Automatic exposure parameter linkage control method for photographic equipment

Abstract

The invention discloses an automatic exposure parameter linkage control method for photographic equipment, and aims to solve the problem of image quality degradation caused by shutter, aperture and sensitivity parameter split adjustment in existing automatic exposure. According to the method, a comprehensive cost function fusing an exposure physical model, a multi-dimensional image quality loss function and scene semantic priori is constructed, a real-time gradient descent algorithm with projection is adopted under the physical constraint of equipment, and the shutter speed, the aperture relative aperture and the sensitivity gain are synchronously optimized; therefore, the optimal parameter combination which meets the target exposure and considers the requirements of noise, motion blur and depth of field can be output in milliseconds. Through adoption of the technical scheme, the imaging quality and the parameter adjustment smoothness in complex illumination and dynamic scenes can be remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of photography technology, and more specifically, relates to an automatic exposure parameter linkage control method for photographic equipment. Background Technology

[0002] In today's era of rapid development in digital imaging and intelligent photography technologies, automatic exposure control, as one of the core functions of photographic equipment, directly determines the stability and visual expressiveness of image quality. Automatic exposure systems dynamically adjust three main parameters—shutter speed, aperture, and ISO—to maintain image brightness within an ideal range under complex and changing lighting conditions, thus ensuring both usability and artistic quality of the image. Traditional automatic exposure mechanisms often employ metering zone-weighted averaging or spot metering strategies, combined with preset exposure target values ​​(such as 18% grayscale reflectance), and use feedback control loops to adjust exposure parameters independently or sequentially. This approach performs reasonably well in static or slowly changing lighting scenarios. Its design logic is based on the simplified assumption of "single parameter dominating adjustment, other parameters passively adapting," meaning that at any given moment, only the shutter speed is adjusted first to respond to brightness changes; if mechanical or motion blur limitations are exceeded, the aperture is then adjusted, and only lastly is ISO increased to compensate for underexposure. While this strategy was feasible in terms of engineering implementation under the limited computing power of early hardware, its inherent coupling relationship was artificially severed, making it difficult to achieve a multi-dimensional balance between image clarity, noise level and motion capture capability in complex shooting conditions such as high dynamic range, fast motion or low light.

[0003] Specifically, existing automatic exposure algorithms generally rely on the following form of exposure equation as the control basis:

[0004]

[0005] in This indicates the effective exposure received by the sensor. For scene brightness, For shutter speed, It is the reciprocal of the aperture F-number (i.e., the relative aperture). These are equipment-related calibration constants. Under this model, the control system typically sets the target exposure. The parameter combination is determined by iteratively solving the above equation. However, this equation only describes the static exposure balance relationship and does not embed the secondary influence function of each parameter on image quality. For example, shortening the shutter speed t can suppress motion blur, but it will reduce the signal-to-noise ratio; increasing the aperture A increases the amount of light entering the camera, but it will reduce the depth of field and may introduce optical aberrations; while increasing the ISO directly amplifies sensor readout noise and fixed-mode noise. Therefore, if only considering... If the sole optimization objective is ignored and the nonlinear effect of parameter selection on image perceived quality is ignored, it is very easy to lead to a suboptimal result that is "technically correct in exposure but visually poor".

[0006] Furthermore, with the deep integration of computational photography and artificial intelligence technologies, modern photographic equipment places higher demands on automatic exposure systems: they not only need to make parameter decisions within milliseconds, but also achieve Pareto optimality under multiple objective constraints. However, existing linkage control strategies mostly employ staged threshold judgment mechanisms, for example:

[0007]

[0008] While such rules are interpretable, their discretized and discontinuous decision boundaries struggle to adapt to continuously changing scene semantics. Furthermore, potential conflicts exist between rules, requiring extensive empirical weight tuning and lacking a unified mathematical optimization framework. Fundamentally, existing methods fail to construct a joint optimization objective that organically integrates exposure physics models, image quality evaluation functions, and prior knowledge of user intent. For example, a comprehensive cost function could be defined:

[0009]

[0010] in , , The quantization losses for noise, motion blur, and depth-of-field defocusing are respectively characterized. Adaptive weights are used for different scenarios. However, current mainstream devices are limited by real-time requirements and algorithm architecture inertia, and rarely adopt this kind of end-to-end optimization paradigm. Instead, they follow decoupled and heuristic parameter adjustment processes, which leads to frequent problems such as exposure fluctuations, abrupt parameter switching, or loss of control of key image quality indicators when facing scenarios such as high dynamic range video shooting, low-light portrait or high-speed sports photography.

[0011] Therefore, how to construct an automatic exposure linkage control method that can deeply integrate exposure physical constraints and multi-dimensional image quality goals and support continuous collaborative optimization between parameters has become a key challenge and a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0012] To achieve the above-mentioned objectives, this invention provides an automatic exposure parameter linkage control method for photographic equipment. The method constructs a joint optimization objective that integrates an exposure physics model, a multidimensional image quality loss function, and scene semantic priors. It also employs a continuous parameter collaborative adjustment mechanism based on real-time gradient descent to simultaneously determine the optimal combination of shutter speed, aperture value, and ISO within a millisecond-level response time. This minimizes the overall image quality degradation caused by noise, motion blur, and depth-of-field defocusing while satisfying the target exposure constraints.

[0013] The method of this invention specifically includes the following steps: First, acquiring raw image data of the current frame through an image sensor, and calculating scene brightness distribution characteristics, estimated subject motion speed, and depth-of-field requirement level based on the data; Second, defining shutter speed t, relative aperture A (i.e., the reciprocal of the F-value), and sensitivity gain G according to a preset set of device physical constraints. ISO The feasible region is determined; furthermore, a comprehensive cost function is constructed that includes exposure bias term, noise loss term, motion blur loss term, and depth-of-field defocusing loss term. Finally, within the feasible region, J is optimized in real time using a constrained gradient descent algorithm, outputting an optimal parameter triplet that satisfies multi-objective balance. It also drives the actuator to complete the parameter settings.

[0014] Furthermore, the scene brightness distribution characteristics are obtained in the following way: the original image of the current frame is divided into N×M metering units, and the brightness value of each unit is... Calculated by the following formula:

[0015]

[0016] in Indicates the first Line number The set of pixels in the metering area. , , These are the original pixel values ​​for the red, green, and blue channels, respectively. , , This refers to the normalized weighting coefficients of the CIE 1931 standard observer color matching function under the device's spectral response. Their values ​​are determined by factory calibration, with a typical value being [value missing]. =0.299, =0.587, =0.114. Global scene brightness. Defined as the weighted median of the brightness of all metering units:

[0017]

[0018] Among them, weight Generated by a face detection module or deep learning attention map, if a face region is detected, the corresponding unit weight is increased to 1.0, and the weights of other regions are distributed according to Gaussian decay, with the center located in the center of the image and the standard deviation being 1 / 6 of the image width.

[0019] Furthermore, the estimated velocity of the main body. It is obtained through optical flow calculation. Specifically, pyramid Lucas-Kanade optical flow calculations are performed on the current frame and the previous frame to obtain the displacement vector of each pixel. Then the local velocity modulus is:

[0020]

[0021] Main body speed Defined as a high gradient region (i.e., edge intensity greater than a threshold). (pixels) 90th percentile:

[0022]

[0023] in The gradient magnitude of the image is calculated using the Sobel operator.

[0024] Furthermore, the depth-of-field requirement level Output from the scene semantic analysis module. This module performs semantic segmentation on the current frame based on a lightweight convolutional neural network (CNN), identifying the foreground subject and background region. If the foreground subject occupies a certain percentage of the image area... Furthermore, the background region exhibits significant texture complexity (i.e., the local binary pattern LBP variance of the background region is greater than the threshold). If the subject is ____, it is determined to be "Portrait Mode", and the depth-of-field requirement level is set to high. If objects at multiple distance levels are detected (judged using binocular parallax or ToF depth maps, the standard deviation of depth variation is considered); If the field depth requirement is set to low, then it is determined to be "landscape mode". All other cases are set to medium, and recorded as follows: .

[0025] Furthermore, the set of physical constraints for the device is defined as follows:

[0026] Shutter speed satisfy ,in , The current video frame rate, ;

[0027] Aperture relative aperture satisfy ,in , , and This is the physical aperture limit of the lens;

[0028] Sensitivity gain satisfy Among them, the minimum gain value (Corresponding to ISO 100), Maximum Gain Value (Corresponding to ISO 6400), and For discrete adjustable values, the step size is... This corresponds to a 1 / 3 EV increase. Integer index.

[0029] Furthermore, the comprehensive cost function Defined as:

[0030]

[0031] in , , , For scene-adaptive weight coefficients, satisfying Its value is dynamically adjusted according to the semantics of the scene; For exposure deviation, For noise loss, For motion fuzzing loss term, This is the depth-of-field defocusing loss item.

[0032] Furthermore, the exposure deviation term Defined as:

[0033]

[0034] in This is the equipment calibration constant, measured at the factory under a standard light source (D65, 1000 lux), with a typical value of [value missing]. =12.5; The target exposure value is dynamically set based on the metering mode: if matrix metering is used, then... If face-priority metering is used, then ,in The proportion of the image area occupied by the face. The baseline exposure, corresponding to 18% grayscale reflectance, has a value of =0.18Q, where Q is the digital value corresponding to the full well capacity of the sensor, in DN (Digital Number), with a typical value of Q=16383 (14-bit ADC).

[0035] Furthermore, the noise loss term Modeling as photosensitivity gain Nonlinear functions:

[0036]

[0037] in To read the noise standard deviation, the unit is electrons. Typical value =2.1 ; For the variance of shot noise, For quantum efficiency (typical value 0.6). The average number of photons is given by Calculations show that S is the effective photosensitive area of ​​a single pixel (typical value). ), The photon energy (3.61 × 10⁻¹⁹ J at a wavelength of 550 nm) is given. This is the fixed-mode noise figure, typically valued at... =0.001. The model was verified by sensor noise calibration experiments, and the goodness of fit R2>0.98.

[0038] Furthermore, the motion blur loss term Defined as:

[0039]

[0040] in This is a motion blur tolerance scale, measured in pixels. Its value is determined by the image resolution and display size, with a typical value of [value missing]. =1.5 pixels. This function is in When it is approximately 0, When the value approaches 1, it aligns with the human eye's perception of motion blur.

[0041] Furthermore, the depth-of-field defocusing loss term Defined as:

[0042]

[0043] in For reference relative aperture, its value is determined by the focusing distance. With focal length Decide:

[0044]

[0045] Acceptable circle diameter, in millimeters, typically [value missing]. =0.03 mm; This is the steepness coefficient, with a typical value of [value missing]. =4. When When =0 (scenery mode), =0, meaning no penalty for large depth of field; when When =1 (Portrait Mode), Follow Increase the aperture while decreasing it, encouraging the use of a large aperture to blur the background.

[0046] Furthermore, the scene adaptive weight coefficients , , , Set according to the following rules:

[0047] like (Low illumination), then ;

[0048] like Pixels / frame (high-speed motion), then ;

[0049] like =1 (Portrait Mode), then ;

[0050] In other cases (normal scenarios), then .

[0051] Furthermore, the constrained gradient descent algorithm is implemented using the projected gradient descent method. The specific steps are as follows:

[0052] 1. Initialize parameters: , where the superscript (k) indicates the k-th iteration;

[0053] 2. Regarding the number of iterations a. Calculate the gradient vector ;

[0054] b. Perform an unconstrained update:

[0055]

[0056] Learning rate =0.01 seconds, =0.005, =0.1;

[0057] c. Project to the feasible region:

[0058]

[0059] in For discrete ISO gain set, Map continuous values ​​to the nearest discrete gear;

[0060] Output final parameters

[0061] Furthermore, the partial derivatives are calculated as follows:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] All other partial derivatives are zero. The overall gradient is the weighted sum of the components:

[0068]

[0069]

[0070]

[0071] Furthermore, to improve the robustness of the algorithm, a parameter smoothing constraint is introduced. A smoothing cost term is defined. :

[0072]

[0073] in and will Included in the total cost function:

[0074]

[0075] Accordingly, a smoothing term is added to the gradient update:

[0076]

[0077]

[0078]

[0079] Furthermore, the method is executed in hardware by the exposure control module in the image signal processor of the photographic device. This exposure control module includes the following functional modules: a metering feature extraction unit, a motion estimation unit, a semantic analysis unit, a constraint management unit, a cost function calculation unit, and an optimization solution unit. The modules communicate via an on-chip interconnect bus, and the data format uses fixed-point numbers to reduce power consumption. Brightness values ​​are represented as 16-bit unsigned integers, motion speeds as 12-bit fixed-point decimals, and parameter variables as 32-bit floating-point numbers.

[0080] Furthermore, the photometric feature extraction unit receives the original Bayer format image, generates an RGB image after de-mosaicing, and then calculates the partitioned brightness and global brightness according to the aforementioned formula. The motion estimation unit uses a hardware-accelerated optical flow engine, supporting a maximum displacement of 16 pixels with a computation latency of less than 2 milliseconds. The semantic analysis unit deploys a quantized MobileNetV2 backbone network with an input resolution of 224×224 and outputs three types of semantic labels (portrait, landscape, and ordinary), with an inference latency of less than 5 milliseconds. The constraint management unit stores the physical parameter tables of the lens and sensor and queries the current feasible domain boundary in real time. The cost function calculation unit performs parallel computation. The gradient is calculated using a lookup table (LUT) to accelerate exponential and logarithmic operations. The optimization unit executes the aforementioned projective gradient descent algorithm, iterating 5 times, with a single iteration delay of less than 1 millisecond.

[0081] Furthermore, to address extreme lighting abrupt changes, a feedforward compensation mechanism is introduced. When a global brightness change rate between adjacent frames is detected... At this time, activate fast response mode: set initial parameters Set as , remain unchanged. This keeps the exposure level constant, thus approaching exposure equilibrium in the first iteration and avoiding response delays caused by multiple iterations.

[0082] Furthermore, the method supports user intent intervention. If the user manually locks a parameter (e.g., fixes the shutter speed), that parameter is treated as a constant during the optimization process, and only the other two degrees of freedom are optimized. For example, if If locked, the optimization variable is only... and The feasible region is reduced to a two-dimensional subspace, and gradient calculations retain only the relevant components.

[0083] Furthermore, the method introduces a temporal consistency constraint in video shooting mode. A temporal smoothing factor is defined. =0.8, using the optimal parameters of the previous frame as the soft constraint center for optimization in the current frame:

[0084]

[0085] in This is the theoretical shutter speed value, solved solely from the exposure equation. This center value is used for initialization. and replace it in the smoothing cost term. This helps suppress parameter jitter in video sequences.

[0086] Furthermore, the method has been verified by actual tests. In low light (1 lux), high-speed motion (100 pixels / frame) and portrait scenes, compared with the traditional staged thresholding method, the image signal-to-noise ratio is improved by 3.2 dB, the motion blur area is reduced by 41%, the background blur satisfaction (subjective rating by MOS) is improved by 1.8 points (out of 5), and the parameter switching transition is smooth with no visible flickering or jumps.

[0087] The beneficial effects of this invention are:

[0088] This invention constructs a joint cost function that is physically interpretable, mathematically optimizable, and engineering feasible, and employs an efficient constraint optimization algorithm to achieve continuous, coordinated, and real-time linkage control of three parameters: shutter speed, aperture value, and ISO. This fundamentally solves the problems of parameter fragmentation, rule conflict, and suboptimal decision-making in existing technologies, and significantly improves automatic exposure performance and user experience in complex scenarios. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the system architecture and data flow of the automatic exposure parameter linkage control method for photographic equipment described in this invention.

[0090] The attached figures are labeled as follows: 1. Image sensor; 2. Metering feature extraction unit; 3. Motion estimation unit; 4. Semantic analysis unit; 5. Constraint management unit; 6. Cost function calculation unit; 7. Optimization solution unit; 8. Actuator; 9. Image signal processor; 10. Exposure control module. Detailed Implementation

[0091] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0092] This invention provides an automatic exposure parameter linkage control method for photographic equipment. Its core lies in constructing a joint optimization objective that integrates an exposure physics model, a multi-dimensional image quality loss function, and scene semantic priors. It employs a continuous parameter collaborative adjustment mechanism based on real-time gradient descent to simultaneously determine the optimal combination of shutter speed, aperture value, and ISO within a millisecond-level response time. This minimizes the overall image quality degradation caused by noise, motion blur, and depth-of-field defocusing while satisfying the target exposure constraint. The following section will elaborate on the technical solution of this invention, including system architecture, data processing flow, mathematical modeling details, hardware implementation, and typical application scenarios.

[0093] First, the overall execution environment of the method described in this invention is the exposure control module 10 within the image signal processor 9 of the photographic equipment. This exposure control module 10 is integrated within the image signal processor 9, possesses an independent data path and computing unit, and can process the raw data stream from the image sensor in parallel, completing the entire process from feature extraction to parameter output within a single frame period. (See attached...) Figure 1 As shown, the coprocessor includes a photometric feature extraction unit 2, a motion estimation unit 3, a semantic analysis unit 4, a constraint management unit 5, a cost function calculation unit 6, and an optimization solution unit 7. The modules communicate through an on-chip interconnect bus, and the data format uses fixed-point numbers to reduce power consumption. The brightness value uses a 16-bit unsigned integer, the motion speed uses a 12-bit fixed-point decimal, and the parameter variables use 32-bit floating-point numbers.

[0094] In one specific embodiment, image sensor 1 outputs raw image data in Bayer format with a resolution of 4096×2160 and a frame rate of 30 fps. This raw data first enters metering feature extraction unit 2. This unit performs zone metering on the current frame's raw image, dividing it into N×M=8×8 metering units, each covering a 512×270 pixel area of ​​the image. For the first... Line number Metering area Its brightness value Calculate using the following formula:

[0095]

[0096] in , , These are the red, green, and blue channel pixel values ​​obtained after de-mosaicing. =0.299, =0.587, =0.114 is the normalized weighting coefficient of the CIE 1931 standard observer color matching function under the device's spectral response. This set of values ​​was obtained during the device's factory calibration phase using an integrating sphere light source and a standard color chart, and is stored in the coprocessor's read-only memory. The global scene brightness L is defined as the weighted median of the brightness of all metering units.

[0097]

[0098] Weight Generated by the face detection module: If a face is detected in the corresponding region (using a cascaded classifier based on Haar features, with a threshold set to 0.85 confidence), then... =1.0; If no face is detected, the weight decays according to a Gaussian distribution, centered at the center of the image coordinates (2048, 1080), with a standard deviation of 1 / 6 of the image width, i.e., σ = 4096 / 6 ≈ 683 pixels. The specific expression is as follows:

[0099]

[0100] in( , ( ) is the photometer unit The geometric center coordinates.

[0101] Meanwhile, motion estimation unit 3 receives grayscale images (converted from RGB images) of the current frame and the previous frame, and performs pyramid Lucas-Kanade optical flow calculations. This unit employs a three-level pyramid structure, with the bottom layer at the original resolution and the top layer at 256×135. The displacement vector for each pixel... The local motion velocity magnitude is obtained by solving a least-squares problem within a local window, with a window size of 15×15 pixels.

[0102]

[0103] Main body speed Then take the high gradient region The 90th percentile. High gradient regions are detected by the Sobel operator: for grayscale images. Calculate gradient magnitude ,like (Based on an 8-bit image), the pixel is then included in the statistical set. Ultimately:

[0104]

[0105] This unit uses a hardware-accelerated optical flow engine, supports a maximum displacement of 16 pixels, and has a computation latency of less than 2 milliseconds.

[0106] Semantic analysis unit 4 deploys a quantized MobileNetV2 lightweight convolutional neural network. The input is an RGB image downsampled to 224×224 pixels from the current frame, and the output is the probability distribution of three semantic labels: portrait, landscape, and ordinary. After pre-training on ImageNet, the network is fine-tuned using a self-built photographic scene dataset containing 100,000 labeled images, with portrait samples accounting for 40%, landscape for 30%, and ordinary for 30%. During inference, if the probability of the portrait class... Then, the portrait mode judgment logic is entered: calculate the area ratio of the foreground subject. (Statistics on the proportion of non-background pixels using semantic segmentation masks) If >0.3, and the variance of the local binary pattern (LBP) in the background region is greater than 0.3. If so, it is determined to be "Portrait Mode", with a depth-of-field requirement level. If the probability of scenery is... If the device is equipped with a binocular camera or a ToF depth sensor, it can read the depth map and calculate the standard deviation of the depth values ​​across the entire map. ;like If the distance is greater than 0.5 meters, it is determined to be in "landscape mode". =0. All other cases are set to medium depth of field. =0.5. The inference latency of this unit is less than 5 milliseconds.

[0107] Constraint Management Unit 5 maintains a device physical parameter table, storing key specifications of the lens and sensor. For a device equipped with an f=50 mm fixed-focus lens and a maximum aperture... Minimum aperture For a camera with an aperture ratio of 16, the feasible region of its aperture relative aperture is... =1 / 16=0.0625 to =1 / 1.8≈0.5556. The lower limit of shutter speed is determined by the video frame rate: if the current frame rate... =30 Hz, then =1 / (2×30)≈0.0167 seconds (i.e., 1 / 60 of a second), upper limit =30 seconds. Sensitivity gain discrete sets From the initial value =1 (ISO 100), step size ≈1.26 and upper limit =64 (ISO 6400) is a common definition, containing a total of There are several levels, with specific values ​​{1, 1.26, 1.59, 2.0, ..., 64}. Before each optimization, this unit queries the current feasible region boundary in real time and passes it to subsequent modules.

[0108] Cost function calculation unit 6 performs parallel calculation of the comprehensive cost function. and its gradient. The function is defined as:

[0109]

[0110] in This is the smoothing cost term. The components are calculated as follows:

[0111] Exposure deviation item Depends on target exposure Baseline exposure =0.18Q, where Q=16383 (full scale of a 14-bit ADC), therefore =2949 DN. If face-priority metering is used and the face region accounts for... =0.25, then =2949×(1+0.2×0.25)=3096 DN. Equipment calibration constant. =12.5, obtained at the factory under a D65 light source and 1000 lux illuminance. Therefore:

[0112]

[0113] Noise loss item Based on sensor noise model:

[0114]

[0115] in =2.1 , =0.001. Average photon number , , quantum effect Therefore For example, when =100L = 100 cd / m², When, it can be calculated that ,and then .

[0116] Motion fuzzing loss term Using an exponential decay model:

[0117]

[0118] in =1.5 pixels is the motion blur tolerance scale.

[0119] Depth of field defocus loss item Only Activated when >0. Reference relative aperture. Where f = 50 mm, =0.03 mm, focusing distance =2 m, then 0.0427. The kurtosis coefficient k=4, therefore:

[0120]

[0121] Smoothing cost term Introducing time consistency:

[0122]

[0123] The scene-adaptive weighting coefficients are dynamically set based on the current scene. For example, in a low-light scene. The weights are α=0.3, β=0.4, γ=0.2, δ=0.1; in high-speed motion scenarios ( =60 pixels / frame > 50), weight is In portrait mode ( =1), weight is In typical scenarios, it is .

[0124] The optimization unit 7 executes the projected gradient descent algorithm, with a maximum number of iterations. During initialization, if the brightness change rate of adjacent frames... ,but ;like Then feedforward compensation is activated: All other parameters remain unchanged. In each iteration, the gradient is calculated as follows:

[0125]

[0126] in In video mode , =0.8, Similarly, calculate... and Unconstrained updates use the learning rate. =0.01 seconds, =0.005, =0.1, and then projected to the feasible region: shutter speed and aperture are limited by the clip function, and ISO gain is mapped to the nearest discrete stop by a lookup table.

[0127] After the optimization and solution unit 7 outputs the optimal parameters, the physical settings of the aperture blades or shutter curtain are completed by the actuator 8 (such as a stepper motor drive circuit). The entire calculation process is uniformly scheduled by the exposure control module 10 in the image signal processor 9.

[0128] In a preferred embodiment of the present invention, when the user manually locks the shutter speed (e.g., sets...), =1 / 500 seconds), then the optimization variable is reduced to and In gradient calculation If ignored, the feasible region becomes a two-dimensional rectangular region, and the algorithm can still converge within 3 iterations.

[0129] To verify the effectiveness of the present invention, the following embodiments and comparative examples were designed:

[0130] Example: In low light (1 lux), subject movement speed 100 pixels / frame, portrait mode ( Under condition =1), the method of the present invention is run on a prototype machine equipped with the aforementioned coprocessor. Global brightness is measured. , =102 pixels / frame Background LBP variance = 180. Weights are set to... (Due to the simultaneous requirements of low light intensity and high-speed motion, parameters are superimposed according to priority). Optimized output parameters are as follows: .

[0131] Comparative example: Using the traditional staged threshold method: First, set the base exposure combination based on the metering results (e.g., t=1 / 60, A=0.2, G=8), then check whether the motion blur exceeds the limit (if...). If the resolution is >2 pixels, increase the shutter speed to 1 / 500 and simultaneously increase the ISO to 32. Finally, if it's portrait mode and the aperture isn't at its maximum, open the aperture to F2.0 and decrease the ISO to 20 accordingly. The final parameters are: .

[0132] The images captured using the two sets of parameters were objectively and subjectively evaluated, and the results are shown in the table below:

[0133]

[0134] Experiments show that this invention achieves superior overall performance in multi-objective conflict scenarios. The fundamental reason is that the joint optimization framework avoids the local optimum trap in staged decision-making, such as the excessive increase in ISO to suppress ambiguity in comparative studies, which leads to increased noise; while this invention finds a globally better solution by weighing various losses using gradient information.

[0135] Furthermore, the present invention performs particularly well in video shooting mode. Taking a video recording at 30 fps that quickly moves from indoors (100 lux) to outdoors (10000 lux) as an example, the traditional method, due to independent calculation for each frame, causes the ISO to jump from 800 to 100 within a few frames, resulting in obvious brightness flicker; while the present invention, by introducing time consistency constraints, achieves smooth parameter transitions, continuous brightness changes, and no discomfort for subjective viewing.

[0136] In summary, this invention, through rigorous engineering implementation, transforms a complex multi-objective optimization problem into an algorithmic process that can be efficiently executed on embedded platforms. From feature extraction, semantic understanding, and physical modeling to real-time optimization, each step has been carefully designed and parameter calibrated, ensuring the feasibility and robustness of the technical solution. Those skilled in the art can reproduce this invention on different hardware platforms based on this embodiment, requiring only adjustments to relevant parameters (such as...) according to the specific sensor, lens, and computing resources. By doing so, the same technical effect can be achieved.

[0137] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for automatic exposure parameter linkage control of photographic equipment, characterized in that, The method is executed by the exposure control module (10) of the image signal processor (9) and includes the following steps: S1: Acquire the raw image data of the current frame through the image sensor (1); S2: Based on the original image data, calculate the scene brightness distribution characteristics and the estimated value of the subject's motion speed. and depth of field requirement level ; S3: Determine the shutter speed t, relative aperture A, and ISO gain G based on the preset set of device physical constraints. ISO The feasible domain; S4: Construct the comprehensive cost function J(t, A, G) ISO ), which includes the exposure deviation term J E Noise loss item J N Motion fuzzing loss term J B and depth of field defocus loss item J D ; S5: Within the feasible region, the comprehensive cost function is optimized in real time using a constrained gradient descent algorithm, and the optimal parameter triplet is output. And drive the actuator (8) to complete the parameter setting; Wherein, the relative aperture A is defined as the reciprocal of the F value, and the weights of each component of the comprehensive cost function are dynamically adjusted according to the scene semantics.

2. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The acquisition of the scene brightness distribution features includes: Divide the current frame's original image into N×M metering units. The brightness value of each unit Calculate using the following formula: in , , These are the pixel values ​​for the red, green, and blue channels. , , These are normalized color weight coefficients, with typical values ​​of 0.299, 0.587, and 0.114, respectively. The global scene brightness L is defined as the weighted median of the brightness of all metering units, with the weights... The weight is determined by the face detection results or the Gaussian decay distribution in the center of the image. If a face region is detected, the weight of the corresponding unit is set to 1.0, and the weight of the remaining regions decays according to a Gaussian function with a standard deviation of 1 / 6 of the image width.

3. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The estimated speed of the main body Calculated using the optical flow method: Pyramid Lucas-Kanade optical flow calculations are performed on the current frame and the previous frame to obtain the displacement vector of each pixel. ; Calculate the local motion velocity modulus Main body speed Defined as image gradient magnitude Greater than the threshold In the high gradient region of = 30 The 90th percentile, of which The Sobel operator is used for computation.

4. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The depth of field requirement level Output from semantic analysis unit (4) based on a lightweight convolutional neural network: If the area ratio of the foreground subject If the variance of the local binary pattern (LBP) in the background region is greater than 150 and the value is greater than 0.3, then it is determined to be a portrait pattern. = 1; If the standard deviation of depth variation is determined by binocular parallax or ToF depth maps... If the value is in meters, it is determined to be in landscape mode. = 0; In all other cases, a medium depth of field is required. = 0.

5.

5. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The set of physical constraints for the equipment is defined as follows: Shutter speed satisfy ,in , The current video frame rate, = 30 seconds; The relative aperture A satisfies ,in , , and This is the physical aperture limit of the lens; Sensitivity gain These are discrete adjustable values, taken from the set Among them, the minimum gain value = 1, maximum gain value = 64, step size , Integer index.

6. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The comprehensive cost function It also includes a smoothing cost term. Defined as: in , , , The parameter value is from the previous frame; The total cost function is and introduce it into gradient calculation The partial derivative terms.

7. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, The constrained gradient descent algorithm employs the projected gradient method and includes the following steps: Initialization parameters , , ; Iteration In each iteration: a. Calculate the comprehensive cost function for t, A, G ISO The gradient; b. Perform an unconstrained update: Learning rate = 0.01 seconds, = 0.005, = 0.1; c. Project to the feasible region: Output the final parameter triplet .

8. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, Introduce time consistency constraints in video recording mode: Define time smoothing factor =0.8, calculate the soft constraint center ,in This is the theoretical shutter value obtained solely from the exposure equation; As the initial value for the current frame optimization and in the smoothing cost term China alternative This is to suppress parameter jitter.

9. The automatic exposure parameter linkage control method for photographic equipment according to claim 1, characterized in that, Supports user intent intervention: If the user manually locks any one of the parameters, such as shutter speed, aperture value, or ISO, then that parameter is treated as a constant during the optimization process. Only the remaining two degrees of freedom are jointly optimized, and the feasible region is reduced to a two-dimensional subspace. The gradient calculation only retains the partial derivative components of the relevant variables.

Images