Petrochemical engineering operation monitoring method based on image enhancement and identification joint optimization

Abstract

The invention discloses a petrochemical engineering operation monitoring method based on image enhancement and recognition joint optimization, and the method comprises the following steps: S1, carrying out the low-light adaptive enhancement, so as to improve the image visibility; s2, respectively extracting brightness and chrominance characteristics through a double-branch network, and realizing cross-branch alignment so as to enhance complementarity of information in a low-light environment; s3, an enhancement and recognition integrated output mode is adopted, and an enhanced image and behavior recognition features are output at the same time; and S4, identifying behaviors by using the space-time modeling network, and outputting a classification result. According to the method, the image definition in a low-illumination environment is remarkably improved, the real-time response capability of a video analysis system is enhanced, and abnormal behaviors and potential safety hazards can be accurately identified in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent monitoring technology, and in particular to a method for monitoring petrochemical operations based on joint optimization of image enhancement and recognition. Background Technology

[0002] Due to its high-risk nature, the petrochemical industry commonly involves high temperatures, high pressures, flammable, explosive, and toxic substances during production, necessitating extremely stringent requirements for safe production and environmental protection. Traditional video surveillance equipment produces blurry images at night or in low-light conditions, making it difficult to meet the 24-hour monitoring needs of critical areas.

[0003] Currently, the petrochemical industry faces numerous challenges in video surveillance and security management, particularly in low-light environments where video image quality is poor, behavior recognition accuracy is insufficient, and system real-time performance is inadequate. To address these issues, existing technologies typically rely on simple image enhancement algorithms or shallow video analysis methods, which offer limited effectiveness and perform poorly in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a petrochemical operation monitoring method based on joint optimization of image enhancement and recognition, to solve the technical problem of how to achieve efficient video monitoring and safety assurance in petrochemical equipment and production environments. This technology can be applied to real-time monitoring of petrochemical facilities under low-light or severe lighting conditions, improving the visibility of equipment operating status, and using behavior recognition technology to identify potential safety hazards or abnormal operations in advance, effectively enhancing production safety and emergency response capabilities.

[0005] This invention is achieved using the following technical solution: a petrochemical operation monitoring method based on joint optimization of image enhancement and recognition, comprising the following steps: S1: Low-light adaptive enhancement to improve image visibility; S2: Brightness and chromaticity features are extracted separately through a dual-branch network and cross-branch alignment is achieved to enhance the complementarity of information in low-light environments; S3: Employs an integrated approach to enhancement and recognition output, simultaneously outputting enhanced images and behavioral recognition features; S4: Use a spatiotemporal modeling network to identify behaviors and output classification results.

[0006] Furthermore, step S1 includes the following steps: S11: Linearize the input image by applying inverse gamma correction to restore the linear photometric response; S12: Image enhancement is performed using a deep learning-based learnable strength compression method.

[0007] Furthermore, the specific calculation method for the image enhancement is as follows: Calculate local contrast, noise estimation, and global average brightness based on the linearized brightness image; Construct an adaptive correction factor for the environment: ; in, , , , These are learnable parameters; Enhanced brightness is achieved by performing intensity compression based on a logarithmic function and incorporating environmental correction. ; Where α is the enhancement coefficient, L(x,y) is the brightness value, L'(x,y) is the enhanced brightness value, C(x,y) is the standard deviation of the brightness in the neighborhood centered at pixel (x,y), N(x,y) is the noise estimate obtained based on local variance or a pre-trained noise estimation network, and L_mean is the global average brightness of the current image.

[0008] Furthermore, step S2 includes the following steps: S21: Extract the brightness and chromaticity features of the image separately using an independent encoder to obtain brightness features and chromaticity features; S22: Align and interact the luminance and chrominance features through a cross-branch attention mechanism; S23: Through adaptive cross-branch fusion, the aligned luminance and chrominance features are synthesized into joint features with spatiotemporal information.

[0009] Furthermore, the specific calculation method for the cross-branch fusion is as follows: A correction matrix M_e is generated based on the environment estimation vector v_e of the current image or video clip. The environment estimation vector v_e includes at least one of the following: global average brightness, smoke or dust index, and proportion of flame or bright light area. The formula for calculating the correction matrix is ​​as follows: M_e = φ(v_e; Θ); Where φ(·) is a nonlinear mapping composed of fully connected layers, and Θ is its parameter; the channel dimension of M_e is consistent with the number of channels of the luminance and chrominance features; Channel correction is applied to the luminance and chrominance features to obtain the corrected luminance feature ~F_L and the corrected chrominance feature ~F_C. The calculation formulas are as follows: ~F_L = γ_L·(M_e·F_L); ~F_C = γ_C·(M_e·F_C); Wherein, γ_L and γ_C are the channel calibration factors for the luminance branch and chrominance branch, respectively; The modified features are fused using dynamic weights to obtain the fused comprehensive feature F_fusion, which is calculated using the following formula: F_fusion = g(w_L ⊙ ~F_L ⊕ w_C ⊙ ~F_C); Where w_L and w_C are the dynamically learned weights obtained through the cross-branch attention mechanism, ⊙ represents element-wise multiplication, ⊕ represents feature concatenation or element-wise addition, and g(·) is a non-linear transformation function.

[0010] Furthermore, step S3 includes the following steps: decoding the feature map fused in step S2 into an enhanced image through a decoder, and extracting temporal features F_seq for behavior recognition from the intermediate layer output of the decoder to provide semantic information for subsequent behavior recognition.

[0011] Furthermore, step S3 also includes the following steps: through multi-task joint optimization, image enhancement and behavior recognition work together to achieve integrated processing of enhancement and recognition, thereby improving the accuracy and stability of low-light behavior recognition.

[0012] Furthermore, step S4 includes the following steps: S41: The key region features of the image are extracted in a fine manner through the self-attention mechanism, and the behavior is spatiotemporally modeled by combining temporal convolution and Transformer network models; S42: Output classification results using a spatiotemporal model.

[0013] Furthermore, the calculation method for the spatiotemporal model is as follows: The region features enhanced by self-attention are subjected to 3D convolution and self-attention processing to obtain the spatiotemporal feature representation F_st; Global average pooling is performed on the spatiotemporal feature representation F_st to obtain the global feature vector h, which is calculated using the following formula: h = GAP(F_st); The global feature vector h is input into the classifier to obtain the behavior category prediction probability distribution p, which is calculated as follows: p = Softmax(W_c·h + b_c); Where GAP(·) represents the global average pooling operation, W_c and b_c are the weights and bias parameters of the classifier, respectively, and p is the predicted probability distribution of the behavior category.

[0014] Furthermore, the multi-task joint optimization employs an environment-adaptive task weight correction factor λ_e to weight and combine the image enhancement loss L_enh and the action recognition loss L_cls, resulting in the following total loss function: L = λ_e·L_enh + (1 - λ_e)·L_cls; Where L_enh is the reconstruction loss between the enhanced image and the reference image, L_cls is the cross-entropy loss for behavior classification, and λ_e ∈ (0,1) are weight coefficients that are adaptively adjusted according to environmental conditions and recognition uncertainty.

[0015] Furthermore, the calculation method for the environment adaptive task weight correction factor λ_e is as follows: The environmental difficulty index D_e is constructed based on the global brightness standard deviation σ_L, the global noise estimate N_global, and the proportion of flame or strong light areas R_flare. Its calculation formula is as follows: D_e = a_1·σ_L + a_2·N_global + a_3·R_flare; The identification uncertainty U_cls is calculated based on the predicted probability distribution p of the behavior category. The formula is as follows: U_cls = - Σ_k p_k·log p_k; The environmental difficulty index D_e is linearly combined with the identification uncertainty U_cls and mapped using the Sigmoid function to obtain λ_e, which is calculated as follows: λ_e = σ(b_0 + b_1·D_e + b_2·U_cls); Where a_1, a_2, a_3, b_0, b_1, and b_2 are learnable parameters, σ(·) is the Sigmoid activation function, and p_k is the predicted probability of the k-th class of behavior.

[0016] Furthermore, the method also includes calculating a risk score correction factor r_t based on behavioral prediction probability and environmental characteristics, used to quantify the operational safety risk at the current time t. The calculation method for the risk score correction factor is as follows: At time t, obtain the behavior category prediction probability vector p_t, where p_t^(abn) represents the prediction probability of abnormal or illegal behavior; Construct an environmental risk feature vector z_t = [D_e, L_mean,t, R_flammable,t], where L_mean,t is the average brightness at the current moment, and R_flammable,t is the proportion of flammable or dangerous areas in the current image; Inputting p_t^(abn) and z_t into the risk assessment network yields the risk score correction factor r_t, which is calculated using the following formula: r_t = ψ(p_t^(abn), z_t; Ω); Wherein, ψ(·) is a nonlinear mapping composed of fully connected layers, Ω is its parameter, and r_t ∈ [0,1] represents the comprehensive safety risk score at the current moment. The method outputs the corresponding risk level or alarm information according to the magnitude of r_t.

[0017] The beneficial effects of this invention are as follows: This invention combines low-light video enhancement with deep learning behavior recognition technology to effectively improve image clarity under low-light conditions, ensure the temporal consistency of video sequences, and significantly improve behavior recognition accuracy in petrochemical scenarios. It solves the problems of poor monitoring effect, insufficient real-time performance, and low recognition accuracy of existing technologies in high-risk environments, and provides a more efficient and reliable intelligent monitoring solution for the safe production of petrochemical enterprises.

[0018] This invention not only significantly improves image clarity in low-light environments but also enhances the real-time response capabilities of video analytics systems, ensuring accurate identification of abnormal behavior and potential security risks in complex environments. Through optimization of the deep learning model, the system can automatically adapt to different lighting, smoke, or obstruction conditions, reducing manual intervention and improving monitoring efficiency and security. Furthermore, this solution achieves efficient data processing and storage through intelligent analysis of monitoring data, reducing network load and hardware resource consumption. Attached Figure Description

[0019] 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 drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the risk scoring process for this invention; Figure 3 This is a flowchart of the image enhancement process of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] See Figures 1 to 3 A petrochemical operation monitoring method based on joint optimization of image enhancement and recognition includes the following steps: S1: Low-light adaptive enhancement to improve image visibility; S2: Brightness and chromaticity features are extracted separately through a dual-branch network and cross-branch alignment is achieved to enhance the complementarity of information in low-light environments; S3: Employs an integrated approach to enhancement and recognition output, simultaneously outputting enhanced images and behavioral recognition features; S4: Use a spatiotemporal modeling network to identify behaviors and output classification results.

[0025] In typical petrochemical plant areas, pipeline corridors, and loading / unloading areas, surveillance videos taken at night or in adverse weather conditions often suffer from insufficient brightness, high noise, and color distortion, leading to frequent false positives and false negatives in subsequent video-based behavior recognition. To address this, this invention, based on the traditional low-light enhancement and behavior recognition process, introduces a series of learnable correction parameters, including an environmental perception enhancement correction factor, a cross-branch alignment correction matrix, task adaptive weights, and risk correction factors. This allows the entire monitoring method to adaptively adjust the enhancement intensity and recognition strategy according to the actual monitoring environment, improving the robustness and accuracy of the overall system.

[0026] In this embodiment, step S1 first performs inverse gamma correction on the input image to restore its linear photometric response. Then, it is enhanced using a deep learning-based learnable intensity compression method, introducing an environment-adaptive correction factor to adapt to different lighting and noise conditions, especially expanding the details of low-light areas (such as shadows) to improve image visibility.

[0027] Linearization is performed using inverse gamma correction for the input low-light RGB image. Its pixel values ​​are normalized to After interval correction, inverse gamma correction is performed to obtain a linearized image, resulting in: ; In the formula, For low-light input images at location Pixel values; The gamma parameter of a camera or display device, typically ranging from... ; These are the pixel values ​​of the linearized image after correction.

[0028] Calculate the luminance component: ; In the formula, Image at location The pixel values ​​of the red, green, and blue channels; coefficients 0.299, 0.587, 0.114: conforming to the weighted ratio of human eye sensitivity to the three colors; Grayscale or luminance component.

[0029] Meanwhile, in low-light environments, the brightness and color distribution of an image may be uneven, which can affect the accuracy of subsequent feature extraction and behavior recognition. By adaptively enhancing the brightness and adjusting the contrast of the image, the problem of uneven brightness caused by low light can be effectively alleviated, thereby improving image quality.

[0030] Specifically, the image enhancement calculation method is as follows: First, based on the linearized brightness image L(x,y), calculate the local contrast C(x,y), noise estimate N(x,y), and global average brightness L_mean of the current image within the neighborhood centered at pixel (x,y); then, construct the environment adaptive correction factor κ_e(x,y), the calculation formula of which is: κ_e(x,y) = κ_0 + κ_1·C(x,y) + κ_2·N(x,y) + κ_3·L_mean; Then, brightness enhancement is performed based on the logarithmic intensity compression model to obtain the enhanced brightness: L'(x,y) = κ_e(x,y)·ln(1 + α·L(x,y)); In the formula, α is the enhancement coefficient, L(x,y) is the brightness value, and L'(x,y) is the enhanced brightness value; C(x,y) is the local contrast, N(x,y) is the local noise estimate, and L_mean is the global average brightness of the current image; κ_0, κ_1, κ_2, and κ_3 are learnable parameters. By using the environment-adaptive correction factor κ_e(x,y), noise and overexposure are suppressed while ensuring the enhancement of details in dark areas, resulting in better robustness of the enhancement results under different petrochemical conditions.

[0031] The above methods enhance details in low-light images, avoiding the problems of missing information in dark areas and overexposure in bright areas, thereby ensuring input quality and providing a high-quality foundation for subsequent feature extraction. Specifically, this invention uses inverse gamma correction and learnable intensity compression to adaptively enhance images in low-light environments, effectively improving details in dark areas and avoiding overexposure in bright areas. This solves the problems of poor input image quality and high noise under low-light conditions, providing a reliable and high-quality data foundation for subsequent feature extraction and behavior recognition.

[0032] In this embodiment, step S2 mainly involves bi-branch feature extraction and cross-branch alignment. Specifically, the luminance and chrominance features of the image are extracted separately by an independent encoder to obtain luminance feature F_L and chrominance feature F_C; the luminance feature F_L and chrominance feature F_C are aligned and interacted through a cross-branch attention mechanism to highlight behavior-related regions; and the luminance and chrominance information are synthesized into a joint feature with spatiotemporal information through adaptive cross-branch fusion.

[0033] The specific calculation method for cross-branch fusion is as follows: The correction matrix M_e is generated based on the environment estimation vector v_e of the current image or video clip (including at least one feature among global average brightness, smoke or dust index, and proportion of flame or bright light area), and the calculation formula is as follows: M_e = φ(v_e; Θ); Where φ(·) is a nonlinear mapping composed of fully connected layers, Θ is its parameter, and the channel dimension of M_e is consistent with the number of channels of the luminance feature F_L and the chrominance feature F_C; then, channel correction is performed on the luminance and chrominance features to obtain the corrected luminance feature ~F_L and the corrected chrominance feature ~F_C, and their calculation formulas are as follows: ~F_L = γ_L·(M_e·F_L), ~F_C = γ_C·(M_e·F_C), Where γ_L and γ_C are the channel calibration factors for the luminance and chrominance branches, respectively; then, the corrected features are fused using dynamic weights to obtain the fused comprehensive feature F_fusion, the calculation formula of which is: F_fusion = g(w_L ⊙ ~F_L ⊕ w_C ⊙ ~F_C), In the formula, F_fusion represents the fused integrated features, w_L and w_C are the dynamically learned weights obtained through the cross-branch attention mechanism, ⊙ represents element-wise multiplication, ⊕ represents feature concatenation or element-wise addition, and g(·) is the nonlinear transformation function. By introducing the correction matrix M_e and the channel calibration factors γ_L and γ_C, joint correction and adaptive fusion of the luminance and chrominance channels are achieved.

[0034] In this embodiment, step S3 specifically combines image enhancement and behavior recognition. The feature map fused in step S2 is decoded into an enhanced image using a decoder, while temporal features are extracted to provide semantic information for subsequent behavior recognition. Through multi-task joint optimization, image enhancement and behavior recognition work collaboratively, achieving integrated processing of enhancement and recognition, improving the accuracy and stability of low-light behavior recognition. The multi-task joint optimization process introduces an environment-adaptive task weight correction factor λ_e to weight and combine the image enhancement loss L_enh and the behavior recognition loss L_cls, with the total loss function being: L = λ_e·L_enh + (1 - λ_e)·L_cls; Where L_enh is the reconstruction loss between the enhanced image and the reference image, L_cls is the cross-entropy loss for behavior classification, and λ_e ∈ (0,1) are weight coefficients adaptively adjusted according to environmental conditions and recognition uncertainty. The environmental difficulty index D_e is obtained by linearly combining the global brightness standard deviation σ_L, the global noise estimate N_global, and the proportion of flame or strong light areas R_flare. D_e = a_1·σ_L + a_2·N_global + a_3·R_flare; The identification uncertainty U_cls is defined by the entropy of the predicted probability distribution p of the behavior category: U_cls = - Σ_k p_k·ln p_k; By linearly combining the environmental difficulty D_e with the recognition uncertainty U_cls and mapping it using the Sigmoid function, we obtain λ_e: λ_e = σ(b_0 + b_1·D_e + b_2·U_cls), Where a_1, a_2, a_3, b_0, b_1, and b_2 are learnable parameters, σ(·) is the Sigmoid activation function, and p_k is the predicted probability of the k-th behavior. By dynamically adjusting λ_e, the training weights of the image enhancement sub-network can be enhanced in extremely low light or strong interference conditions, while focusing more on recognition accuracy in better environments.

[0035] Through multi-task joint optimization, the enhancement and recognition modules work collaboratively to achieve integrated "enhancement equals recognition" processing, improving the accuracy and stability of low-light behavior recognition. Specifically, this invention employs a dual-branch network combining luminance and chrominance branches, utilizing a cross-branch attention mechanism to dynamically fuse luminance and chrominance information, enhancing feature representation capabilities under low-light conditions. Simultaneously, the system is designed as an integrated enhancement and recognition joint optimization framework, enabling image enhancement and behavior recognition to mutually promote each other, avoiding the task fragmentation problem in traditional methods, and significantly improving overall recognition performance.

[0036] In this embodiment, step S4 specifically involves using a spatiotemporal modeling network to recognize behavior. Key region features of the image are meticulously extracted using a self-attention mechanism, and the behavior is modeled using a combination of temporal convolution and Transformer network models. The spatiotemporal model effectively captures the temporal and spatial information of actions, especially for recognizing dynamic behavior in low-light environments. During spatiotemporal modeling, local self-attention mechanisms further emphasize the features of key regions, while global modeling ensures the long-term dependencies of dynamic behaviors. Through this method, the system can more accurately classify and predict behavior under low-light conditions.

[0037] Based on the extracted behavioral features, spatiotemporal modeling is performed to obtain the spatiotemporal feature representation F_st, and the classification result is output. Specifically, the calculation method of the spatiotemporal model is as follows: First, global average pooling is performed on the spatiotemporal feature representation F_st to obtain the global feature vector h, and its calculation formula is as follows: h = GAP(F_st); The global feature vector h is then input into the classifier to obtain the behavior category prediction probability distribution p, which is calculated using the following formula: p = Softmax(W_c·h + b_c); In the formula, GAP(·) is the global average pooling operation, W_c and b_c are the classifier weights and biases, respectively, and p is the predicted probability distribution of the behavior category.

[0038] Furthermore, to meet the safety early warning needs of petrochemical operations, this invention introduces a risk scoring correction factor r_t based on behavior classification to quantify the safety risks of different behaviors under different environments. At time t, the behavior category prediction probability vector is denoted as p_t, where p_t^(abn) represents the predicted probability of abnormal or illegal behavior; an environmental risk feature vector z_t = [D_e, L_mean,t, R_flammable,t] is constructed, where L_mean,t is the average brightness at the current time, and R_flammable,t is the proportion of flammable or dangerous areas in the current image; p_t^(abn) and z_t are input into the risk assessment network to obtain the risk scoring correction factor r_t, whose calculation formula is: r_t = ψ(p_t^(abn), z_t; Ω), Where ψ(·) is a nonlinear mapping composed of fully connected layers, Ω is its parameter, and r_t ∈ [0,1] represents the comprehensive safety risk score at the current moment. The system can output the corresponding risk level or alarm information according to the magnitude of r_t.

[0039] By combining 3D convolution with Transformer architecture, this invention captures both short-term action features and models long-term dependencies, achieving accurate recognition of complex behaviors. Specifically, in the behavior recognition stage, this invention combines 3D convolution with Transformer networks to achieve unified modeling of local temporal features and global long-term dependencies, enabling the simultaneous recognition of short-term actions and complex long-term behaviors. This method is particularly suitable for low-light scenarios such as petrochemical plants and nighttime security, enabling accurate classification of personnel behavior and detection of abnormal behaviors, and has high practical value and promotional significance.

[0040] Based on the above embodiments, the present invention has advantages in at least the following aspects: Regarding image enhancement, traditional low-light enhancement methods often rely on histogram equalization or fixed contrast stretching, which can easily lead to amplified noise in dark areas and overexposure in bright areas. Furthermore, the enhancement results lack adaptability to subsequent recognition tasks. This invention, however, employs an adaptive enhancement method combining inverse gamma correction with learnable logarithmic intensity compression. This not only restores details in dark areas and suppresses overexposure in bright areas but also, through a joint optimization mechanism, makes the enhancement results more suitable for the recognition task, fundamentally improving recognition accuracy under low-light conditions.

[0041] Regarding feature extraction and fusion, most methods rely solely on a single luminance or chrominance channel for feature extraction, which can easily lead to information loss or color distortion under low-light conditions, resulting in insufficient feature representation. This invention, however, designs a dual-channel feature extraction network with luminance and chrominance branches and introduces a cross-branch attention mechanism for dynamic weighted fusion, fully leveraging the complementary effects of luminance and chrominance information to significantly enhance feature integrity and robustness.

[0042] In terms of behavior recognition modeling, most existing behavior recognition methods rely on two-dimensional convolution or single-frame image features, making it difficult to capture temporal information, especially under low-light conditions, resulting in low accuracy for recognizing complex actions or long-term behaviors. This invention, however, combines three-dimensional convolution with a Transformer model, enabling the extraction of both local temporal features and the modeling of long-term global dependencies. This allows for accurate recognition of both short-term rapid actions and long-term complex behaviors, significantly improving the accuracy and applicability of dynamic behavior recognition in low-light environments.

[0043] For the foregoing embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to this application.

[0044] The above embodiments describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the invention should be within the protection scope of the appended claims.

Claims

1. A method for monitoring petrochemical operations based on joint optimization of image enhancement and recognition, characterized in that, Includes the following steps: S1: Low-light adaptive enhancement to improve image visibility; S2: Brightness and chromaticity features are extracted separately through a dual-branch network and cross-branch alignment is achieved to enhance the complementarity of information in low-light environments; S3: Employs an integrated approach to enhancement and recognition output, simultaneously outputting enhanced images and behavioral recognition features; S4: Use a spatiotemporal modeling network to identify behaviors and output classification results.

2. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 1, characterized in that, Step S1 includes the following steps: S11: Linearize the input image by applying inverse gamma correction to restore the linear photometric response; S12: Image enhancement is performed using a deep learning-based learnable strength compression method.

3. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 2, characterized in that, The specific calculation method for the image enhancement is as follows: Calculate local contrast, noise estimation, and global average brightness based on the linearized brightness image; Construct an adaptive correction factor for the environment: ； in, , , , These are learnable parameters; Enhanced brightness is achieved by performing intensity compression based on a logarithmic function and incorporating environmental correction. ； Where α is the enhancement coefficient, L(x,y) is the brightness value, L'(x,y) is the enhanced brightness value, C(x,y) is the standard deviation of the brightness in the neighborhood centered at pixel (x,y), N(x,y) is the noise estimate obtained based on local variance or a pre-trained noise estimation network, and L_mean is the global average brightness of the current image.

4. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 1, characterized in that, Step S2 includes the following steps: S21: Extract the brightness and chromaticity features of the image separately using an independent encoder to obtain brightness features and chromaticity features; S22: Align and interact the luminance and chrominance features through a cross-branch attention mechanism; S23: Through adaptive cross-branch fusion, the aligned luminance and chrominance features are synthesized into joint features with spatiotemporal information.

5. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 4, characterized in that, The specific calculation method for cross-branch fusion is as follows: A correction matrix M_e is generated based on the environment estimation vector v_e of the current image or video clip. The environment estimation vector v_e includes at least one of the following: global average brightness, smoke or dust index, and proportion of flame or bright light area. The formula for calculating the correction matrix is ​​as follows: M_e = φ(v_e; Θ); Where φ(·) is a nonlinear mapping composed of fully connected layers, and Θ is its parameter; the channel dimension of M_e is consistent with the number of channels of the luminance and chrominance features; Channel correction is applied to the luminance and chrominance features to obtain the corrected luminance feature ~F_L and the corrected chrominance feature ~F_C. The calculation formulas are as follows: ~F_L = γ_L·(M_e·F_L); ~F_C = γ_C·(M_e·F_C); Wherein, γ_L and γ_C are the channel calibration factors for the luminance branch and chrominance branch, respectively; The modified features are fused using dynamic weights to obtain the fused comprehensive feature F_fusion, which is calculated using the following formula: F_fusion = g(w_L ⊙ ~F_L ⊕ w_C ⊙ ~F_C); Where w_L and w_C are the dynamically learned weights obtained through the cross-branch attention mechanism, ⊙ represents element-wise multiplication, ⊕ represents feature concatenation or element-wise addition, and g(·) is a non-linear transformation function.

6. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 1, characterized in that, Step S3 includes the following steps: decoding the feature map fused in step S2 into an enhanced image through a decoder, and extracting temporal features F_seq for behavior recognition from the intermediate layer output of the decoder to provide semantic information for subsequent behavior recognition.

7. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 6, characterized in that, Step S3 further includes the following steps: Through multi-task joint optimization, image enhancement and behavior recognition work together to achieve integrated processing of enhancement and recognition, improving the accuracy and stability of low-light behavior recognition; the multi-task joint optimization uses an environment-adaptive task weight correction factor λ_e to weight and combine the image enhancement loss L_enh and the behavior recognition loss L_cls, with the total loss function being: L = λ_e·L_enh + (1 - λ_e)·L_cls; Where L_enh is the reconstruction loss between the enhanced image and the reference image, L_cls is the cross-entropy loss for behavior classification, and λ_e ∈ (0,1) are weight coefficients that are adaptively adjusted according to environmental conditions and recognition uncertainty.

8. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 1, characterized in that, Step S4 includes the following steps: S41: Key region features of the image are finely extracted using a self-attention mechanism, and the behavior is spatiotemporally modeled by combining temporal convolution and Transformer network models; the calculation method of the spatiotemporal model is as follows: The region features enhanced by self-attention are subjected to 3D convolution and self-attention processing to obtain the spatiotemporal feature representation F_st; Global average pooling is performed on the spatiotemporal feature representation F_st to obtain the global feature vector h, which is calculated using the following formula: h = GAP(F_st); The global feature vector h is input into the classifier to obtain the behavior category prediction probability distribution p, which is calculated as follows: p = Softmax(W_c·h + b_c); Where GAP(·) represents the global average pooling operation, W_c and b_c are the weights and bias parameters of the classifier, respectively, and p is the probability distribution of the predicted behavior category; S42: Output classification results using a spatiotemporal model.

9. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in claim 7, characterized in that, The calculation method for the environment adaptive task weight correction factor λ_e is as follows: The environmental difficulty index D_e is constructed based on the global brightness standard deviation σ_L, the global noise estimate N_global, and the proportion of flame or strong light areas R_flare. Its calculation formula is as follows: D_e = a_1·σ_L + a_2·N_global + a_3·R_flare; The identification uncertainty U_cls is calculated based on the predicted probability distribution p of the behavior category. The formula is as follows: U_cls = - Σ_k p_k·log p_k; The environmental difficulty index D_e is linearly combined with the identification uncertainty U_cls and mapped using the Sigmoid function to obtain λ_e, which is calculated as follows: λ_e = σ(b_0 + b_1·D_e + b_2·U_cls); Where a_1, a_2, a_3, b_0, b_1, and b_2 are learnable parameters, σ(·) is the Sigmoid activation function, and p_k is the predicted probability of the k-th class of behavior.

10. The petrochemical operation monitoring method based on joint optimization of image enhancement and recognition as described in any one of claims 8 and 9, characterized in that, The method further includes calculating a risk score correction factor r_t based on behavioral prediction probability and environmental characteristics, used to quantify the operational safety risk at the current time t. The calculation method for the risk score correction factor is as follows: At time t, obtain the behavior category prediction probability vector p_t, where p_t^(abn) represents the prediction probability of abnormal or illegal behavior; Construct an environmental risk feature vector z_t = [D_e, L_mean,t, R_flammable,t], where L_mean,t is the average brightness at the current moment, and R_flammable,t is the proportion of flammable or dangerous areas in the current image; Inputting p_t^(abn) and z_t into the risk assessment network yields the risk score correction factor r_t, which is calculated using the following formula: r_t = ψ(p_t^(abn), z_t; Ω); Wherein, ψ(·) is a nonlinear mapping composed of fully connected layers, Ω is its parameter, and r_t ∈ [0,1] represents the comprehensive safety risk score at the current moment. The method outputs the corresponding risk level or alarm information according to the magnitude of r_t.

Images