving the edge blurring and target adhesion problems caused by traditional algorithms.

[0065] III. Adaptive image contrast enhancement to adapt to complex weather and road conditions. Traditional histogram equalization (GTHE) applied to infrared images often results in over-enhanced backgrounds, localized highlights, and information overload, with performance deteriorating drastically under conditions such as rain, fog, haze, nighttime, and long-distance imaging. This disclosed GTHE algorithm achieves adaptive enhancement through gradient constraints, preserving the image's natural grayscale distribution for more uniform contrast and a more reasonable dynamic range. Simultaneously, Gamma distribution modeling suppresses interference from emissivity, ambient temperature, humidity, dust, and distance, ensuring stable output even in complex environments. Whether in high-temperature daytime backgrounds, low-temperature nighttime scenes, or under rain, fog, or long-distance imaging conditions, it can output clear and stable infrared images, meeting the all-weather requirements of vehicle night vision, roadside monitoring, and autonomous driving.

[0066] IV. Significantly improves target recognition accuracy, especially suitable for small sample learning scenarios. The quality of infrared images directly determines the performance of detection, classification, and recognition models. Traditional processing methods suffer from incomplete denoising and unreasonable enhancement, leading to high false positive and false negative rates. This disclosure utilizes integrated Gamma FoE-GTHE processing to significantly improve image signal-to-noise ratio, edge integrity, and feature recognition, resulting in more prominent target features and more standardized data distribution, thus greatly reducing the learning difficulty of the model. In CNN classification tests, the processed images show improved classification accuracy, with even more significant improvements under small sample training conditions such as 50, 100, and 200 samples, effectively alleviating the industry pain points of high cost and insufficient sample size for infrared data annotation. Improved image quality also reduces model complexity and shortens inference time, facilitating real-time deployment on automotive embedded systems and low-computing-power platforms.

[0067] V. Improve the stability and accuracy of temperature measurement and enhance quantitative analysis capabilities. Infrared thermal imaging possesses both qualitative observation and quantitative temperature measurement capabilities, but noise, artifacts, and grayscale distortion can directly lead to measurement errors. This disclosure fits the infrared speckle noise pattern based on the Gamma distribution and recovers the true image signal through Bayesian maximum a posteriori estimation (MAP), making the mapping relationship between grayscale values ​​and temperature more stable, resulting in smaller measurement errors and stronger consistency. Compared with traditional methods, the processed image provides more accurate and repeatable temperature measurements, meeting the high accuracy requirements of scenarios such as vehicle security, industrial inspection, and fire rescue, thus promoting the upgrade of infrared thermal imaging from "qualitative observation" to "quantitative analysis."

[0068] VI. The algorithm is highly efficient and converges quickly, meeting the requirements of real-time processing. This invention employs the L-BFGS quasi-Newton algorithm to solve the energy minimization problem, which is numerically stable, converges quickly, and has low memory consumption, enabling rapid denoising optimization. The GTHE enhancement stage reduces invalid pixel calculations through gradient filtering, further lowering the computational load. The overall algorithm is highly modular, with a simple flow, easy to implement in code and accelerate with GPUs / FPGAs. It can run in real-time on embedded devices and automotive chips, meeting the stringent requirements of low latency and high real-time performance in applications such as automotive infrared night vision, intelligent transportation, and autonomous driving.

[0069] VII. It has strong versatility and wide applicability, and can be deployed at scale across different scenarios. This disclosure is based on the general mechanism of infrared imaging, and is not dependent on specific equipment or scenarios. It can be adapted to infrared thermal imagers of different resolutions, bands, and brands, and is suitable for many fields such as vehicle night vision, intelligent transportation, security monitoring, industrial inspection, fire rescue, and outdoor inspection. It does not require significant parameter modifications for specific scenarios, has low algorithm portability costs, and possesses strong versatility and scalability, enabling large-scale application of infrared thermal imaging technology across multiple industries.

[0070] 8. Address existing technological shortcomings and improve the infrared image processing technology system. Existing infrared image processing technologies are mostly derived from optical algorithms, which suffer from poor adaptability, difficulty in balancing denoising and enhancement, and weak environmental robustness. This disclosure, starting from the statistical characteristics of infrared images and the requirements of road scenes, constructs a dedicated Gamma FoE-GTHE framework. This framework achieves synergistic denoising and enhancement, separation of details and noise, and a balance between anti-interference and high precision. It fundamentally breaks through the limitations of traditional methods, filling the gap in high-quality infrared processing algorithms specifically for road scenes. This has significant technical value and application implications for improving environmental perception capabilities in fields such as intelligent transportation, autonomous driving, and security monitoring.

[0071] In summary, this invention achieves efficient noise reduction, precise enhancement, edge enhancement, detail preservation, temperature measurement optimization, and real-time operation of infrared images through an integrated algorithm. Its overall performance is significantly better than traditional methods, which can effectively improve the perception reliability and stability of infrared systems under complex road conditions and severe weather. It has outstanding technological advancement, practicality, and promotional value, and can bring significant economic and social benefits.

[0072] It should be noted that, in this document, 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. Furthermore, 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 limitations, 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.

[0073] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. 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 this disclosure. Therefore, this disclosure 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 claimed herein.

Claims

1. An image processing method based on infrared thermal imaging, characterized in that, Includes the following steps: An expert domain FoE model is constructed using the expert experience PoE method. The expert domain FoE model is optimized and trained to obtain the optimized FoE model; Based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, a Gamma FoE denoising algorithm is constructed. The denoised image is defined based on the posterior distribution of the image, and an image processing model is constructed. Based on the image processing model, and combining the optimized FoE model and gamma distribution, an energy minimization problem is constructed to solve for the optimal denoised image. The optimal denoised image is enhanced to output the processed infrared thermal imaging image.

2. The image processing method based on infrared thermal imaging according to claim 1, characterized in that, Expert domain FoE models are constructed using the expert experience PoE method, including: Based on the expert experience PoE framework, the Student-t distribution is used as the expert function to fit the statistical characteristics of infrared images to obtain the image probability density function. Then, according to the Hammersley-Clifford theorem, it is converted into the Gibbs distribution form to calculate the PoE energy function. Based on the PoE energy function, an expert domain FoE model is constructed. A neighborhood system is built using image pixels as graph nodes. The potential function of the Markov random field (MRF) is characterized by an expert function to complete the modeling of the global statistical characteristics of the image, thus obtaining the final expert domain FoE model.

3. The image processing method based on infrared thermal imaging according to claim 1, characterized in that, The expert domain FoE model is optimized and trained to obtain an optimized FoE model, including: The FoE model parameters in the expert domain are optimized by gradient updates, the sampling process is accelerated by hybrid Monte Carlo sampling, and the hidden variables are estimated iteratively by the EM algorithm to optimize the gamma distribution parameters, resulting in the optimized FoE model.

4. The image processing method based on infrared thermal imaging according to claim 1, characterized in that, Based on the prior indices of the optimized FoE model and the likelihood function of the gamma distribution, a Gamma FoE denoising algorithm is constructed. The denoised image is defined based on the posterior distribution of the image, and an image processing model is constructed, including: Using image pixels as observation points, a probability density model of a mixed gamma distribution is constructed; Based on the probability density model, a latent state variable is introduced to label the pixel source distribution, and the joint conditional density function of the complete data is obtained. Based on the joint conditional density function, and using the Bayesian maximum a posteriori probability (MAP) as the criterion, an objective function for infrared image denoising is established, resulting in an image processing model.

5. The image processing method based on infrared thermal imaging according to claim 1, characterized in that, The energy minimization problem integrates the prior energy of the expert domain FoE model with the gamma distribution likelihood energy, and achieves the optimal solution by iteratively updating the gradient.

6. The image processing method based on infrared thermal imaging according to claim 1, characterized in that, Enhancement of the optimal denoised image includes: The gradient value of each pixel in the image is calculated using the median difference method; Set a gradient threshold and retain pixels with gradients greater than the threshold for histogram statistics; Histogram equalization is performed on the selected pixels to complete image enhancement.

7. The image processing method based on infrared thermal imaging according to claim 6, characterized in that, The gradient threshold is adaptively adjusted based on the infrared image scene and noise level to remove background areas with no information while preserving target edge details.

8. An image processing system based on infrared thermal imaging, characterized in that, include: The expert model building module is used to build expert domain FoE models using the expert experience PoE method. The model optimization module is used to optimize and train the expert domain FoE model to obtain an optimized FoE model. The image processing model construction module is used to construct a Gamma FoE denoising algorithm based on the prior index of the optimized FoE model and the likelihood function of the gamma distribution, define the denoised image based on the posterior distribution of the image, and construct the image processing model. The denoising solution module is used to construct an energy minimization problem based on the image processing model, combined with the optimized FoE model and gamma distribution, and solve for the optimal denoised image. An enhanced output module is used to enhance the optimal denoised image and output the processed infrared thermal imaging image.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor is configured to execute a program stored in a memory to implement the image processing method based on infrared thermal imaging as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the image processing method based on infrared thermal imaging as described in any one of claims 1-7.

Images