Image acquisition enhancement processing method and system of camera based on deep learning

Abstract

The invention discloses an image acquisition enhancement processing method and system for a camera based on deep learning, and relates to the field of image acquisition enhancement, and the method comprises the steps: carrying out the preliminary feature extraction according to an acquisition device state, an acquisition content state and an acquisition environment state; boundary limitation is set, so that the accuracy of predicting the parameter time sequence of the point spread function (PSF) is improved; constructing a PSF parameter time sequence prediction model by using a deep learning algorithm, and performing time sequence prediction on PSF parameters; constructing a device dominant motion sub-library, a content dominant motion sub-library and a mixed dominant motion sub-library; constructing a lightweight neural network model, and quickly deducing PSF parameters; and on the basis of Wiener filtering deconvolution, the dynamically acquired image is corrected and enhanced. The method has the advantages that the problem that PSF parameters cannot adapt to a complex dynamic environment in image enhancement processing of Wiener filtering deconvolution is effectively solved, and the dynamic image acquisition capability of a dynamic camera is enhanced.

Description

Technical Field

[0001] This invention relates to the field of image acquisition enhancement, specifically to an image acquisition enhancement processing method and system based on deep learning for cameras. Background Technology

[0002] With the rapid development of fields such as intelligent monitoring, autonomous driving, and drone aerial photography, the image acquisition quality in dynamic scenes has become a key technical bottleneck. During the acquisition process, dynamic cameras often suffer from motion blur and loss of detail due to equipment movement (such as hand shaking or carrier bumps), movement of the subject (such as high-speed vehicles or moving pedestrians), and interference from complex environments (such as sudden changes in lighting or fog scattering). These degraded images seriously affect the accuracy of subsequent target detection, tracking, and recognition. Therefore, the demand for real-time enhancement processing of dynamic images is becoming increasingly urgent.

[0003] The point spread function (PSF) is a core parameter describing the image degradation process, and its accuracy directly determines the effect of deconvolution enhancement. Traditional methods restore images by pre-setting a fixed PSF model (such as a Gaussian blur kernel). However, in dynamic scenes, the PSF is dynamically affected by the device's motion state, the motion characteristics of the captured content, and environmental interference. Fixed models are difficult to adapt to complex changes. In recent years, deep learning technology has provided a new approach for PSF prediction, but it still faces challenges in multi-source state feature fusion, temporal dynamic modeling, and lightweight deployment. Therefore, how to construct a PSF parameter derivation mechanism that is both robust and real-time, and combine it with Wiener filtering deconvolution to achieve dynamic image enhancement, has become a research hotspot in the field of computer vision.

[0004] Traditional point spread function (PSF) combined with Wiener filtering deconvolution technology is difficult to adapt to complex dynamic environments. Due to the separate analysis of PSF and Wiener filtering deconvolution, physical constraints are not integrated into the processing flow, which can easily lead to artifacts or loss of details in the deconvolution results. Secondly, although optimizing PSF with deep learning improves the prediction accuracy of PSF, the large number of parameters and computational overhead brought by the complex network structure makes it difficult to deploy in embedded cameras with limited computing power, and cannot meet the real-time requirements of dynamic acquisition. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a deep learning-based image acquisition enhancement processing method and system for cameras. This technical solution solves the problems mentioned in the background, such as the fragmented analysis of PSF and Wiener filtering deconvolution, which fails to incorporate physical constraints into the processing flow, easily leading to artifacts or loss of details in the deconvolution results, and the large number of parameters and computational overhead brought by the complex network structure, making it difficult to deploy in embedded cameras with limited computing power and failing to meet the real-time requirements of dynamic acquisition.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based image acquisition enhancement processing method for cameras includes: Preliminary feature extraction is performed based on the status of the acquisition device, the status of the acquired content, and the status of the acquisition environment. Based on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment, the boundary constraints of the point spread function (PSF) are set to improve the accuracy of time series prediction of PSF parameters. Based on the extracted preliminary features, a PSF parameter time series prediction model is constructed using deep learning algorithms to achieve time series prediction of PSF parameters. Based on the technical architecture of predicting PSF parameters, we construct device-dominated motion sub-libraries, content-dominated motion sub-libraries, and hybrid-dominated motion sub-libraries. A lightweight neural network model is built based on the motion sub-library, and the PSF parameters are quickly derived by using the PSF parameter template matching results and combining them with actual data. Based on Wiener filtering and deconvolution processing, combined with the derived PSF parameter data, dynamically acquired images are corrected and enhanced.

[0007] Preferably, the step of setting boundary constraints for the point spread function (PSF) based on the status of the acquisition device, the status of the acquired content, and the status of the acquisition environment to improve the accuracy of time-series prediction of PSF parameters specifically includes: Based on the physical correlation between the camera's optical characteristics and motion parameters, a maximum physical upper limit threshold for the PSF blur kernel size is set. Based on the relative motion direction between the acquisition device state and the acquired content state, a range interval for the PSF blur kernel angle and a deviation threshold for the velocity direction are set, where the angle range is... If the deviation between the predicted PSF fuzzy kernel angle and the relative velocity direction angle exceeds the deviation threshold, the deviation needs to be corrected. Calculate the mean and standard deviation of the PSF blur intensity in the dataset, and restrict the predicted PSF parameter values ​​to within a certain range. The portion exceeding the range is truncated and resampled, forcibly set to [a certain value]. ; We independently statistically analyzed each parameter of PSF under different scenarios, calculated the quantile of each parameter, and took the 95th quantile of historical data as the hard upper limit for PSF parameter prediction to ensure that the model output does not exceed this value. The expression for setting the maximum physical upper limit threshold of the PSF fuzz kernel size is: In the formula, This represents the upper limit of the PSF blur kernel size length. For camera focal length, In pixels For maximum relative velocity, This is the maximum exposure time for the camera; The expression for correcting the deviation is as follows: In the formula, To predict the deviation between the PSF blur kernel angle and the relative velocity direction angle, The deviation threshold, To predict the PSF blur kernel angle, The angle is the relative velocity direction. This is the attenuation coefficient.

[0008] Preferably, the step of constructing a PSF parameter time-series prediction model based on the extracted preliminary features using a deep learning algorithm to achieve time-series prediction of PSF parameters specifically includes: A three-channel LSTM neural network is set up to extract temporal features from the preliminary features of the acquisition device status, the acquisition content status, and the acquisition environment status; An attention mechanism is set up to perform feature fusion on the temporal features of the acquisition device status and the acquisition content status, the temporal features of the acquisition device status and the acquisition environment status, and the temporal features of the acquisition content status and the acquisition environment status, respectively. Based on the relative motion entropy and the environmental scattering coefficient, and using the sigmoid function, fusion weights for the three fusion features are constructed respectively. Based on the fusion weights of the three fusion features, a fusion feature is constructed that integrates the state of the acquisition device, the state of the acquisition content, and the state of the acquisition environment. With the aim of minimizing the error between the PSF predicted value and the true value, a PSF loss function is constructed based on the mean square error formula and combined with boundary constraints. A PSF parameter time series prediction model is constructed based on deep learning algorithms to achieve time series prediction of PSF parameters; The fusion weight expression for constructing the three fusion features is as follows: In the formula, Weights are used to fuse the temporal features of the acquisition device status and the acquired content status. Weights are used to fuse the temporal features of the acquisition device status and the acquisition environment status. To fuse the temporal features of the collected content status and the collected environment status into weights, , , The normalized fusion weights The normalized relative motion entropy, The normalized environmental scattering coefficient is... , This is a constant term, which can be determined through a grid search of the validation set; The fusion feature expression of the acquisition device status, acquisition content status, and acquisition environment status is as follows: In the formula, The integration of installation, content, and environment. It is a fusion feature of device state characteristics and content state characteristics. It is a fusion of device state characteristics and environmental state characteristics. This is a fusion feature of content state characteristics and environmental state characteristics; ⊙ is the symbol for the Hadamard product. The expression for the PSF loss function is: In the formula, This is the combined loss value between the PSF predicted value and the actual value. The loss value is based on the mean squared error. The PSF blur kernel size penalty loss term is used. The angle penalty loss term for the PSF blur kernel. The loss term is the PSF fuzziness intensity penalty term. The PSF quantile hard upper limit penalty loss term, , , , The penalty term weights are determined as hyperparameters through a grid search on the validation set.

[0009] Preferably, the technical architecture based on predicted PSF parameters, which constructs a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library, specifically includes: Based on the PSF parameter time series prediction model, adversarial test experiments were conducted, introducing parameters under different scenarios and environments to carry out simulation experiments and obtain the simulation data; By comparing the simulated data with the actual data, the difference is used as feedback data to correct the PSF parameter time series prediction model. Obtain the corrected simulation data and construct device-dominated motion sub-libraries, content-dominated motion sub-libraries, and hybrid-dominated motion sub-libraries respectively. Based on the device-dominated motion sub-library, content-dominated motion sub-library, and hybrid-dominated motion sub-library, motion pattern recognition, PSF parameter template calling strategy, template fast indexing mechanism, and library update mechanism are established respectively.

[0010] Preferably, the step of constructing a lightweight neural network model based on a motion sub-library and rapidly deriving PSF parameters using PSF parameter template matching results and actual data specifically includes: Based on filtering algorithms and normalization formulas, the actual acquisition device status, acquisition content status, and acquisition environment status data are filtered and normalized. Based on the actual acquisition device status, acquisition content status, and acquisition environment status, the PSF parameter template in the motion sub-library is matched, and the matched PSF parameter template parameters are output. The data acquisition device status, data acquisition content status, data acquisition environment status, and the matched PSF parameter template parameters, similarity, and template priority are used as inputs, and the optimized PSF core parameters are used as outputs. Three depthwise separable convolutional layers are set up to encode features of data on acquisition device status, acquisition content status, and acquisition environment status. A two-layer MLP network is set up. By inputting the data of the acquisition device status, the acquisition content status and the acquisition environment status, the matched PSF parameter template parameter data, as well as the similarity and template priority, the output is the normalized PSF parameter activated by Sigmoid. The normalized PSF parameters are mapped to the actual physical domain through a linear transformation, and the deviation rate between the mapped PSF parameters and the PSF parameter template matching results is calculated. Based on the deviation rate between the mapped PSF parameters and the matching results of the PSF parameter template, a PSF parameter fusion model is constructed, and the PSF parameters are output through the model. If the deviation rate exceeds the preset value, the current PSF parameters and image features are fed back to the corresponding motion sub-library, triggering the library update mechanism; The expression for the PSF parameter fusion model is: In the formula, The final output of the model is the PSF parameter. These are the mapped PSF parameters. The PSF parameters are the results of matching the PSF parameter template. These are weighting coefficients, which can be tuned using the validation set. These are correction coefficients used to limit the magnitude of corrections to the model parameters. , The set deviation rate threshold, and , This represents the deviation rate between the mapped PSF parameters and the PSF parameters of the template matching result.

[0011] Furthermore, this solution proposes a deep learning-based image acquisition enhancement processing system for cameras, used to implement the deep learning-based image acquisition enhancement processing method for cameras described above, including: The feature extraction module is used to perform preliminary feature extraction based on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment. The PSF parameter derivation module is used to set boundary constraints of the point spread function (PSF) based on the state of the acquisition device, the state of the acquired content, and the state of the acquisition environment, thereby improving the accuracy of PSF parameter time-series prediction. Based on the extracted preliminary features, a deep learning algorithm is used to construct a PSF parameter time-series prediction model to achieve time-series prediction of PSF parameters. Based on the technical architecture for predicting PSF parameters, a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid dominant motion sub-library are constructed. A lightweight neural network model is constructed based on the motion sub-libraries, and PSF parameters are quickly derived using PSF parameter template matching results and actual data. An image enhancement module is used to correct and enhance dynamically acquired images based on Wiener filter deconvolution processing and combined with derived PSF parameter data.

[0012] Preferably, the PSF parameter derivation module includes: A boundary limiting unit is used to set boundary limits for the point spread function (PSF) based on the state of the acquisition device, the state of the acquisition content, and the state of the acquisition environment, thereby improving the accuracy of time-series prediction of PSF parameters. The parameter prediction unit is used to construct a PSF parameter time series prediction model based on the extracted preliminary features using a deep learning algorithm, so as to achieve time series prediction of PSF parameters. The library building unit is used to construct a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library based on the technical architecture of the predicted PSF parameters. A fast parameter derivation unit is used to construct a lightweight neural network model based on a motion sub-library, and to quickly derive PSF parameters using PSF parameter template matching results and actual data.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This solution provides a deep learning-based image acquisition enhancement method for cameras. By setting boundary constraints for the point spread function (PSF) and employing deep learning algorithms to construct a temporal prediction model for PSF parameters, it effectively predicts the temporal parameters of the PSF, thereby improving the accuracy of generated PSF parameters. Simultaneously, it effectively solves the problem of artifacts or loss of detail in the deconvolution results caused by the fragmented analysis of PSF and Wiener filtering deconvolution. Furthermore, to address the issue that deep learning-based PSF prediction models have a large number of parameters and are computationally time-consuming, making them difficult to deploy in embedded cameras with limited computing power and unable to meet the real-time requirements of dynamic acquisition, this solution conducts adversarial testing experiments using the PSF parameter temporal prediction model, introducing parameters under different scenarios and environments. Simulation experiments were conducted, and simulated data was acquired. Through analysis and processing of the simulation data, a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library were constructed. This motion sub-library provides PSF parameter templates for real-time prediction of subsequent PSF parameters, thereby addressing the real-time requirements of dynamic acquisition. Simultaneously, to improve the accuracy of dynamic image acquisition, a lightweight neural network model was constructed. Using the PSF parameter template matching results and combining them with actual data, PSF parameters were quickly derived, ensuring the accuracy of PSF parameter acquisition. Finally, Wiener filtering deconvolution technology was used, combined with PSF parameters, to correct and enhance dynamically acquired images, improving the camera's ability to acquire and adapt dynamic images in dynamic scenes. Attached Figure Description

[0014] Figure 1 This is a flowchart of the deep learning-based image acquisition and enhancement processing method for cameras according to the present invention. Figure 2 The flowchart of the present invention describes the preliminary feature extraction process based on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment. Figure 3 The flowchart of the present invention describes how to set boundary constraints of the point spread function (PSF) based on the state of the acquisition device, the state of the acquisition content, and the state of the acquisition environment to improve the accuracy of time series prediction of PSF parameters. Figure 4 Based on the extracted preliminary features, the present invention uses a deep learning algorithm to construct a PSF parameter time series prediction model to realize the time series prediction flowchart of PSF parameters; Figure 5 To construct a lightweight neural network model based on a motion sub-library in this invention, the PSF parameter template matching results are used to quickly derive the PSF parameter flowchart by combining actual data. Figure 6The flowchart below illustrates the process of correcting and enhancing dynamically acquired images using Wiener filtering deconvolution processing based on derived PSF parameter data, as described in this invention. Detailed Implementation

[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0016] Reference Figure 1 As shown, a deep learning-based image acquisition enhancement processing method for cameras includes: Preliminary feature extraction is performed based on the status of the acquisition device, the status of the acquired content, and the status of the acquisition environment. Based on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment, the boundary constraints of the point spread function (PSF) are set to improve the accuracy of time series prediction of PSF parameters. Based on the extracted preliminary features, a PSF parameter time series prediction model is constructed using deep learning algorithms to achieve time series prediction of PSF parameters. Based on the technical architecture of predicting PSF parameters, we construct device-dominated motion sub-libraries, content-dominated motion sub-libraries, and hybrid-dominated motion sub-libraries. A lightweight neural network model is built based on the motion sub-library, and the PSF parameters are quickly derived by using the PSF parameter template matching results and combining them with actual data. Based on Wiener filtering and deconvolution processing, combined with the derived PSF parameter data, dynamically acquired images are corrected and enhanced.

[0017] Understandably, combining PSF parameters with Wiener filtering deconvolution technology primarily addresses the image quality degradation caused by imaging system blurring in dynamic scenes, such as motion blur and optical system distortion. PSF parameters accurately describe the blurring characteristics during the imaging process, and Wiener filtering deconvolution, based on these parameters, can effectively reverse-correct blurred images, reducing blur and artifacts caused by dynamic acquisition and restoring image details and sharpness. However, traditional deconvolution is susceptible to problems such as poor image enhancement, detail loss, or distortion due to dynamic and environmental interference. Therefore, this solution leverages the powerful inference capabilities and fitting ability of deep learning to construct a PSF parameter time-series prediction model. Adversarial testing experiments are then conducted using this model, introducing... Simulation experiments were conducted on parameters under the same scenario and environment, and the simulated data was obtained. Through the analysis and processing of the simulated data, a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library were constructed. This motion sub-library provides PSF parameter templates for subsequent real-time prediction of PSF parameters, thereby solving the real-time requirement of dynamic acquisition. At the same time, to improve the accuracy of actual image dynamic acquisition, a lightweight neural network model was constructed. The PSF parameter template matching results were used to quickly derive PSF parameters in combination with actual data, thereby quickly and accurately obtaining PSF parameters. Finally, through the synergy of Wiener filtering deconvolution and accurate PSF parameters, dynamic image correction and enhancement were achieved, improving the camera's ability to acquire and adapt dynamic images in dynamic scenes.

[0018] Reference Figure 2 As shown, the preliminary feature extraction based on the status of the acquisition device, the status of the acquired content, and the status of the acquisition environment specifically includes: By triggering hardware synchronization, data on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment are acquired respectively. Linear interpolation is used to align asynchronously arriving state data to the target time sequence, ensuring that the temporal dimension of the feature vectors in each frame is consistent. The categorical features in the data of the acquisition device status, the acquisition content status, and the acquisition environment status are converted into binary vector form through one-hot encoding. By using filtering algorithms and normalization formulas, the collected device status, collected content status, and collected environment status data are filtered and normalized to eliminate the influence of data peaks and data dimensions. Based on the Lie group feature extraction method, the geometric structure and group action invariance of the Lie group manifold are used to perform preliminary feature extraction on the processed state of the acquisition device. Based on the optical flow field vorticity-divergence decomposition method, the optical flow field of the acquired content state is analyzed into rotational and stretching components, thereby realizing the preliminary feature extraction of the processed acquired content state. Based on differential coding, the difference sequence between adjacent data points is used to perform preliminary feature extraction on the processed collection environment status. Based on the initially extracted feature data, a feature dataset is constructed that includes the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment.

[0019] Understandably, the original data regarding the acquisition device status, acquired content status, and acquisition environment status often contains a large amount of noise, redundancy, or irrelevant information. This information can interfere with the training and prediction capabilities of subsequent PSF parameter time series prediction models, leading to decreased model performance and an inability to accurately capture key features in the data. However, preliminary feature extraction can filter out useless information, improve data quality, enhance the model's adaptability to the data, and thus improve the accuracy of PSF parameter time series prediction. This lays a solid foundation for subsequent image correction and enhancement based on Wiener filtering and deconvolution. The specific implementation method of the Lie group feature extraction method is as follows: The motion parameters of the device are mapped to the SO(3) Lie group space to construct feature points on the manifold; By utilizing the group action invariance of Lie groups, the features of the manifold are projected onto the tangent space through a logarithmic mapping to obtain the eigenvectors in the Euclidean space. The magnitude, angle, and other statistical properties of the tangent space vectors are extracted as the final features; This method can effectively preserve the geometric symmetry of the device's motion, avoid feature distortion caused by coordinate system transformation, and is directly related to the physical causes of the fuzzy kernel angle and size in PSF. Specific implementation method of the optical flow field vorticity-divergence decomposition: Based on the Helmholtz decomposition theorem, the optical flow field vector field is decomposed to obtain rotational and scaling components. Gaussian filtering is used to smooth the optical flow field and reduce noise interference; The mean value of vorticity, directional entropy, and extreme values ​​of divergence are calculated as features describing the motion pattern of the content.

[0020] This decomposition directly corresponds to the rotation angle and size scaling of the PSF blur kernel, providing the model with feature dimensions that are directly related to the cause of blur. The specific implementation method of the differential encoding is as follows: Perform first-order difference on environmental time-series data; Sliding window statistics are performed on the mean, variance, and peak value within the difference sequence window; By combining the baseline values ​​of the original data, feature vectors for the baseline and differential trends are constructed. This method highlights the dynamic impact of environmental changes on image quality, avoiding the limitation of static features in capturing transient interference.

[0021] Reference Figure 3 As shown, the step of setting boundary constraints for the point spread function (PSF) based on the status of the acquisition device, the status of the acquired content, and the status of the acquisition environment to improve the accuracy of time-series prediction of PSF parameters specifically includes: Based on the physical correlation between the camera's optical characteristics and motion parameters, a maximum physical upper limit threshold for the PSF blur kernel size is set. Based on the relative motion direction between the acquisition device state and the acquired content state, a range interval for the PSF blur kernel angle and a deviation threshold for the velocity direction are set, where the angle range is... If the deviation between the predicted PSF fuzzy kernel angle and the relative velocity direction angle exceeds the deviation threshold, the deviation needs to be corrected. Calculate the mean and standard deviation of the PSF blur intensity in the dataset, and restrict the predicted PSF parameter values ​​to within a certain range. The portion exceeding the range is truncated and resampled, forcibly set to [a certain value]. ; We independently statistically analyzed each parameter of PSF under different scenarios, calculated the quantile of each parameter, and took the 95th quantile of historical data as the hard upper limit for PSF parameter prediction to ensure that the model output does not exceed this value. The expression for setting the maximum physical upper limit threshold of the PSF fuzz kernel size is: In the formula, This represents the upper limit of the PSF blur kernel size length. For camera focal length, In pixels For maximum relative velocity, This is the maximum exposure time for the camera; The expression for correcting the deviation is as follows: In the formula, To predict the deviation between the PSF blur kernel angle and the relative velocity direction angle, The deviation threshold, To predict the PSF blur kernel angle, The angle is the relative velocity direction. This is the attenuation coefficient.

[0022] Understandably, the PSF parameter is essentially a mathematical description of "motion blur and optical distortion" during the imaging process. Its values ​​must conform to optical principles and kinematic laws. Without boundary constraints, deep learning models may output parameters that violate physical laws due to training data bias or extreme sample interference, leading to the failure of subsequent deconvolution processing or a break in the physical correlation between PSF and Wiener filtering deconvolution, resulting in artifacts or loss of detail in the deconvolution results. Therefore, this solution analyzes and limits the physical correlation between camera optical characteristics and motion parameters, the relative motion direction between the acquisition device state and the acquired content state, the mean and standard deviation of PSF blur intensity, and the quantiles of various PSF parameters under different scenarios to ensure the physical validity of the PSF parameter. Among these, the deviation threshold... The angle of the PSF fuzzy kernel can be determined through a validation set grid search. The PSF parameters can be derived through subsequent steps.

[0023] Reference Figure 4 As shown, the step of constructing a PSF parameter time-series prediction model based on the extracted preliminary features using a deep learning algorithm to achieve time-series prediction of PSF parameters specifically includes: A three-channel LSTM neural network is set up to extract temporal features from the preliminary features of the acquisition device status, the acquisition content status, and the acquisition environment status; An attention mechanism is set up to perform feature fusion on the temporal features of the acquisition device status and the acquisition content status, the temporal features of the acquisition device status and the acquisition environment status, and the temporal features of the acquisition content status and the acquisition environment status, respectively. Based on the relative motion entropy and the environmental scattering coefficient, and using the sigmoid function, fusion weights for the three fusion features are constructed respectively. Based on the fusion weights of the three fusion features, a fusion feature is constructed that integrates the state of the acquisition device, the state of the acquisition content, and the state of the acquisition environment. With the aim of minimizing the error between the PSF predicted value and the true value, a PSF loss function is constructed based on the mean square error formula and combined with boundary constraints. A PSF parameter time series prediction model is constructed based on deep learning algorithms to achieve time series prediction of PSF parameters; The fusion weight expression for constructing the three fusion features is as follows: In the formula, Weights are used to fuse the temporal features of the acquisition device status and the acquired content status. Weights are used to fuse the temporal features of the acquisition device status and the acquisition environment status. To fuse the temporal features of the collected content status and the collected environment status into weights, , , The normalized fusion weights The normalized relative motion entropy, The normalized environmental scattering coefficient is... , This is a constant term, which can be determined through a grid search of the validation set; The fusion feature expression of the acquisition device status, acquisition content status, and acquisition environment status is as follows: In the formula, The integration of installation, content, and environment. It is a fusion feature of device state characteristics and content state characteristics. It is a fusion of device state characteristics and environmental state characteristics. This is a fusion feature of content state characteristics and environmental state characteristics; ⊙ is the symbol for the Hadamard product. The expression for the PSF loss function is: In the formula, This is the combined loss value between the PSF predicted value and the actual value. The loss value is based on the mean squared error. The PSF blur kernel size penalty loss term is used. The angle penalty loss term for the PSF blur kernel. The loss term is the PSF fuzziness intensity penalty term. The PSF quantile hard upper limit penalty loss term, , , , The penalty term weights are determined as hyperparameters through a grid search on the validation set.

[0024] Understandably, traditional deconvolution is susceptible to problems such as poor image enhancement, loss of detail, or distortion due to dynamic and environmental interference. By leveraging the powerful inference capabilities and fitting abilities of deep learning to various parameters, a temporal prediction model for PSF parameters is constructed, thereby improving the accuracy of PSF parameters and enhancing the camera's ability to acquire and adapt dynamic images in dynamic scenes. The three-channel LSTM neural network specifically includes: Dynamic Time Warping (DTW) aligns the initial characteristics of the acquisition device status, acquisition content status, and acquisition environment status to a unified timestamp. Based on the preliminary characteristics of the acquisition device status, acquisition content status, and acquisition environment status after timestamp alignment, a fixed-dimensional three-channel tensor is constructed. Based on the LSTM neural network, temporal features are extracted from the preliminary features in each channel, and then uniformly mapped through a fully connected layer. The attention-based mechanism specifically includes: The temporal features output by the three-channel LSTM are expanded into a sequence of feature vectors; Based on the Transformer-based attention mechanism algorithm, attention weights are constructed for the temporal features of the acquisition device state and the acquisition content state, the temporal features of the acquisition device state and the acquisition environment state, and the temporal features of the acquisition content state and the acquisition environment state. The data acquisition device status, data acquisition content status, and data acquisition environment status features are combined into three feature pairs and then concatenated into a query matrix. The key and value both use the device status, data acquisition content status, and environment features. The weights of each pair of features are calculated using scaled dot product attention. The outputs are concatenated along the channel dimension and then integrated and enhanced through a linear transformation to obtain the final three pairs of fused features.

[0025] The technical architecture based on predicted PSF parameters, which constructs a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library, specifically includes: Based on the PSF parameter time series prediction model, adversarial test experiments were conducted, introducing parameters under different scenarios and environments to carry out simulation experiments and obtain the simulation data; By comparing the simulated data with the actual data, the difference is used as feedback data to correct the PSF parameter time series prediction model. Obtain the corrected simulation data and construct device-dominated motion sub-libraries, content-dominated motion sub-libraries, and hybrid-dominated motion sub-libraries respectively. Based on the device-dominated motion sub-library, content-dominated motion sub-library, and hybrid-dominated motion sub-library, motion pattern recognition, PSF parameter template calling strategy, template fast indexing mechanism, and library update mechanism are established respectively.

[0026] Understandably, the PSF parameter temporal prediction model is limited by its large number of parameters and computational time, making it difficult to deploy in embedded cameras with limited computing power and thus unable to meet the real-time requirements of dynamic acquisition. However, the PSF parameter temporal prediction model can be used to preprocess image acquisition under different environments and scenarios, thereby obtaining simulated data through adversarial testing experiments. This allows for the construction of device-dominated motion sub-libraries, content-dominated motion sub-libraries, and hybrid-dominated motion sub-libraries. These libraries, combined with subsequent lightweight neural network models, can effectively solve the problem of the PSF parameter temporal prediction model being limited by its large number of parameters and computational time, making it difficult to deploy in embedded cameras with limited computing power and unable to meet the real-time requirements of dynamic acquisition. Specifically, the motion pattern recognition includes: By using a linear weighted formula, the relationship between linear velocity, angular velocity and kinetic energy is fitted, and the ratio of kinetic energy between the state of the acquisition device and the state of the acquired content is obtained through kinetic energy. Based on expert scoring, corresponding intervals and boundary thresholds are set for different motion modes. , ,and ; Based on the kinetic energy ratio, and combined with the corresponding intervals and boundary thresholds for different kinetic modes, the kinetic mode is determined, as follows: when At that time, it was determined that the motion was dominated by the data acquisition device; when At that time, it was determined that the movement was dominated by a mixture of the data acquisition device and the data being acquired; when At that time, it is determined that the content being collected is the dominant movement; The relationship between the linear velocity and angular velocity of motion and the kinetic energy, and the expression for the kinetic energy ratio, are as follows: In the formula, This represents the energy value of movement. The weights for the linear velocities of motion can be determined using the least squares method. The normalized linear velocity of the motion. The normalized angular velocity, The ratio of kinetic energy. To collect the motion energy value of the content state, The value represents the kinetic energy of the data acquisition device, where the subscript... and This indicates the status of the data acquisition device and the status of the acquired data. and Based on The formula is obtained from it; The PSF parameter template invocation strategy specifically includes: Based on the device-dominated motion sub-library, the content-dominated motion sub-library, and the hybrid-dominated motion sub-library, generate corresponding PSF parameter templates; Based on the identified dominant motion type, PSF parameter templates in the corresponding sub-library are retrieved first. If no matching template is found in the corresponding sub-database, record the missing type, trigger other sub-databases to perform a search, record the search results of other sub-databases, and mark them. The PSF parameter template fast indexing mechanism specifically includes: The dominant motion type, motion state, and scene features are encoded and fused to generate a unique label; The hash forest algorithm is used to map high-dimensional composite tags to low-dimensional hash codes, and multiple hash trees are used for parallel retrieval and location. When new templates are added or the scenario changes, the index tags are automatically updated, and recall and precision are regularly calculated. The automatic index tag update process is as follows: Generate exclusive composite tags for new templates based on new templates or changes in scenarios; Replace the old label with a new composite label, and recalculate the low-dimensional hash code using the Hash Forest algorithm; The library update mechanism specifically includes: When a new relative motion pattern is detected, its feature vector is automatically calculated and added to the corresponding sub-library using the PSF parameter time series prediction model; Set template call priority by setting a tiered range and combining the number of times the same template is called to increase the template call priority; Incremental learning is adopted, and the template is only partially updated for new data, avoiding full database retraining and reducing the computing power consumption of edge devices.

[0027] Reference Figure 5 As shown, the construction of a lightweight neural network model based on a motion sub-library, and the rapid derivation of PSF parameters using PSF parameter template matching results and actual data, specifically includes: Based on filtering algorithms and normalization formulas, the actual acquisition device status, acquisition content status, and acquisition environment status data are filtered and normalized. Based on the actual acquisition device status, acquisition content status, and acquisition environment status, the PSF parameter template in the motion sub-library is matched, and the matched PSF parameter template parameters are output. The data acquisition device status, data acquisition content status, data acquisition environment status, and the matched PSF parameter template parameters, similarity, and template priority are used as inputs, and the optimized PSF core parameters are used as outputs. Three depthwise separable convolutional layers are set up to encode features of data on acquisition device status, acquisition content status, and acquisition environment status. A two-layer MLP network is set up. By inputting the data of the acquisition device status, the acquisition content status and the acquisition environment status, the matched PSF parameter template parameter data, as well as the similarity and template priority, the output is the normalized PSF parameter activated by Sigmoid. The normalized PSF parameters are mapped to the actual physical domain through a linear transformation, and the deviation rate between the mapped PSF parameters and the PSF parameter template matching results is calculated. Based on the deviation rate between the mapped PSF parameters and the matching results of the PSF parameter template, a PSF parameter fusion model is constructed, and the PSF parameters are output through the model. If the deviation rate exceeds the preset value, the current PSF parameters and image features are fed back to the corresponding motion sub-library, triggering the library update mechanism; The expression for the PSF parameter fusion model is: In the formula, The final output of the model is the PSF parameter. These are the mapped PSF parameters. The PSF parameters are the results of matching the PSF parameter template. These are weighting coefficients, which can be tuned using the validation set. These are correction coefficients used to limit the magnitude of corrections to the model parameters. , The set deviation rate threshold, and , This represents the deviation rate between the mapped PSF parameters and the PSF parameters of the template matching result.

[0028] Understandably, due to the complexity of the environment, acquisition device status, and content status during actual dynamic image acquisition, calling the corresponding PSF parameter template based solely on the device-dominated motion sub-library, content-dominated motion sub-library, and hybrid dominant motion sub-library may lead to deviations between the PSF parameters and the actual situation, resulting in poor quality of the final dynamic image acquisition. Therefore, to improve the accuracy of actual dynamic image acquisition, it is necessary to construct a lightweight neural network model. This model combines real-time acquired acquisition device status, content status, and environment status data with the matched PSF parameter template parameters to output the final PSF parameters. Therefore, this solution provides a PSF parameter fusion model. By judging the deviation rate threshold range between the mapped PSF parameters and the PSF parameter template matching results, different PSF parameter output models are set to improve the accuracy of the final output PSF parameters. Here, the deviation rate refers to the relative error between the model-mapped PSF parameters and the template matching results, and the set deviation rate threshold... , This can be determined through a validation set grid search.

[0029] Reference Figure 6 As shown, the process of correcting and enhancing dynamically acquired images based on Wiener filtering deconvolution, combined with the derived PSF parameter data, specifically includes: The PSF parameters output by the lightweight model are converted into a Wiener-compatible format to generate a two-dimensional Gaussian blur kernel matrix and a noise power spectrum matrix. Perform range validation on PSF parameters, and replace outliers with the median value of the same type of template in the sub-library; The PSF parameters are decomposed in the frequency domain based on the Fourier transform formula, and the Wiener filter kernel is constructed based on the decomposed PSF parameters. The weight of the noise-to-signal power ratio is dynamically adjusted based on the adaptation of the motion sub-library type. Based on the Wiener filter kernel and combined with the acquired content status image, the corrected acquired content status image is obtained through Wiener filter deconvolution processing. Set a residual threshold, calculate the residual energy between the first corrected image and the original image, determine whether the residual energy between the first corrected image and the original image is greater than the residual threshold, if so, update the PSF parameters and rebuild the filter kernel to perform a second deconvolution, iterate until the residual energy is less than the residual threshold or the maximum number of iterations is reached, if not, end the deconvolution process and output the corrected image. If the structural similarity of the corrected image is below the threshold, the current PSF parameters and image features are fed back to the corresponding sub-library, triggering the library update mechanism.

[0030] It is understandable that Wiener filtering deconvolution technology is an effective technique for recovering motion blur, reducing blur and artifacts caused by dynamic acquisition, and restoring image details and sharpness. Its image recovery accuracy is directly affected by the PSF parameters. Therefore, this scheme can effectively correct and enhance dynamically acquired images by improving the quality of the PSF parameters. Furthermore, it judges the quality of the corrected image based on the residual energy between the corrected and original images, combined with a residual threshold, and incorporates a feedback loop in the motion sub-library update mechanism, greatly improving the adaptability and update capability of the motion sub-library. The expression for the residual energy between the first corrected image and the original image is: In the formula, The residual energy between the corrected image and the original image, For the original image in coordinates Pixel value at that location, To correct the image in coordinates Pixel value at that location, The size of the captured image.

[0031] Furthermore, based on the same inventive concept as the aforementioned deep learning-based image acquisition and enhancement processing method for cameras, this solution proposes a deep learning-based image acquisition and enhancement processing system for cameras, comprising: The feature extraction module is used to perform preliminary feature extraction based on the status of the acquisition device, the status of the acquisition content, and the status of the acquisition environment. The PSF parameter derivation module is used to set boundary constraints of the point spread function (PSF) based on the state of the acquisition device, the state of the acquired content, and the state of the acquisition environment, thereby improving the accuracy of PSF parameter time-series prediction. Based on the extracted preliminary features, a deep learning algorithm is used to construct a PSF parameter time-series prediction model to achieve time-series prediction of PSF parameters. Based on the technical architecture for predicting PSF parameters, a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid dominant motion sub-library are constructed. A lightweight neural network model is constructed based on the motion sub-libraries, and PSF parameters are quickly derived using PSF parameter template matching results and actual data. An image enhancement module is used to correct and enhance dynamically acquired images based on Wiener filter deconvolution processing and combined with derived PSF parameter data. The PSF parameter derivation module includes: A boundary limiting unit is used to set boundary limits for the point spread function (PSF) based on the state of the acquisition device, the state of the acquisition content, and the state of the acquisition environment, thereby improving the accuracy of time-series prediction of PSF parameters. The parameter prediction unit is used to construct a PSF parameter time series prediction model based on the extracted preliminary features using a deep learning algorithm, so as to achieve time series prediction of PSF parameters. The library building unit is used to construct a device-dominated motion sub-library, a content-dominated motion sub-library, and a hybrid-dominated motion sub-library based on the technical architecture of the predicted PSF parameters. A fast parameter derivation unit is used to construct a lightweight neural network model based on a motion sub-library, and to quickly derive PSF parameters using PSF parameter template matching results and actual data.

[0032] In summary, the advantages of this invention are: it effectively solves the problem that the PSF parameters in Wiener filtering deconvolution cannot adapt to complex dynamic environments in image enhancement processing, and greatly enhances the ability of dynamic cameras to acquire dynamic images.

[0033] The foregoing has shown and described 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 principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A deep learning-based camera image acquisition enhancement processing method, characterized in that, The application relates to a method for predicting a point spread function (PSF) parameter time sequence, and belongs to the technical field of image processing. According to the acquisition device state, the acquisition content state and the acquisition environment state, preliminary feature extraction is performed on the states respectively; According to the acquisition device state, the acquisition content state and the acquisition environment state, the boundary limit of a point spread function (PSF) is set, and the accuracy of PSF parameter time sequence prediction is improved; Based on the extracted preliminary features, a PSF parameter time sequence prediction model is constructed by using a deep learning algorithm to realize the time sequence prediction of the PSF parameter; Based on the technical architecture of the predicted PSF parameter, a device dominant motion sub-library, a content dominant motion sub-library and a mixed dominant motion sub-library are constructed; Based on the motion sub-library, a lightweight neural network model is constructed, and the PSF parameter template matching result is used to quickly deduce the PSF parameter in combination with actual data; Based on Wiener filter deconvolution processing, the PSF parameter data deduced is combined to correct and enhance the dynamic acquisition image.

2. The image acquisition enhancement processing method based on deep learning of the camera according to claim 1, characterized in that, The preliminary feature extraction of the acquisition device state, the acquisition content state and the acquisition environment state specifically comprises the following steps: The acquisition device state, the acquisition content state and the acquisition environment state data are acquired through hardware synchronization triggering; Linear interpolation is adopted to align the asynchronously-arriving state data to a target time sequence, so that the time dimension of each feature vector is consistent; The category type features in the acquisition device state, the acquisition content state and the acquisition environment state data are converted into binary vector forms through one-hot encoding; The acquisition device state, the acquisition content state and the acquisition environment state data are filtered and normalized through a filtering algorithm and a normalization formula, so that the influence of data peak values and data dimensions is eliminated; Based on the Lie group feature extraction method, the geometric structure and group action invariance of the Lie group manifold are used to perform preliminary feature extraction on the processed acquisition device state; Based on the light flow field vortex-diffusion decomposition method, the light flow field of the acquisition content state is analyzed into a rotation component and a stretching component, so that preliminary feature extraction is realized on the processed acquisition content state; Based on the difference coding method, the difference sequence between adjacent data points is used to perform preliminary feature extraction on the processed acquisition environment state; According to the preliminary feature data extracted, the acquisition device state, the acquisition content state and the acquisition environment state feature data set is constructed.

3. The image acquisition enhancement processing method based on deep learning of the camera according to claim 2, characterized in that, The boundary limit of the PSF is set according to the acquisition device state, the acquisition content state and the acquisition environment state, and the accuracy of PSF parameter time sequence prediction is improved, and the specific steps comprise the following steps: Based on the physical correlation between the camera optical characteristics and the motion parameters, the maximum physical upper limit threshold of the PSF blur kernel size is set; Based on the relative motion direction of the acquisition device state and the acquisition content state, the range interval of the PSF blur kernel angle and the deviation threshold of the speed direction are set, wherein the range interval of the angle is , the deviation between the PSF blur kernel angle and the relative speed direction angle exceeds the deviation threshold, and the deviation needs to be corrected. The mean and standard deviation of the PSF blur strength in the data set are computed, and the PSF parameter prediction is limited to the interval, with the excess part forced to zero by truncation and resampling. The PSF parameters in different scenes are independently counted, the quantile of each parameter is calculated, and the 95% quantile of the historical data is taken as the hard upper limit of the PSF parameter prediction, so that the model output is ensured to be not more than the value; The expression of the maximum physical upper limit threshold of the PSF blur kernel size is as follows: wherein is the PSF blur kernel size length upper limit, is the camera focal length, is the pixel size, is the maximum relative speed, is the camera maximum exposure time; The expression of the need for deviation correction is as follows: wherein is a predicted PSF blur kernel angle and is a deviation threshold, is a predicted PSF blur kernel angle, is a relative velocity direction angle, is a decay coefficient.

4. The image acquisition enhancement processing method based on deep learning of the camera according to claim 3, characterized in that, The PSF parameter time sequence prediction model is constructed by using the deep learning algorithm based on the extracted preliminary features, and the specific steps comprise the following steps: A three-channel LSTM neural network is set to extract time sequence features of the preliminary features of the acquisition device state, acquisition content state and acquisition environment state; An attention mechanism is set to fuse features of the acquisition device state and acquisition content state time sequence features, acquisition device state and acquisition environment state time sequence features, and acquisition content state and acquisition environment state time sequence features; Based on the relative motion entropy and the environmental scattering coefficient, a fusion weight of the three fused features is constructed based on a sigmoid function; Based on the fusion weight of the three fused features, a fusion feature of the acquisition device state, acquisition content state and acquisition environment state is constructed; Based on the mean square error formula and combined with boundary constraints, a PSF loss function is constructed to minimize the error between the predicted value and the true value of the PSF; Based on the deep learning algorithm, a PSF parameter time sequence prediction model is constructed to realize the time sequence prediction of the PSF parameter; The expression of the fusion weight of the three fused features is: In the formula, is the fusion weight of the time sequence characteristics of the acquisition device state and the acquisition content state, is the fusion weight of the time sequence characteristics of the acquisition device state and the acquisition environment state, is the fusion weight of the time sequence characteristics of the acquisition content state and the acquisition environment state, , , is the normalized fusion weight, is the normalized relative motion entropy, is the normalized environment scattering coefficient, , is a constant term, which can be determined by grid search verification set; The expression of the fusion feature of the acquisition device state, acquisition content state and acquisition environment state is: wherein is a fusion feature of device, content and environment, is a fusion feature of device state feature and content state feature, is a fusion feature of device state feature and environment state feature, is a fusion feature of content state feature and environment state feature, and is a Hadamard product symbol. The expression of the PSF loss function is: wherein, is the overall loss value of PSF prediction and true value, is the loss value based on mean square error, is the PSF blur kernel size penalty loss term, is the PSF blur kernel angle penalty loss term, is the PSF blur strength penalty loss term, is the PSF quantile hard upper bound penalty loss term, , , , is the penalty term weight, determined as a hyperparameter by grid search on validation set.

5. The image acquisition enhancement processing method based on deep learning of the camera according to claim 4, characterized in that, The technical architecture based on the predicted PSF parameter includes: Based on the PSF parameter time sequence prediction model, an adversarial testing experiment is carried out, different parameters under different situations and environments are introduced, a simulation test is carried out, and simulated data is obtained; By comparing the simulated data with the actual data, the difference is obtained as feedback data to correct the PSF parameter time sequence prediction model; The corrected simulation data is obtained, and the device dominant motion sub-library, content dominant motion sub-library and mixed dominant motion sub-library are constructed; According to the device dominant motion sub-library, content dominant motion sub-library and mixed dominant motion sub-library, motion pattern recognition, PSF parameter template calling strategy, template fast indexing mechanism and library updating mechanism are established.

6. The image acquisition enhancement processing method based on deep learning of the camera according to claim 5, characterized in that, Based on the motion sub-library, a lightweight neural network model is constructed, and the PSF parameter is quickly derived based on the PSF parameter template matching result and the actual data, which specifically includes: Based on the filtering algorithm and the normalization formula, the actual acquisition device state, acquisition content state and acquisition environment state data are filtered and normalized; According to the actual acquisition device state, acquisition content state and acquisition environment state, the PSF parameter template in the motion sub-library is matched, and the matched PSF parameter template parameter is outputted; The acquisition device state, acquisition content state, acquisition environment state, matched PSF parameter template parameter, similarity and template priority are taken as input, and the optimized PSF core parameter is taken as output; A 3-layer depth separable convolution is set to encode the acquisition device state, acquisition content state and acquisition environment state data; A 2-layer MLP network is set to input the feature encoded acquisition device state, acquisition content state and acquisition environment state data, matched PSF parameter template parameter data, and similarity and template priority, and output the normalized PSF parameter activated by Sigmoid. The normalized PSF parameter is mapped to an actual physical range through a linear transformation, and a deviation rate of a mapping result of the PSF parameter from a PSF parameter template is calculated; A PSF parameter fusion model is constructed according to the deviation rate of the mapping result of the PSF parameter from the PSF parameter template, and the PSF parameter is output through the model; If the deviation rate exceeds a preset value, the current PSF parameter and image features are fed back to a corresponding motion sub-library, and a library updating mechanism is triggered; The PSF parameter fusion model expression is: In the formula, is the final output PSF parameter of the model, is the mapped PSF parameter, is the PSF parameter matched with the PSF parameter template, is the weight coefficient, which can be optimized through a verification set, is the correction coefficient, used to limit the correction range of the model parameter, , is the set deviation rate threshold, and , is the deviation rate of the mapped PSF parameter and the template matched PSF parameter.

7. The image acquisition enhancement processing method based on deep learning of the camera according to claim 6, characterized in that, The Wiener filter deconvolution processing is combined with the derived PSF parameter data to correct and enhance the dynamic acquisition image, and specifically includes: The PSF parameter output by the lightweight model is converted into a Wiener filter compatible format to generate a two-dimensional Gaussian blur kernel matrix and a noise power spectrum matrix; The PSF parameter is range-verified, and an abnormal value is replaced by a median value of a same type template in the sub-library; The PSF parameter is frequency-domain decomposed based on a Fourier transform formula, and a Wiener filter kernel is constructed according to the decomposed PSF parameter; The weight of the noise-signal power ratio is dynamically adjusted according to the adaptation of the motion sub-library type; The Wiener filter kernel is combined with the acquisition content state image to obtain a corrected acquisition content state image through Wiener filter deconvolution processing; A residual threshold is set, the residual energy of the first corrected image and the original image is calculated, and it is determined whether the residual energy of the first corrected image and the original image is greater than the residual threshold. If yes, the PSF parameter is updated and the filter kernel is reconstructed to perform secondary deconvolution, and iteration is performed until the residual energy is less than the residual threshold or the maximum iteration number is reached. If no, the deconvolution processing is ended, and the corrected image is output. If the structural similarity of the corrected image is lower than a threshold, the current PSF parameter and image features are fed back to a corresponding sub-library, and a library updating mechanism is triggered.

8. A deep learning based camera image acquisition enhancement processing system, characterized in that, The image acquisition enhancement processing method based on deep learning of the camera is implemented, and the method includes: A feature extraction module is configured to perform preliminary feature extraction on the acquisition device state, the acquisition content state and the acquisition environment state, respectively; A PSF parameter derivation module is configured to set boundary limits of a point spread function (PSF) according to the acquisition device state, the acquisition content state and the acquisition environment state, to improve the accuracy of the time series prediction of the PSF parameter, to construct a PSF parameter time series prediction model based on the extracted preliminary features and a deep learning algorithm, to realize the time series prediction of the PSF parameter, to construct a device dominant motion sub-library, a content dominant motion sub-library and a mixed dominant motion sub-library based on the technical architecture of the predicted PSF parameter, and to quickly derive the PSF parameter based on the PSF parameter template matching result and actual data; An image enhancement module is configured to correct and enhance the dynamic acquisition image based on the Wiener filter deconvolution processing and the derived PSF parameter data.

9. The deep learning based camera image acquisition enhancement processing system of claim 8, wherein, The PSF parameter derivation module includes: A boundary restriction unit is configured to set a boundary restriction of a point spread function (PSF) according to a state of a collection device, a state of collection content, and a state of a collection environment, and improve accuracy of time sequence prediction of PSF parameters; A parameter prediction unit is configured to construct a PSF parameter time sequence prediction model by using a deep learning algorithm based on the extracted preliminary features, so as to realize time sequence prediction of the PSF parameters; A library establishing unit is configured to construct a device dominant motion sub-library, a content dominant motion sub-library, and a hybrid dominant motion sub-library based on a technical architecture of the predicted PSF parameters; A parameter rapid derivation unit is configured to construct a lightweight neural network model based on the motion sub-library, and rapidly derive the PSF parameters by combining actual data and a PSF parameter template matching result.

Images