Camera module dynamic environment adaptation and image enhancement method

Abstract

The invention discloses a dynamic environment adaptation and image enhancement method for a camera module. The method comprises the following steps: S1, acquiring environment sensing data; s2, inputting a Bayesian module, constructing a Gaussian process model, and generating image enhancement parameter distribution; s3, initializing an image enhancement module according to the parameter distribution; s4, iteratively optimizing the image enhancement parameters by using zebra optimization, performing enhancement processing on the original image, and generating candidate results; s5, performing image quality evaluation on the candidate result, and performing weighted fusion to generate a comprehensive score; s6, feeding back the comprehensive score to a Bayesian module, updating a Gaussian model, and improving the prediction accuracy; and S7, selecting the image with the highest score as the final output, and storing the environment data and the parameter samples into a database. According to the invention, real-time improvement and stable output of the image quality of the camera module in a complex environment are realized through dynamic adaptive parameter adjustment and image enhancement technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and camera modules, and in particular to a camera module dynamic environment adaptation and image enhancement method. Background Art

[0002] With the rapid development of modern electronic devices and intelligent systems, camera modules, as an important component of image acquisition and processing, have a direct impact on the overall performance of downstream image processing, computer vision, and intelligent decision-making systems. At present, camera modules are widely used in smart phones, surveillance cameras, autonomous driving, industrial inspection and other fields. However, as the application environment becomes increasingly complex and changeable, traditional camera modules are often unable to maintain stable and clear imaging effects in dynamic environments (such as rapidly changing lighting conditions, strong contrast environments, low-light scenes, and severe weather conditions). This is mainly manifested in problems such as increased image noise, color deviation, insufficient contrast, and loss of details during imaging, which seriously affects the accuracy and real-time performance of subsequent image analysis, target detection, and recognition tasks.

[0003] Existing technologies address image quality degradation primarily through fixed parameter correction, preset enhancement algorithms, or simple adaptive adjustment strategies. These methods typically rely on traditional image enhancement algorithms, such as Retinex theory, multi-scale Retinex, and histogram equalization. While these methods improve image quality to a certain extent, they are inadequately responsive to environmental changes and lack the ability to deeply explore complex parameter relationships in dynamic environments. Furthermore, while some technologies based on traditional optimization algorithms, such as genetic algorithms and particle swarm optimization, can adjust camera parameters to a certain extent, they often suffer from slow convergence, susceptibility to local optimality, and high computational resource usage, making them difficult to meet the requirements of real-time dynamic environment adaptation.

[0004] In a dynamic environment, the camera module not only needs to perceive environmental factors such as lighting, color temperature, and motion blur in real time, but also needs to quickly adjust the parameters of image acquisition and enhancement to obtain the best imaging effect. Existing technologies mostly use fixed parameter adjustment schemes, which cannot provide sufficient adaptability when the environment changes drastically. This leads to the problem of image quality degradation in complex scenes such as tunnels, at night, backlight, rain and fog, affecting user experience and system performance. In the existing technology, some methods attempt to combine image enhancement with parameter adaptive control, but most of them are simple series relationships, lacking deep fusion and feedback mechanisms, and cannot achieve coordinated optimization between image enhancement and camera parameter adjustment.

[0005] On the other hand, traditional image quality assessment methods rely primarily on single metrics, such as the Structural Similarity Index (SSIM) or the No-Reference Image Quality Evaluation (NIQE). While these methods can reflect the overall quality of an image to a certain extent, they are insufficient for assessing performance in special environments, such as underwater image color. Furthermore, existing technologies lack comprehensive evaluation methods that combine weighted multiple metrics. This makes it difficult to fully reflect all aspects of an image during image quality improvement, thus affecting the effectiveness and robustness of image enhancement algorithms.

[0006] To address these issues, some studies have begun exploring online adjustments to camera module parameters based on meta-heuristic optimization algorithms to improve image enhancement. These methods utilize techniques such as the zebra optimization algorithm and Bayesian optimization to strike a balance between global and local search, continuously optimizing image enhancement parameters through multiple rounds of iteration. However, existing technologies still have many shortcomings in algorithm design: First, most methods operate each optimization module independently and lack a deep fusion mechanism, resulting in limited synergy between image enhancement and environmental adaptation. Second, parameter initialization, search boundary setting, and update strategies are mostly preset fixed values, which cannot fully reflect the real-time changes in image quality and parameter distribution in dynamic environments. Third, the image quality evaluation index is single, making it difficult to fully capture the local details and global quality information of the image, which in turn affects the feedback and adjustment of the optimization process.

[0007] Therefore, how to provide a camera module dynamic environment adaptation and image enhancement method is a problem that technicians in this field urgently need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a method for dynamic environment adaptation and image enhancement of a camera module. The present invention makes full use of the deep fusion technology of Bayesian optimization and improved zebra optimization algorithm, combines the prediction of image quality gradient by deep neural network and multi-index weighted fusion image quality evaluation strategy, and describes in detail the implementation process of adaptive adjustment of acquisition parameters and intelligent image enhancement of the camera module in a complex dynamic environment. By constructing an image enhancement parameter distribution based on a Gaussian process model and using an improved zebra optimization algorithm to perform multiple rounds of iterative optimization on candidate parameters, the present invention can respond to environmental changes in real time and realize closed-loop feedback between image enhancement and parameter adjustment. This method not only has real-time and high precision, but also significantly improves image quality, ensuring that the camera module can still output clear, natural, and detail-rich images under harsh conditions such as low illumination, backlight, rain and fog, thereby improving the overall robustness of the system and the reliability of the application.

[0009] A camera module dynamic environment adaptation and image enhancement method according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect environmental perception data of the camera module in the target environment, including light intensity, color temperature information, blur level and scene dynamic characteristics;

[0011] S2. Input the environmental perception data into the Bayesian optimization algorithm module, build a Gaussian process model, and generate image enhancement parameter distribution results based on the mapping relationship between historical image quality scores and image enhancement parameters;

[0012] S3. Initialize and configure the parameter search space of the image enhancement module according to the image enhancement parameter distribution results, including the initial population position, search boundary range, and update behavior control factor of the zebra optimization algorithm;

[0013] S4. In the image enhancement module, the image enhancement parameters are optimized for multiple rounds of iterations using the zebra optimization algorithm, the original image data is enhanced, and multiple image enhancement candidate results are generated;

[0014] S5. Evaluate the image quality of multiple image enhancement candidate results by extracting information of image structure similarity index, no-reference image quality index, and underwater image color quality index, and generate a comprehensive image quality score based on weighted fusion;

[0015] S6. Feedback the comprehensive image quality score to the Bayesian optimization algorithm module to update the Gaussian process model and improve the prediction accuracy and generalization ability of the next round of image enhancement parameter distribution results;

[0016] S7. The image with the highest score is used as the final image enhancement output result, and the current environment data and the adopted image enhancement parameters constitute a sample pair and are stored in the database.

[0017] Optionally, the S2 specifically includes:

[0018] S21, record the environmental perception data as E={E1, E2, ..., E n}, where E i represents the i-th environmental feature;

[0019] S22. Record the historical image quality score set as Q and the corresponding image enhancement parameter set as P to form a training data set D:

[0020] D={(E i ,Q i ,P i )|i=1,2,…,m};

[0021] Among them, Q i Score the quality of the i-th image, P i is the corresponding image enhancement parameter, m represents the total number of samples in the training data;

[0022] S23. Construct Gaussian process model GP:

[0023] f(E)~GP(m(E),k(E,E ′ ));

[0024] Among them, m(E) represents the mean function, k(E,E ′ ) represents the covariance function, E and E ′ are any two environmental data respectively, f(E) represents the mapping function for predicting image enhancement parameters for the input environmental perception data E;

[0025] S24, use the training data set D to fit the Gaussian process model GP to obtain the prediction function f * (E):

[0026]

[0027] Among them, K represents the training environment dataset E train The covariance matrix on , represents the variance of the observation noise, I is the unit matrix, P train Representation and training environment dataset E train The corresponding image enhancement parameter value, m(E train ) represents the mean function in the training environment dataset E train The value vector on ;

[0028] S25. Record the current environment data as E current , using the prediction function f * (E) Predict the current environment data and obtain the preliminary image enhancement parameter P init , calculate the prediction uncertainty σ corresponding to the current environmental data current , according to the preset dynamic weight function α(σ current ) Fusion of preliminary image enhancement parameters and standard image enhancement parameters P baseline Get the image enhancement parameter distribution results:

[0029] P current =α(σ current )·P init +[1-α(σ current )]·P baseline ;

[0030] Among them, P current is the image enhancement parameter distribution result.

[0031] Optionally, the S3 specifically includes:

[0032] S31, receiving the output image enhancement parameter distribution result Pcurrent , as the basic reference information before the zebra optimization algorithm searches, and as the core basis for constructing the parameter search space of the image enhancement module;

[0033] S32. Constructing initial value ranges for the corresponding parameters based on the image enhancement parameter distribution results, and generating a joint search space covering multiple dimensions including brightness enhancement factor, contrast adjustment coefficient, Retinex component weight, and image detail restoration strength.

[0034] S33. In the constructed joint search space, according to the set population size N, the initial population individual positions of the zebra optimization algorithm are generated using a uniform distribution strategy:

[0035] Z init ={z j ∣z j =P current +∈ j ,∈ j ~U(-Δ,Δ),j=1,2,…,N};

[0036] Among them, Z init Represents the initial population position set of the zebra optimization algorithm, z j Indicates the position of the jth individual in the initial population, ∈ j represents the random variable used to perturb the distribution of image enhancement parameters, Δ represents the search perturbation amplitude, and N represents the number of individuals in the initial population;

[0037] S34, setting the search boundary range of each enhancement parameter, including the maximum adjustable value and the minimum adjustable value, and dynamically adjusting the boundary interval width in combination with the confidence interval of the image enhancement parameter distribution result;

[0038] S35. Configure the update behavior control factors of the zebra optimization algorithm, including jump direction factor, behavior response probability, population update frequency, and local disturbance intensity;

[0039] S36. Input the initialized population position, search boundary range and update behavior control factor into the image enhancement module to complete the parameter search space initialization configuration.

[0040] Optionally, the S4 specifically includes:

[0041] S41, receiving the parameter search space S of the generated image enhancement module search , which includes the initial population position set Z init , search boundary range and initial update behavior control factor and original image data I raw ;

[0042] S42, setting the number of iterative optimization rounds T;

[0043] S43. For the tth iteration, select a candidate solution from the current parameter search space Each candidate solution represents a set of image enhancement parameters;

[0044] S44, using the image quality gradient prediction module based on deep neural network, for each candidate solution Generated candidate enhanced images Calculate local detail gradients And generate the corresponding search step and jump probability prediction value

[0045] S45. Set of candidate solutions Calculate the statistic, the mean is μ t , the standard deviation is σ t and skewness is κ t , and at the same time, calculate the global image quality improvement factor Γ t :

[0046]

[0047] in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the average image quality score of the candidate solution set, Represents the global image quality score, adopts a multi-dimensional adaptive weight adjustment strategy, and dynamically calculates the update factor β t :

[0048]

[0049] Among them, β0 is the preset basic update factor, λ is the global quality improvement weight coefficient, σ t represents the standard deviation of the candidate solution set in the tth iteration, μ t represents the mean of the candidate solution set in the tth iteration, κ t represents the skewness of the candidate solution set in the tth iteration;

[0050] S46. For each candidate solution Generated candidate enhanced images Use the image enhancement function G() to process and calculate the image quality score Q (j,t) :

[0051]

[0052] Among them, α, β, and γ are preset weight coefficients. Represents the candidate enhanced image and the reference image I refThe structural similarity index between represents the no-reference image quality evaluation index of the candidate enhanced image, Represents the color quality evaluation index of the candidate enhanced image, I ref is the reference image, N0 and U0 are normalization constants;

[0053] S47. Construct a mixed reward function R (j,t) The reward function comprehensively considers the image quality improvement rate, local detail restoration effect, and population convergence speed to evaluate each candidate solution:

[0054] R (j,t) =w1ΔQ (j,t) +w2 D (j,t) +w3 C (t) ;

[0055] Among them, D (j,t) represents the local detail recovery index corresponding to the candidate solution, C (t) Indicates the population convergence speed index, w1, w2, w3 are preset weight coefficients, ΔQ (j,t) Represents a candidate solution The image quality improvement of the generated candidate enhanced image in the tth iteration;

[0056] S48, using the update mechanism based on adaptive fuzzy control to update the mixed reward function R (j,t) and global image quality improvement factor Γ t Perform fuzzy membership evaluation and use the preset fuzzy membership function to update the control factor β t , generate the updated control factor β′ t :

[0057] β′ t =β t +K·(μ high (R (t) ,Γ t )-μ low (R (t) ,Γ t ));

[0058] Among them, K is the adjustment coefficient, μ high (R (t) ,Γ t ) and μ low (R (t) ,Γ t ) represent the mixed reward function R (t) and Γ t The calculated high-efficiency and low-efficiency membership values;

[0059] S49, based on the updated control factor βt and the jump probability prediction value Update candidate solutions:

[0060]

[0061] in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the new value of the jth candidate solution in the t+1th iteration after the update;

[0062] S410: After completing T iterations, select the candidate enhanced image with the highest image quality score as the final output result, and output the corresponding multiple image enhancement candidate results.

[0063] Optionally, the S5 specifically includes:

[0064] S51, receiving and collecting multiple image enhancement candidate results, including image enhancement effects obtained by different parameter combinations;

[0065] S52, extracting an index that can reflect the image structure similarity for each image enhancement candidate result by comparing the structural feature information between the candidate image and the reference image;

[0066] S53. Extracting a no-reference image quality index for each image enhancement candidate result. The no-reference image quality index is based on the statistical characteristics and local information of the image itself and can represent the overall visual quality of the image without relying on an external reference.

[0067] S54. extracting an underwater image color quality index for each image enhancement candidate result, where the underwater image color quality index specifically captures the effectiveness of color distribution, saturation, and contrast of the image underwater, reflecting the image's performance in color reproduction;

[0068] S55. According to a pre-set weighting strategy, the weighting strategy sets fixed weights for the structural similarity index, the no-reference image quality index, and the underwater image color quality index, respectively, with the sum of the weight values being 1, reflecting the relative importance of the three indicators in the comprehensive visual quality evaluation, and weightedly fusing the information of the three indicators to obtain a comprehensive evaluation value;

[0069] S56. Generate a comprehensive image quality score for each image enhancement candidate result based on the weighted fusion of various indicator information.

[0070] The beneficial effects of the present invention are:

[0071] This method achieves adaptive parameter adjustment and image enhancement in dynamic and complex environments by deeply fusing a Bayesian optimization algorithm with an improved Zebra optimization algorithm, combined with image quality gradient prediction based on a deep neural network and a multi-index weighted fusion image quality assessment strategy. This method enables the camera module to respond in real time to environmental changes such as lighting, color temperature, and motion blur, effectively avoiding problems such as increased image noise, color deviation, and loss of detail caused by drastic environmental changes, thereby significantly improving the overall quality of image acquisition.

[0072] By constructing a Gaussian process model to obtain the distribution of image enhancement parameters and then using an improved zebra optimization algorithm to perform multiple rounds of iterative optimization of candidate parameters, this invention implements a closed-loop feedback mechanism between image enhancement and parameter adjustment, enabling the system to maintain high image clarity and color reproduction in complex scenes. Furthermore, an image quality gradient prediction module based on a deep neural network and a multi-dimensional adaptive weight adjustment strategy enable the system to dynamically adjust the update step size and jump probability during the search process, further improving the algorithm's global search capabilities and local fine-grained optimization level.

[0073] Overall, this invention not only improves the camera module's imaging performance in harsh environments but also enhances the system's real-time and robustness. This method enables high-quality image output in complex scenarios such as low illumination, backlighting, and rain and fog. This provides more reliable and efficient visual information support for fields such as autonomous driving, security monitoring, and industrial inspection, greatly expanding the applicability and performance advantages of smart camera modules in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0075] Figure 1 This is a flow chart of a camera module dynamic environment adaptation and image enhancement method proposed by the present invention;

[0076] Figure 2 This is a schematic diagram of the overall system architecture of a camera module dynamic environment adaptation and image enhancement method proposed in the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0078] refer to Figure 1 and Figure 2A camera module dynamic environment adaptation and image enhancement method comprises the following steps:

[0079] S1. Collect environmental perception data of the camera module in the target environment, including light intensity, color temperature information, blur level and scene dynamic characteristics;

[0080] S2. Input the environmental perception data into the Bayesian optimization algorithm module, build a Gaussian process model, and generate image enhancement parameter distribution results based on the mapping relationship between historical image quality scores and image enhancement parameters;

[0081] S3. Initialize and configure the parameter search space of the image enhancement module according to the image enhancement parameter distribution results, including the initial population position, search boundary range, and update behavior control factor of the zebra optimization algorithm;

[0082] S4. In the image enhancement module, the image enhancement parameters are optimized for multiple rounds of iterations using the zebra optimization algorithm, the original image data is enhanced, and multiple image enhancement candidate results are generated;

[0083] S5. Evaluate the image quality of multiple image enhancement candidate results by extracting information of image structure similarity index, no-reference image quality index, and underwater image color quality index, and generate a comprehensive image quality score based on weighted fusion;

[0084] S6. Feedback the comprehensive image quality score to the Bayesian optimization algorithm module to update the Gaussian process model and improve the prediction accuracy and generalization ability of the next round of image enhancement parameter distribution results;

[0085] S7. The image with the highest score is used as the final image enhancement output result, and the current environment data and the adopted image enhancement parameters constitute a sample pair and are stored in the database.

[0086] The present invention realizes online adaptive adjustment and image enhancement of data such as illumination, color temperature, blur level and dynamic features collected by the camera module in the target environment by integrating the Bayesian optimization algorithm and the improved zebra optimization algorithm. The system first uses the Bayesian module to construct a Gaussian process model, generates a parameter distribution based on the mapping relationship between historical image quality scores and image enhancement parameters, and then initializes the parameter search space of the image enhancement module based on this. Subsequently, through multiple rounds of iterative optimization of the zebra optimization algorithm, the original image data is enhanced, and multiple image enhancement candidate results are generated. The candidate results are comprehensively evaluated for image quality using a weighted fusion method of structural similarity, no-reference image quality index and underwater image color quality index. The feedback mechanism uses the comprehensive score to update the Gaussian process model in the Bayesian optimization module to form a closed-loop optimization process. Finally, the present invention automatically selects the image with the highest score as output, and stores the environmental data and optimization parameters in a database to provide data support for subsequent system optimization. This method significantly improves the imaging quality of the camera in complex dynamic environments such as low illumination, backlighting, rain and fog, ensuring rich image details, reduced noise, and accurate color reproduction. It also has real-time response and high robustness, meeting the actual needs of high-demand application fields such as autonomous driving and security monitoring.

[0087] In this embodiment, S2 specifically includes:

[0088] S21, record the environmental perception data as E={E1, E2, ..., E n}, where E i represents the i-th environmental feature;

[0089] S22. Record the historical image quality score set as Q and the corresponding image enhancement parameter set as P to form a training data set D:

[0090] D={(E i ,Q i ,P i )|i=1,2,…,m};

[0091] Among them, Q i Score the quality of the i-th image, P i is the corresponding image enhancement parameter, m represents the total number of samples in the training data;

[0092] S23. Construct Gaussian process model GP:

[0093] f(E)~GP(m(E),k(E,E ′ ));

[0094] Among them, m(E) represents the mean function, k(E,E ′ ) represents the covariance function, E and E ′are any two environmental data respectively, f(E) represents the mapping function for predicting image enhancement parameters for the input environmental perception data E;

[0095] S24, use the training data set D to fit the Gaussian process model GP to obtain the prediction function f * (E):

[0096]

[0097] Among them, K represents the training environment dataset E train The covariance matrix on , represents the variance of the observation noise, I is the unit matrix, P train Representation and training environment dataset E train The corresponding image enhancement parameter value, m(E train ) represents the mean function in the training environment dataset E train The value vector on ;

[0098] S25. Record the current environment data as E current , using the prediction function f * (E) Predict the current environment data and obtain the preliminary image enhancement parameter P init , calculate the prediction uncertainty σ corresponding to the current environmental data current , according to the preset dynamic weight function α(σ current ) Fusion of preliminary image enhancement parameters and standard image enhancement parameters P baseline Get the image enhancement parameter distribution results:

[0099] P current =α(σ current )·P init +[1-α(σ current )]·P baseline ;

[0100] Among them, P current is the image enhancement parameter distribution result.

[0101] The present invention uses a Gaussian process model to model the relationship between environmental perception data and image enhancement parameters, and uses historical image quality scores to construct a training data set, thereby achieving accurate prediction and dynamic adjustment of image enhancement parameter distribution results. The method maps current environmental data into preliminary image enhancement parameters, and combines the prediction uncertainty with the standard image enhancement parameters through a preset dynamic weight function to generate accurate image enhancement parameter distribution results. Through this method, not only can the model's response speed to environmental changes be improved, but also the parameter prediction error can be reduced, ensuring that the camera module always outputs high-quality images in complex dynamic environments. Overall, the present invention significantly improves the image enhancement effect and the system's adaptive adjustment capability, so that the image can still maintain clear details and true colors under conditions such as insufficient lighting, backlighting, rain and fog, providing more stable and reliable visual data support for autonomous driving, security monitoring and other high-demand applications, but also provides an accurate data basis for subsequent system optimization, and has high practical application value.

[0102] In this embodiment, S3 specifically includes:

[0103] S31, receiving the output image enhancement parameter distribution result P current , as the basic reference information before the zebra optimization algorithm searches, and as the core basis for constructing the parameter search space of the image enhancement module;

[0104] S32. Constructing initial value ranges for the corresponding parameters based on the image enhancement parameter distribution results, and generating a joint search space covering multiple dimensions including brightness enhancement factor, contrast adjustment coefficient, Retinex component weight, and image detail restoration strength.

[0105] S33. In the constructed joint search space, according to the set population size N, the initial population individual positions of the zebra optimization algorithm are generated using a uniform distribution strategy:

[0106] Z init ={z j ∣z j =P current +∈ j ,∈ j ~U(-Δ,Δ),j=1,2,…,N};

[0107] Among them, Z init Represents the initial population position set of the zebra optimization algorithm, z j Indicates the position of the jth individual in the initial population, ∈ j represents the random variable used to perturb the distribution of image enhancement parameters, Δ represents the search perturbation amplitude, and N represents the number of individuals in the initial population;

[0108] S34, setting the search boundary range of each enhancement parameter, including the maximum adjustable value and the minimum adjustable value, and dynamically adjusting the boundary interval width in combination with the confidence interval of the image enhancement parameter distribution result;

[0109] S35. Configure the update behavior control factors of the zebra optimization algorithm, including jump direction factor, behavior response probability, population update frequency, and local disturbance intensity;

[0110] S36. Input the initialized population position, search boundary range and update behavior control factor into the image enhancement module to complete the parameter search space initialization configuration.

[0111] The present invention constructs a joint search space using image enhancement parameter distribution results, achieving precise initialization of image enhancement module parameters, thereby providing accurate basic reference information for the Zebra optimization algorithm. The method first maps each image enhancement parameter to a corresponding initial value range based on the image enhancement parameter distribution results generated by environmental perception data, effectively covering multiple key dimensions such as brightness enhancement factor, contrast adjustment coefficient, Retinex component weight, and image detail restoration strength. Subsequently, the initial population position is generated through a uniform distribution strategy, allowing the Zebra optimization algorithm to be distributed over a wider parameter space during the initial search phase, thereby avoiding local optimality and accelerating the global search convergence rate. Furthermore, by setting the search boundary range and dynamically adjusting the boundary width based on the confidence interval of the parameter distribution results, the present invention greatly improves the robustness and adaptability of the search process. At the same time, configuring update behavior control factors such as jump direction factor, behavior response probability, population update frequency, and local disturbance intensity provides a flexible control method for subsequent parameter updates. Overall, this solution not only optimizes the setting of initial parameters and shortens the algorithm iteration time, but also provides higher accuracy and stability for image enhancement processing, significantly improving the image output quality of the camera module in complex dynamic environments, fully demonstrating the superior performance of the present invention in practical applications.

[0112] In this embodiment, the S4 specifically includes:

[0113] S41, receiving the parameter search space S of the generated image enhancement module search , which includes the initial population position set Z init , search boundary range and initial update behavior control factor and original image data I raw ;

[0114] S42, setting the number of iterative optimization rounds T;

[0115] S43. For the tth iteration, select a candidate solution from the current parameter search space Each candidate solution represents a set of image enhancement parameters;

[0116] S44, using the image quality gradient prediction module based on deep neural network, for each candidate solution Generated candidate enhanced images Calculate local detail gradients And generate the corresponding search step and jump probability prediction value

[0117] S45. Set of candidate solutions Calculate the statistic, the mean is μ t , the standard deviation is σ t and skewness is κ t , and at the same time, calculate the global image quality improvement factor Γ t :

[0118]

[0119] in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the average image quality score of the candidate solution set, Represents the global image quality score, adopts a multi-dimensional adaptive weight adjustment strategy, and dynamically calculates the update factor β t :

[0120]

[0121] Among them, β0 is the preset basic update factor, λ is the global quality improvement weight coefficient, σ t represents the standard deviation of the candidate solution set in the tth iteration, μ t represents the mean of the candidate solution set in the tth iteration, κ t represents the skewness of the candidate solution set in the tth iteration;

[0122] S46. For each candidate solution Generated candidate enhanced images Use the image enhancement function G() to process and calculate the image quality score Q (j,t) :

[0123]

[0124] Among them, α, β, and γ are preset weight coefficients. Represents the candidate enhanced image and the reference image I ref The structural similarity index between represents the no-reference image quality evaluation index of the candidate enhanced image, Represents the color quality evaluation index of the candidate enhanced image, Iref is the reference image, N0 and U0 are normalization constants;

[0125] S47. Construct a mixed reward function R (j,t) The reward function comprehensively considers the image quality improvement rate, local detail restoration effect, and population convergence speed to evaluate each candidate solution:

[0126] R (j,t) =w1ΔQ (j,t) +w2 D (j,t) +w3 C (t) ;

[0127] Among them, D (j,t) represents the local detail recovery index corresponding to the candidate solution, C (t) Indicates the population convergence speed index, w1, w2, w3 are preset weight coefficients, ΔQ (j,t) Represents a candidate solution The image quality improvement of the generated candidate enhanced image in the tth iteration;

[0128] S48, using the update mechanism based on adaptive fuzzy control to update the mixed reward function R (j,t) and global image quality improvement factor Γ t Perform fuzzy membership evaluation and use the preset fuzzy membership function to update the control factor β t , generate the updated control factor β′ t :

[0129] β′ t =β t +K·(μ high (R (t) ,Γ t )-μ low (R (t) ,Γ t ));

[0130] Among them, K is the adjustment coefficient, μ high (R (t) ,Γ t ) and μ low (R (t) ,Γ t ) represent the mixed reward function R (t) and Γ t The calculated high-efficiency and low-efficiency membership values;

[0131] S49, based on the updated control factor β t and the jump probability prediction value Update candidate solutions:

[0132]

[0133] in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the new value of the jth candidate solution in the t+1th iteration after the update;

[0134] S410: After completing T iterations, select the candidate enhanced image with the highest image quality score as the final output result, and output the corresponding multiple image enhancement candidate results.

[0135] This invention implements online, multi-round iterative optimization of image enhancement parameters by introducing a deep neural network-based image quality gradient prediction and adaptive fuzzy control update mechanism into the image enhancement module. This method first utilizes a pre-generated parameter search space to generate multiple candidate enhanced images based on the original image data. A deep neural network is then used to extract local detail gradients and global quality score information, providing accurate search step and jump probability predictions for the candidate solutions. The system then calculates statistics of the candidate solution set and a global image quality improvement factor, and dynamically determines the update factor using a multidimensional adaptive weight adjustment strategy, effectively balancing global exploration and local optimization. Using an adaptive fuzzy control mechanism, a hybrid reward function is used to comprehensively evaluate the candidate solutions' image quality improvement, local detail restoration, and population convergence speed, enabling online optimization of the update factor and precise updating of the candidate solutions. Finally, after multiple iterations, the system automatically selects the candidate enhanced image with the highest image quality score as the final output. This solution significantly improves the image enhancement performance of camera modules in complex and dynamic environments, ensuring rich image detail, reduced noise, and accurate color reproduction. It also offers high real-time and robustness, providing more reliable visual data support for application scenarios such as autonomous driving, security monitoring, and industrial inspection.

[0136] In this embodiment, the S5 specifically includes:

[0137] S51, receiving and collecting multiple image enhancement candidate results, including image enhancement effects obtained by different parameter combinations;

[0138] S52, extracting an index that can reflect the image structure similarity for each image enhancement candidate result by comparing the structural feature information between the candidate image and the reference image;

[0139] S53. Extracting a no-reference image quality index for each image enhancement candidate result. The no-reference image quality index is based on the statistical characteristics and local information of the image itself and can represent the overall visual quality of the image without relying on an external reference.

[0140] S54. extracting an underwater image color quality index for each image enhancement candidate result, where the underwater image color quality index specifically captures the effectiveness of color distribution, saturation, and contrast of the image underwater, reflecting the image's performance in color reproduction;

[0141] S55. According to a pre-set weighting strategy, the weighting strategy sets fixed weights for the structural similarity index, the no-reference image quality index, and the underwater image color quality index, respectively, with the sum of the weight values being 1, reflecting the relative importance of the three indicators in the comprehensive visual quality evaluation, and weightedly fusing the information of the three indicators to obtain a comprehensive evaluation value;

[0142] S56. Generate a comprehensive image quality score for each image enhancement candidate result based on the weighted fusion of various indicator information.

[0143] The present invention implements a multi-dimensional image quality evaluation on multiple image enhancement candidate results, thereby achieving comprehensive visual quality quantification of the candidate images. This method extracts a structural similarity index, a no-reference image quality index, and an underwater image color quality index from the candidate images. The structural similarity index objectively reflects the consistency between the candidate image and the reference image in terms of detail and texture. The no-reference image quality index comprehensively measures the overall visual quality of the image through the image's own statistical features and local information. The underwater image color quality index specifically captures the image's color distribution, saturation, and contrast performance under specific circumstances. Based on a pre-set weighting strategy, which assigns fixed weights to each of the three indicators and sums them to 1, the method weights and fuses each indicator according to its importance to generate a comprehensive evaluation value, thereby accurately reflecting the overall quality of each candidate image. Through this comprehensive evaluation, the present invention can compare and select the candidate results with the best image quality, providing a reliable basis for subsequent image enhancement output, significantly improving the system's ability to determine image quality in complex environments and its stability. Practical applications have shown that this method has obvious advantages in improving image details, enhancing color reproduction and reducing noise, thereby ensuring that the camera module can output clear, realistic and high-visual-quality images in various complex scenarios, effectively meeting the strict image quality requirements in fields such as autonomous driving, security monitoring and industrial inspection.

[0144] Example 1:

[0145] To verify the feasibility of this invention, the system was applied indoors and outdoors at major traffic intersections and commercial areas in a specific city from October 2024 to March 2025, covering different seasons, lighting conditions, climates, and environmental interference factors. In these scenarios, traditional camera modules often suffer from blurred image details, increased noise, and color distortion due to factors such as instantaneous changes in ambient light, backlighting, and rain and fog. This seriously affects the image processing and subsequent target recognition accuracy of traffic monitoring, security, and autonomous driving assistance systems.

[0146] In this embodiment, we apply the image enhancement method based on the deep fusion of Bayesian optimization and improved zebra optimization algorithm proposed in the present invention to the intelligent traffic monitoring system. In the specific operation process, the real-time image and its corresponding environmental data, such as light intensity, color temperature, blur level and scene dynamic characteristics, are first collected through the camera module. The system uses the Bayesian optimization algorithm to construct a Gaussian process model, and generates the image enhancement parameter distribution in the current environment based on the mapping relationship between historical image quality scores and image enhancement parameters. Subsequently, this distribution result is used as the initialization basis to configure the parameter search space of the zebra optimization algorithm, including the initial population position, search boundary range and update behavior control factor.

[0147] In practical applications, multiple rounds of iterative optimization continuously adjust camera parameters to achieve intelligent optimization of candidate image enhancement results. During each iteration, the system extracts the local detail gradients and global image quality scores of the candidate images based on the deep neural network prediction module, dynamically calculates the update factor using a multi-dimensional adaptive weight adjustment strategy, and evaluates the candidate images using a hybrid reward function. After several iterations, the system automatically selects the candidate with the highest image quality score as the final output. Experimental data shows that this method can significantly improve image detail clarity and color reproduction in low-light, backlit, rainy, foggy, and high-contrast environments, while reducing image noise levels. The overall image quality score is improved by an average of 20% to 35% compared to traditional methods.

[0148] For example, in a backlit night scene at a transportation hub in downtown Beijing, the images output by traditional cameras showed obvious overexposure and local shadows, with a structural similarity index of only 0.68, a NIQE index as high as 5.6, and a UCIQE index of only 0.35, with an overall score below 0.55. After applying the method of the present invention, after automatic adjustment and image enhancement processing, the structural clarity and local detail restoration of the image were significantly improved, with the structural similarity index increased to 0.82, the NIQE reduced to 3.9, the UCIQE increased to 0.48, and the overall image quality score increased to 0.78. At the same time, in the indoor environment of the commercial area, due to insufficient lighting and noise interference, the color performance of the image of the traditional camera module was poor, with an overall score of only 0.60; however, by online optimization of camera parameters and image enhancement strategies, the enhanced image has improved color reproduction and detail performance, with an overall score of more than 0.75.

[0149] In the autonomous driving assistance scenario, we conducted real-time monitoring of the image quality of the camera module in a high-speed driving environment. During the experiment, when the vehicle was driving on the highway, the original image of the camera module was often blurred due to light flickering and high vehicle speed. After processing by this method, the system was able to complete environmental perception, parameter optimization and image enhancement processing within 50 milliseconds. The clarity and contrast of the image were significantly improved, ensuring the accurate recognition of obstacles and lane markings in front of the autonomous driving system. Field test data showed that in the autonomous driving environment, after processing by this method, the image quality was improved, the target detection accuracy rate increased from the original 82% to more than 93%, and the response time was reduced by about 15%, greatly improving driving safety.

[0150] Furthermore, in the field of industrial inspection, camera modules are often used to monitor product defects on production lines. Traditional systems are prone to misjudgments or missed detections due to factors such as lighting fluctuations and mechanical vibrations within the production workshop. The proposed method optimizes the contrast and detail of camera-captured images by adjusting image enhancement parameters in real time. This reduces the detection error rate from 8% to less than 3%, improving the accuracy and efficiency of product inspections.

[0151] Table 1 Statistics of camera module dynamic environment adaptation and image enhancement effect

[0152]

[0153] Table 1 shows the comparative data of the image enhancement effects of the traditional method and the camera module of the present invention in different application scenarios. From the data of the night backlight scene, the structural similarity index of the traditional method is only 0.68, while the present invention reaches 0.82, indicating that through the processing of the present invention, the image has been significantly improved in terms of structural details and texture retention. At the same time, the no-reference image quality index of the traditional method is relatively high, reaching 5.6, while the present invention reduces this index to 3.9, indicating that the noise and blur problems are effectively controlled; in terms of underwater image color index, the traditional method is only 0.35, while the present invention is improved to 0.48, and the comprehensive score is also improved from 0.55 to 0.78, further proving the advantage of the present invention in improving the overall image quality. In addition, in terms of response time, the processing time of the present invention is 60 milliseconds, which has obvious real-time advantages compared with the traditional method.

[0154] In low-light indoor scenes in commercial areas, although the various indicators of the traditional method have improved compared to the night-time backlit scene, its comprehensive score is only 0.60, far lower than the 0.75 of the present invention. The data show that the present invention can better restore image details and color information under low-light conditions, thereby improving image clarity and contrast. In high-speed autonomous driving scenarios, due to factors such as high vehicle speed and unstable lighting, the image quality indicators of the traditional method are even worse, with a comprehensive score of only 0.50. The present invention can increase the structural similarity index to 0.78 and the comprehensive score to 0.77, while reducing the response time to 50 milliseconds, ensuring efficient capture and timely processing of image information in high-speed dynamic environments.

[0155] The application scenario of industrial inspection production lines has extremely high requirements for image quality. Although the traditional method has slightly better indicators in this scenario, the comprehensive score is still only 0.62, while the present invention can achieve a comprehensive score of 0.80, while the response time is only 55 milliseconds. These data fully reflect that the present invention can significantly improve image quality in various typical application scenarios, especially in terms of structural details, noise control, and color reproduction. In general, the present invention not only has a significant improvement in image quality, but also excels in real-time performance, providing more reliable and efficient visual information support for multiple fields such as traffic monitoring, autonomous driving, and industrial inspection.

[0156] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A camera module dynamic environment adaptation and image enhancement method, characterized in that: The steps include: S1. Collect environmental perception data of the camera module in the target environment, including light intensity, color temperature information, blur level and scene dynamic characteristics; S2. Input the environmental perception data into the Bayesian optimization algorithm module, build a Gaussian process model, and generate image enhancement parameter distribution results based on the mapping relationship between historical image quality scores and image enhancement parameters; S3. Initialize and configure the parameter search space of the image enhancement module according to the image enhancement parameter distribution results, including the initial population position, search boundary range, and update behavior control factor of the zebra optimization algorithm; S4. In the image enhancement module, the image enhancement parameters are optimized for multiple rounds of iterations using the zebra optimization algorithm, the original image data is enhanced, and multiple image enhancement candidate results are generated; S5. Evaluate the image quality of multiple image enhancement candidate results by extracting information of image structure similarity index, no-reference image quality index, and underwater image color quality index, and generate a comprehensive image quality score based on weighted fusion; S6. Feedback the comprehensive image quality score to the Bayesian optimization algorithm module to update the Gaussian process model and improve the prediction accuracy and generalization ability of the next round of image enhancement parameter distribution results; S7. The image with the highest score is used as the final image enhancement output result, and the current environment data and the adopted image enhancement parameters constitute a sample pair and are stored in the database.

2. A camera module dynamic environment adaptation and image enhancement method according to claim 1, characterized in that: The S2 specifically includes: S21, record the environmental perception data as E={E1, E2, ..., E n }, where E i represents the i-th environmental feature; S22. Record the historical image quality score set as Q and the corresponding image enhancement parameter set as P to form a training dataset D: D={(E i ,Q i ,P i )∣i=1,2,…,m}; Among them, Q i Score the quality of the i-th image, P i is the corresponding image enhancement parameter, m represents the total number of samples in the training data; S23. Construct Gaussian process model GP: f(E)~GP(m(E),k(E,E ′ )); Among them, m(E) represents the mean function, k(E,E ′ ) represents the covariance function, E and E ′ are any two environmental data respectively, f(E) represents the mapping function for predicting image enhancement parameters for the input environmental perception data E; S24, use the training data set D to fit the Gaussian process model GP to obtain the prediction function f * (E): Among them, K represents the training environment dataset E train The covariance matrix on , represents the variance of the observation noise, I is the unit matrix, P train Representation and training environment dataset E train The corresponding image enhancement parameter value, m(E train ) represents the mean function in the training environment dataset E train The value vector on ; S25. Record the current environment data as E current , using the prediction function f * (E) Predict the current environment data and obtain the preliminary image enhancement parameter P init , calculate the prediction uncertainty σ corresponding to the current environmental data current , according to the preset dynamic weight function α(σ current ) Fusion of preliminary image enhancement parameters and standard image enhancement parameters P baseline Get the image enhancement parameter distribution results: P current =a(s current )·P init +[1-a(s current )]·P baseline ; Among them, P current is the image enhancement parameter distribution result.

3. A camera module dynamic environment adaptation and image enhancement method according to claim 1, characterized in that: The S3 specifically includes: S31, receiving the output image enhancement parameter distribution result P current , as the basic reference information before the zebra optimization algorithm searches, and as the core basis for constructing the parameter search space of the image enhancement module; S32. Constructing initial value ranges for the corresponding parameters based on the image enhancement parameter distribution results, and generating a joint search space covering multiple dimensions including brightness enhancement factor, contrast adjustment coefficient, Retinex component weight, and image detail restoration strength. S33. In the constructed joint search space, according to the set population size N, the initial population individual positions of the zebra optimization algorithm are generated using a uniform distribution strategy: Z init ={z j ∣z j =P current +∈ j ,∈ j ~U(-Δ,Δ),j=1,2,…,N}; Among them, Z init Represents the initial population position set of the zebra optimization algorithm, z j Indicates the position of the jth individual in the initial population, ∈ j represents the random variable used to perturb the distribution of image enhancement parameters, Δ represents the search perturbation amplitude, and N represents the number of individuals in the initial population; S34, setting the search boundary range of each enhancement parameter, including the maximum adjustable value and the minimum adjustable value, and dynamically adjusting the boundary interval width in combination with the confidence interval of the image enhancement parameter distribution result; S35. Configure the update behavior control factors of the zebra optimization algorithm, including jump direction factor, behavior response probability, population update frequency, and local disturbance intensity; S36. Input the initialized population position, search boundary range and update behavior control factor into the image enhancement module to complete the parameter search space initialization configuration.

4. A camera module dynamic environment adaptation and image enhancement method according to claim 1, characterized in that: The S4 specifically includes: S41, receiving the parameter search space S of the generated image enhancement module search , which includes the initial population position set Z init , search boundary range and initial update behavior control factor and original image data I raw ; S42, setting the number of iterative optimization rounds T; S43. For the tth iteration, select a candidate solution from the current parameter search space Each candidate solution represents a set of image enhancement parameters; S44, using the image quality gradient prediction module based on deep neural network, for each candidate solution Generated candidate enhanced images Calculate local detail gradients And generate the corresponding search step length and jump probability prediction value S45. Set of candidate solutions Calculate the statistic, the mean is μ t , the standard deviation is σ t and skewness is κ t , and at the same time, calculate the global image quality improvement factor Γ t : in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the average image quality score of the candidate solution set, Represents the global image quality score, adopts a multi-dimensional adaptive weight adjustment strategy, and dynamically calculates the update factor β t : Among them, β0 is the preset basic update factor, λ is the global quality improvement weight coefficient, σ t represents the standard deviation of the candidate solution set in the tth iteration, μ t represents the mean of the candidate solution set in the tth iteration, κ t represents the skewness of the candidate solution set in the tth iteration; S46. For each candidate solution Generated candidate enhanced images Use the image enhancement processing function G() to process and calculate the image quality score Q (j,t) : Among them, α, β, and γ are preset weight coefficients. Represents the candidate enhanced image and the reference image I ref The structural similarity index between represents the no-reference image quality evaluation index of the candidate enhanced image, Represents the color quality evaluation index of the candidate enhanced image, I ref is the reference image, N0 and U0 are normalization constants; S47. Construct a mixed reward function R (j,t) The reward function comprehensively considers the image quality improvement rate, local detail restoration effect, and population convergence speed to evaluate each candidate solution: R (j,t) =w1ΔQ (j,t) +w2 D (j,t) +w3 C (t) ; Among them, D (j,t) represents the local detail recovery index corresponding to the candidate solution, C (t) Indicates the population convergence speed index, w1, w2, w3 are preset weight coefficients, ΔQ (j,t) Represents a candidate solution The image quality improvement of the generated candidate enhanced image in the tth iteration; S48, using the update mechanism based on adaptive fuzzy control to update the mixed reward function R (j,t) and the global image quality improvement factor Γ t Perform fuzzy membership evaluation and update the control factor using the preset fuzzy membership function β t , generate the updated control factor β′ t : b′ t =b t +K·(μ high (R (t) ,C t )-m low (R (t) ,C t )); Among them, K is the adjustment coefficient, μ high (R (t) ,Γ t ) and μ low (R (t) ,Γ t ) represent the mixed reward function R (t) and Γ t The calculated high-efficiency and low-efficiency membership values; S49, based on the updated control factor β′ t and the jump probability prediction value Update candidate solutions: in, represents the candidate solution with the highest image quality score in the t-th iteration, represents the new value of the jth candidate solution in the t+1th iteration after the update; S410: After completing T iterations, select the candidate enhanced image with the highest image quality score as the final output result, and output the corresponding multiple image enhancement candidate results.

5. A camera module dynamic environment adaptation and image enhancement method according to claim 1, characterized in that: The S5 specifically includes: S51, receiving and collecting multiple image enhancement candidate results, including image enhancement effects obtained by different parameter combinations; S52, extracting an index that can reflect the image structure similarity for each image enhancement candidate result by comparing the structural feature information between the candidate image and the reference image; S53. Extracting a no-reference image quality index for each image enhancement candidate result. The no-reference image quality index is based on the statistical characteristics and local information of the image itself and can represent the overall visual quality of the image without relying on an external reference. S54. extracting an underwater image color quality index for each image enhancement candidate result, where the underwater image color quality index specifically captures the effectiveness of color distribution, saturation, and contrast of the image underwater, reflecting the image's performance in color reproduction; S55. According to a pre-set weighting strategy, the weighting strategy sets fixed weights for the structural similarity index, the no-reference image quality index, and the underwater image color quality index, respectively, with the sum of the weight values being 1, reflecting the relative importance of the three indicators in the comprehensive visual quality evaluation, and weightedly fusing the information of the three indicators to obtain a comprehensive evaluation value; S56. Generate a comprehensive image quality score for each image enhancement candidate result based on the weighted fusion of various indicator information.

Images