Unmanned aerial vehicle aerial image imaging optimization method and device fusing deep learning perception mechanism and physical modeling

Abstract

The invention discloses an unmanned aerial vehicle aerial image imaging optimization method and device fusing a deep learning perception mechanism and physical modeling. The method comprises the following steps: acquiring an original image frame obtained in a flight process of an unmanned aerial vehicle; inputting the image into a MobileViT illumination estimation network, extracting local convolution perception and multi-scale global semantic features, and outputting a scene illumination intensity estimation value; constructing a differentiable imaging parameter reasoning module based on an illumination physical modeling relationship, reversely deducing an optimal exposure parameter combination of a current frame, and constructing a parameter optimization module based on a perceptual error; combining the difference between the reconstructed image and the target image in the semantic perception space to construct a multi-loss function joint training model, and optimizing an exposure combination; deploying an edge computing platform for the trained network model to complete parameter prediction, control feedback and image acquisition link closed loop; according to the method, exposure optimization is realized before imaging, image gamma decoding and target enhancement are realized after imaging, and the image quality in low-light and backlight scenes is improved.

Description

Technical Field

[0001] This invention relates to the field of image perception and intelligent imaging technology, and more specifically to a method and apparatus for optimizing aerial image imaging by unmanned aerial vehicles by integrating deep learning perception mechanisms and physical modeling. Background Technology

[0002] With the widespread application of drones in complex environments such as security patrol, traffic monitoring, power line inspection, emergency response, and agricultural remote sensing, the requirements for image quality of airborne vision systems are becoming increasingly stringent. However, in actual flight missions, drones often need to pass through areas with extreme lighting changes, such as tunnel entrances, shaded areas under trees, twilight skylines, or nighttime streets, which are not ideal lighting conditions. Affected by the drastic fluctuations in ambient light, the acquired images often suffer from severe underexposure, image blurring, loss of detail, and even local saturation. Especially in areas with weak target light, the target outline is unclear and the contrast is low, which seriously affects the accuracy and stability of subsequent target detection, recognition, and tracking tasks.

[0003] The root cause of the above problems is that most current UAV vision systems use fixed or simplified imaging parameter control strategies, that is, image acquisition is performed under preset exposure time, shutter speed and ISO sensitivity, and the camera parameters are not adaptively adjusted according to the actual ambient lighting conditions, resulting in limited image quality at the source of imaging; and even if some cameras use automatic exposure (AE) mechanisms, their exposure adjustment strategies are mostly based on global brightness histogram analysis, lacking an understanding of the target area, image semantics or task-related areas, and still cannot guarantee that the imaging is friendly to the task target.

[0004] Currently, most common image enhancement methods in the industry focus on the post-processing stage after image acquisition, including histogram equalization, Retinex brightness reconstruction, denoising filtering, gamma correction, and low-light enhancement. While these methods can improve the visual quality of images to some extent, they generally have the following three limitations:

[0005] (1) Lost image information cannot be recovered: If the original signal in the dark area is severely underexposed or saturated during imaging, even post-enhancement algorithms will find it difficult to recover the original details.

[0006] (2) The enhancement process is uncontrollable and easily introduces artifacts: Over-enhancement can lead to side effects such as image noise amplification, color distortion, and edge blurring.

[0007] (3) Ignoring the structural mapping between imaging parameters and image quality: Most image enhancement models do not consider how the image is formed and cannot reflect the actual impact of exposure combination on the image.

[0008] Traditional rule-driven or physical modeling-based methods are no longer adequate for the increasingly complex scenarios, dynamic environments, and diverse tasks required for UAV applications. With the rapid development of deep learning technology in image perception, using neural network models to automatically learn the complex nonlinear mapping between image content and the physical exposure process has become a forward-looking development direction. In recent years, some studies have begun to focus on tasks such as illumination estimation, automatic exposure recommendation, and image gamma inverse correction. However, these studies are still mainly based on static image analysis and have not yet formed a complete closed-loop framework of "image → illumination → parameters → feedback control". In particular, they lack comprehensive consideration of the real-time performance, adaptability, and deployment efficiency of UAV flight missions.

[0009] Therefore, how to dynamically learn the ambient light intensity based on images or video streams, predict the key physical parameters that affect imaging quality, and achieve low-latency, high-stability automatic adjustment of imaging parameter combinations during UAV flight, ultimately enabling adaptive enhancement of dim targets during imaging, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] In view of this, the present invention provides a method and apparatus for optimizing UAV aerial imagery by integrating deep learning perception mechanisms and physical modeling to solve some of the technical problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] An imaging optimization method for UAV aerial images that integrates deep learning perception mechanisms and physical modeling includes the following steps:

[0013] S1. Collect raw image frames acquired by the drone during flight;

[0014] S2. Input the image into the illumination estimation network built on MobileViT, extract local convolutional perception and multi-scale global semantic features, and output the scene illumination intensity estimate;

[0015] S3. Construct a differentiable imaging parameter inference module. Based on the quantization relationship between illumination intensity and imaging parameters, decode the optimal combination of exposure parameters for the current frame in reverse according to the predicted scene illumination intensity estimate and the expected exposure amount, and construct a parameter optimization module based on perceptual error.

[0016] S4. Based on the illumination error loss and gamma recovery loss, combined with the receptive domain reconstruction constraint, construct a joint loss function, jointly train the network model including the illumination estimation network, the imaging parameter inference module and the imaging parameter optimization module, and optimize the exposure parameter combination;

[0017] S5. Real-time acquisition of the current frame image during the drone's flight, input into the trained network model for online inference and dynamic adjustment of exposure parameter combinations, to complete parameter exposure prediction, control feedback and acquisition of the next frame image.

[0018] Preferably, the illumination estimation network built on MobileViT includes a local perception module, a Transformer coding module, and a multi-scale fusion module;

[0019] The specific content of step S2 is as follows:

[0020] S21. Local Perceptual Coding Module: Utilizes deep separable convolutions to extract fine-grained brightness, edge, and texture features from the image while maintaining computational efficiency, resulting in an initial representation feature map F. c ;

[0021] S22. Lightweight Transformer Encoding Module: Transforms the local feature map F c The space is divided into fixed-size, non-overlapping patch blocks, which are then linearly embedded and input into a simplified multi-head self-attention module to model the illumination consistency and brightness shift characteristics of distant regions in the image, and output the global context feature F. t ;

[0022] S23. Multi-scale Enhanced Fusion Module: Based on a cross-scale semantic fusion strategy, it integrates local features F c With global context features F t The data are concatenated and then remapped and weighted using 1×1 convolution to output the final fused features.

[0023] Preferably, the initial representation feature map is as follows:

[0024] F c =Conv 3×3 (I input )

[0025] The output features of the multi-head self-attention module are:

[0026] F t =MHSA(LN(F c ))+F c

[0027] Where MHSA represents a simplified self-attention mechanism, and LN represents a layer normalization operation;

[0028] Output final fused features:

[0029] F out =Concat(F c ,F t →Conv 1×1 .

[0030] Preferably, in step S3, the quantization relationship between light intensity and imaging parameters is as follows:

[0031]

[0032] Where E is the effective exposure, L is the scene light intensity, t is the shutter speed, and f is the aperture value;

[0033] In the inference phase, based on the predicted light intensity estimate... With expected exposure E target By working backward, we can obtain the optimal parameter combination {t,f,ISO} that satisfies the current shooting task, where ISO is the gain factor involved in the optimization.

[0034] Preferably, in step 3, the parameter optimization module based on perceptual error introduces perceptual domain reconstruction constraints to minimize the structural differences between the predicted image and the reference image in the high-level semantic space. By introducing a semantically perceptual feature extraction function, the texture, edge and structural information in the image are explicitly modeled to guide the search for the optimal solution of the exposure parameters.

[0035] The optimization objective for the exposure parameter combination is:

[0036]

[0037] Among them, I pred For an image captured based on the current exposure combination {t,f,ISO}, I gt φ is a reference image acquired under standard imaging conditions; φ(·) is a high-order feature extraction function of a pre-trained feature extraction network used to capture the structural semantics of the image. is the Euclidean distance norm, used to measure image dissimilarity in a high-dimensional semantic space.

[0038] Preferably, the joint loss function is:

[0039]

[0040] Where λ1, λ2, and λ3 are loss weighting factors used to dynamically adjust the network to achieve the optimal balance between physical accuracy and perceived visual quality. For illumination estimation error, For gamma-ray inverse reconstruction error, To predict the structural differences between the image and the reference image in the high-level semantic space.

[0041] Preferably, the illumination estimation error Used to measure the light intensity estimate output by network regression. With realistic lighting label L i Mean square error between:

[0042]

[0043] The gamma-ray inverse reconstruction error term is used to evaluate the brightness fidelity of the image after inverse gamma mapping.

[0044]

[0045] Among them, I input For the input image luminance channel, I target For the reference brightness image, the gamma coefficient is γ; is a hyperparameter, and is a non-negative hyperparameter used to balance the contribution ratio of illumination regression and gamma reconstruction in the total loss.

[0046] Preferably, step S5 includes the following:

[0047] S51. Deploy the trained network model to the embedded edge computing platform on the drone;

[0048] S52. Acquire image frame I from the camera at the current timestamp t according to the set frame rate. t The forward prediction process is completed within a total delay of no more than 15ms from image acquisition to processing output, to obtain the estimated light intensity value of the current frame.

[0049] S53. Based on the predicted illumination value Automatically match the preset exposure lookup table or perform real-time numerical solving to obtain a set of optimal solutions {t} that can be used to adjust imaging parameters. * ,f * ISO *} t To meet the target exposure constraint:

[0050]

[0051] Among them, t * For exposure time, f * This refers to the aperture value and ISO. * For sensor gain parameters;

[0052] S54. The obtained exposure parameter combination is transmitted to the UAV camera control unit via the interface control module. The exposure settings are preloaded before the next frame image is acquired through the standard camera API or by directly accessing the CMOS image sensor driver interface.

[0053] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for optimizing drone aerial imagery by integrating deep learning perception mechanisms and physical modeling.

[0054] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the aforementioned method for optimizing drone aerial image imaging by integrating deep learning perception mechanisms and physical modeling.

[0055] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and apparatus for optimizing UAV aerial image imaging by integrating deep learning perception mechanism and physical modeling, which has the following beneficial effects:

[0056] (1) A lightweight illumination estimation network is constructed based on MobileViT to learn the nonlinear relationship between image and ambient illumination. An exposure decoder is constructed based on the physical imaging formula to realize the reverse calculation and optimization of imaging parameters. This realizes a bidirectional mapping mechanism from image to exposure parameters, fills the technical gap of traditional enhancement methods ignoring the control of imaging mechanism, and improves the initiative and interpretability of image quality control in low-light environments.

[0057] (2) By jointly introducing illumination error, gamma restoration error and perceptual loss function, the prediction results can achieve a balance between physical accuracy and semantic consistency, thereby improving the clarity and structural restoration capability of low-light images.

[0058] (4) The illumination estimation model constructed using the MobileViT network has the characteristics of being lightweight, high-precision, and highly deployable, meeting the strict requirements of embedded UAV platforms for computing resources and response speed. Through end-to-end deployment on the embedded platform, it can dynamically sense changes in ambient illumination during flight and adjust the exposure strategy in real time, thereby improving image acquisition quality and task completion. It is suitable for application scenarios such as night reconnaissance, underground inspection, and backlight recognition. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0060] Figure 1 Flowchart of the UAV aerial image imaging optimization method that integrates deep learning perception mechanism and physical modeling provided by the present invention;

[0061] Figure 2 A schematic diagram of the illumination estimation network based on MobileViT provided by this invention;

[0062] Figure 3 This is a schematic diagram illustrating the reasoning behind the combination of exposure parameters provided by the present invention;

[0063] Figure 4 This is a schematic diagram of the principle of the embedded edge computing platform for drones after the deployment model provided by the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] This invention discloses an imaging optimization method for UAV aerial images that integrates deep learning perception mechanisms and physical modeling, such as... Figure 1 This includes the following steps:

[0066] S1. Collect raw image frames acquired by the drone during flight;

[0067] S2. Input the image into the illumination estimation network built on MobileViT, extract local convolutional perception and multi-scale global semantic features, and output the scene illumination intensity estimate;

[0068] S3. Construct a differentiable imaging parameter inference module. Based on the quantization relationship between illumination intensity and imaging parameters, decode the optimal combination of exposure parameters for the current frame in reverse according to the predicted scene illumination intensity estimate and the expected exposure amount, and construct a parameter optimization module based on perceptual error.

[0069] S4. Based on the illumination error loss and gamma recovery loss, combined with the receptive domain reconstruction constraint, construct a joint loss function, jointly train the network model including the illumination estimation network, the imaging parameter inference module and the imaging parameter optimization module, and optimize the exposure parameter combination;

[0070] S5. Real-time acquisition of the current frame image during the drone's flight, input into the trained network model for online inference and dynamic adjustment of exposure parameter combinations, to complete parameter exposure prediction, control feedback and acquisition of the next frame image.

[0071] To further implement the above technical solutions, such as Figure 2To balance model accuracy and deployment performance, a lightweight Transformer structure, MobileViT, is used as the perceptual backbone. MobileViT integrates the local modeling capabilities of CNN and the global semantic modeling capabilities of Transformer.

[0072] The illumination estimation network built on MobileViT includes a local perception module, a Transformer coding module, and a multi-scale fusion module.

[0073] The specific content of step S2 is as follows:

[0074] S21. Local Perceptual Coding Module: Utilizes deep separable convolutions to extract fine-grained brightness, edge, and texture features from the image while maintaining computational efficiency, resulting in an initial representation feature map F. c ;

[0075] S22. Lightweight Transformer Encoding Module: Transforms the local feature map F c The space is divided into fixed-size, non-overlapping patch blocks, which are then linearly embedded and input into a simplified multi-head self-attention module to model the illumination consistency and brightness shift characteristics of distant regions in the image, and output the global context feature F. t ;

[0076] S23. Multi-scale Enhanced Fusion Module: Based on a cross-scale semantic fusion strategy, it integrates local features F c With global context features F t The data are concatenated and then remapped and weighted using 1×1 convolution to output the final fused features.

[0077] In this embodiment, MobileViT has advantages such as compact parameters, embeddable structure, and strong context awareness. It can effectively capture brightness changes and uneven regional illumination under complex lighting conditions, significantly improve the accuracy and robustness of illumination intensity estimation, and is suitable for real-time UAV imaging modeling tasks in edge computing scenarios.

[0078] To further implement the above technical solution, the initial characterization feature map is as follows:

[0079] F c =Conv 3×3 (I input )

[0080] The output features of the multi-head self-attention module are:

[0081] F t =MHSA(LN(F c ))+F c

[0082] Where MHSA represents a simplified self-attention mechanism, and LN represents a layer normalization operation;

[0083] Output final fused features:

[0084] F out =Concat(F c ,F t →Conv 1×1 .

[0085] To further implement the above technical solutions, such as Figure 3 In step S3, the quantization relationship between light intensity and imaging parameters is as follows:

[0086]

[0087] Where E is the effective exposure, L is the scene light intensity, t is the shutter speed, and f is the aperture value;

[0088] In the inference phase, based on the predicted light intensity estimate... With expected exposure E target By working backward, we can obtain the optimal parameter combination {t,f,ISO} that satisfies the current shooting task. ISO is the gain factor involved in the optimization, which can be obtained, for example, by looking up a table / noise cost function.

[0089] To further implement the above technical solution, in step 3, the parameter optimization module based on perceptual error introduces perceptual domain reconstruction constraints to minimize the structural differences between the predicted image and the reference image in the high-level semantic space. By introducing a semantic perceptual feature extraction function, the texture, edge and structural information in the image are explicitly modeled to guide the search for the optimal solution of the exposure parameters.

[0090] The optimization objective for the exposure parameter combination is:

[0091]

[0092] Among them, I pred For an image captured based on the current exposure combination {t,f,ISO}, I gt φ is a reference image acquired under standard imaging conditions; φ(·) is a high-order feature extraction function of a pre-trained feature extraction network used to capture the structural semantics of the image. is the Euclidean distance norm, used to measure image dissimilarity in a high-dimensional semantic space;

[0093] Compared to traditional optimization strategies based on pixel-level errors, this perception-driven objective function can more effectively constrain the performance of predicted images in terms of structure preservation, edge contrast, and semantic consistency. It is particularly suitable for handling complex scenes such as low light or backlight that cause blurred boundaries of dark targets.

[0094] To further implement the above technical solution, the joint loss function is:

[0095]

[0096] Where λ1, λ2, and λ3 are loss weighting factors used to dynamically adjust the network to achieve the optimal balance between physical accuracy and perceived visual quality. For illumination estimation error, For gamma-ray inverse reconstruction error, To predict the structural differences between the image and the reference image in the high-level semantic space.

[0097] To further implement the above technical solution, the illumination estimation error... Used to measure the light intensity estimate output by network regression. With realistic lighting label L i Mean square error between:

[0098]

[0099] The gamma-ray inverse reconstruction error term is used to evaluate the brightness fidelity of the image after inverse gamma mapping.

[0100]

[0101] Among them, I input For the input image luminance channel, I target For the reference brightness image, the gamma coefficient is γ; is a hyperparameter, and is a non-negative hyperparameter used to balance the contribution ratios of illumination regression and gamma reconstruction in the total loss;

[0102] By combining illumination error loss and gamma recovery loss, the network is not only guided to learn the ability to accurately invert the real physical illumination intensity from the image, but also further ensures the reversibility of the gamma mapping on the reverse path and the fidelity of the image brightness structure, thereby improving the robustness and perceptual consistency of the overall system in imaging parameter decoding and exposure control.

[0103] To further implement the above technical solutions, such as Figure 4 The specific content of step S5 includes:

[0104] S51. Edge Deployment and Model Optimization Acceleration: Deploy the trained network model to an embedded edge computing platform (such as NVIDIA Jetson Orin NX) on a drone in ONNX or TensorRT format;

[0105] The platform has FP16 / INT8 mixed precision inference capability, supports tensor-level parallel computing and high-speed caching of image data streams, and enables low-latency inference execution of illumination estimation models.

[0106] S52. Real-time image acquisition and prediction triggering mechanism: Acquire image frames I from the camera at the current timestamp t according to the set frame rate. t The forward prediction process is completed within a total delay of no more than 15ms from image acquisition to processing output, to obtain the estimated light intensity value of the current frame.

[0107] S53. Solving Exposure Combinations under Imaging Physical Constraints: Based on the predicted illumination values... Automatically match the preset exposure lookup table or perform real-time numerical solving to obtain a set of optimal solutions {t} that can be used to adjust imaging parameters. * ,f * ISO *} t To meet the target exposure constraint:

[0108]

[0109] Among them, t * For exposure time, f * This refers to the aperture value and ISO. * For sensor gain parameters;

[0110] Enhance the salience of low-light targets by dynamically adjusting the target exposure based on the task context;

[0111] S54. Imaging parameter writing and hardware-level exposure control: The obtained exposure parameter combination is transmitted to the UAV camera control unit through the interface control module. The exposure setting is preloaded before the next frame image is acquired through the standard camera API or by directly accessing the CMOS image sensor driver interface.

[0112] In this embodiment, step S5 further includes timing optimization and robustness mechanisms, specifically:

[0113] To ensure latency is kept below 15ms, a parallel thread architecture and a double-buffered data mechanism are used for scheduling. In complex environments (such as rapid light changes, dynamic occlusion, and backlighting scenes), a dynamic exposure trend filtering of historical frames and a confidence level discrimination strategy are introduced to improve the system's stability and response robustness under non-ideal conditions. Actual lighting changes are often "abrupt and fluctuating." If the system parameters are drastically adjusted every frame, the image will flicker. Therefore, historical frame sequences are used for filtering to extract the trend of lighting changes, rather than blindly following the instantaneous prediction value. In addition, lighting prediction is sometimes unreliable (such as occlusion, strong reflection, and overexposed areas). A confidence level is calculated for the prediction result of each frame. When the confidence level is low, it is not used directly, but the previous frame is retained or a smooth transition is used to avoid parameter jitter caused by misjudgment of lighting.

[0114] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an imaging optimization method for drone aerial images that integrates deep learning perception mechanisms and physical modeling.

[0115] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements an imaging optimization method for UAV aerial photography that integrates deep learning perception mechanisms and physical modeling.

[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0117] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An unmanned aerial vehicle aerial image imaging optimization method fusing a deep learning perception mechanism and physical modeling, characterized in that, The method comprises the following steps: S1. Collecting original image frames obtained by a UAV during flight; S2. Inputting the image into a light estimation network constructed based on MobileViT, extracting local convolutional perception and multi-scale global semantic features, and outputting scene light intensity estimation values; S3. Constructing a differentiable imaging parameter inference module, based on the quantitative relationship between light intensity and imaging parameters, decoding the optimal exposure parameter combination of the current frame according to the predicted scene light intensity estimation values and the expected exposure amount, and constructing a parameter optimization module based on perception error; S4. Constructing a joint loss function based on light error loss and gamma recovery loss, combining perception domain reconstruction constraints, and jointly training a network model including the light estimation network, the imaging parameter inference module, and the imaging parameter optimization module to optimize the exposure parameter combination; S5. Real-time collection of current frame images during UAV flight, input into the trained network model for online inference and dynamic adjustment of exposure parameter combination, completing parameter exposure prediction, control feedback, and next frame image collection. 2.The UAV aerial image imaging optimization method of claim 1, wherein, The light estimation network constructed based on MobileViT includes a local perception module, a Transformer encoding module, and a multi-scale fusion module; The specific content of step S2 is: S21. Local perception encoding module: utilize deep separable convolution to extract fine-grained brightness, edge and texture features in the image while maintaining computational efficiency, obtaining initial representation feature map F c ; S22. Lightweight Transformer encoding module: the local feature map F c The space is divided into fixed-size non-overlapping patch blocks, and after linear embedding, it is input into a simplified multi-head self-attention module to model the illumination consistency and brightness offset characteristics of long-distance regions in the image, and output global context features F t ; S23. Multi-scale enhanced fusion module: based on the cross-scale semantic fusion strategy, the local features F c are spliced with the global context features F t , and channel remapping and weighting are performed through 1x1 convolution to output the final fusion features. 3.The UAV aerial image imaging optimization method of claim 2, wherein, The initial representation feature map is: F c = Conv 3×3 (I input ) The output feature of the multi-head self-attention module is: F t = MHSA(LN(F c ))+ F c Where MHSA represents a simplified self-attention mechanism, and LN is a layer normalization operation; The final fused feature is output: F out = Concat(F c ,F t ) → Conv 1×1 .

4. The method of claim 1, wherein the method further comprises: In step S3, the quantitative relationship between light intensity and imaging parameters is: Where E is the effective exposure amount, L is the scene light intensity, t is the shutter time, and f is the aperture value; In the inference phase, the estimated value of the predicted light intensity is used to estimate the exposure value with the desired exposure E target , the optimal parameter combination {t, f, ISO} that satisfies the current shooting task is obtained by back calculation, and ISO is the gain factor involved in optimization.

5. The method of claim 1, wherein the method further comprises: In step 3, the parameter optimization module based on perception error introduces a perception domain reconstruction constraint to minimize the structural difference between the predicted image and the reference image in the high-level semantic space. By introducing a semantic perception feature extraction function, the texture, edge, and structural information in the image are explicitly modeled to guide the optimal solution search for the exposure parameters. The optimization target of the exposure parameter combination is: where I pred is the image taken according to the current exposure combination {t, f, ISO}, I gt is the reference image taken under standard imaging conditions; φ(·) is a high-order feature extraction function of a pre-trained feature extraction network, which is used to capture the structural semantics of the image; is the Euclidean distance norm, which is used to measure the difference between images in the high-dimensional semantic space.

6. The method of claim 1, wherein the method further comprises: The joint loss function is: wherein λ1, λ2, λ3 are loss weighting factors used to dynamically adjust the optimal balance between physical accuracy and perceptual visual quality of the network, is the illumination estimation error, is the gamma inverse reconstruction error, is the structural difference between the predicted image and the reference image in the high-level semantic space.

7. The method of claim 6, wherein the method further comprises: Illumination estimation error For metric network regression output illumination intensity estimates The mean squared error between the true illumination label L i and the predicted illumination label L Gamma inverse reconstruction error term, for evaluating the network on the image luminance fidelity after inverse gamma mapping: where I input is the input image luminance channel, I target is the reference luminance image, and the gamma coefficient is γ; is a hyper-parameter, is a non-negative hyper-parameter balancing the contribution of the illumination regression and the gamma reconstruction in the total loss. 8.The UAV aerial image imaging optimization method of claim 1, wherein, The specific content of step S5 includes: S51. Deploying the trained network model to the UAV end embedded edge computing platform; S52. Obtain an image frame I at current timestamp t from the camera at a set frame rate t and complete the forward prediction process to obtain the current frame illumination intensity estimate under the constraint that the image acquisition to processing output total latency does not exceed 15ms S53. According to the predicted illumination value automatically match a pre-set exposure lookup table or perform real-time numerical solving to obtain a set of optimal solutions for regulating imaging parameters * ,f * ,ISO *} t , meet the target exposure constraint: where t * is the exposure time, f * is the aperture value, ISO * is the sensor gain parameter; S54. The obtained exposure parameter combination is transmitted to the UAV camera control unit through the interface control module, and the exposure is preloaded before the next frame image collection through the standard camera API or direct access to the CMOS image sensor driver interface.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of claim 1-8.

10. A processing terminal comprising a memory and a processor, the memory having stored therein a computer program executable on the processor, characterized in that, The computer program is executed by the processor to implement the method of claim 1-8.

Images