Method for dynamically adjusting picture according to watching condition based on self-adaptive naked-eye 3D

Abstract

The invention relates to the technical field of three-dimensional display, and discloses a method for dynamically adjusting a picture according to a viewing condition based on self-adaptive naked-eye 3D, and the method comprises the following steps: obtaining a current viewer state of a viewer; on the basis of the state and the rendering confidence map of the previous period, predicting and generating a prediction viewpoint of the next moment by fusing scene significance analysis and confidence feedback weighted correction; based on the predicted viewpoint, left and right eye images are generated through neural light field rendering, and a current period rendering confidence map representing the rendering quality is synchronously generated; and finally, based on the generated image and the current period rendering confidence map, adaptive optical compensation and display are carried out through an active microlens array. According to the method, a closed-loop system from intention prediction to rendering quality feedback to optical compensation is constructed, so that active and forward-looking response to viewing condition changes is realized, and the smoothness, the stability and the perception image quality of naked eye 3D display are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional display technology, specifically to a method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D. Background Technology

[0002] Glasses-free 3D display technology, by eliminating the need for specialized glasses and providing users with a natural and immersive 3D visual experience, is increasingly becoming a research hotspot and development direction in the display technology field. To overcome the bottleneck of limited viewing position in traditional glasses-free 3D display technology and give viewers greater freedom of movement, adaptive glasses-free 3D display technology has emerged. This technology typically integrates sensors such as cameras to track the viewer's eye position in real time and dynamically adjusts the rendered content and optical system based on this positional information, thereby accurately projecting the corresponding left and right eye images onto the viewer's moving eyes.

[0003] However, existing adaptive glasses-free 3D display technologies still face a series of interconnected technical challenges in practical applications, which collectively limit the smoothness and comfort of the viewing experience. A core challenge stems from its inherently responsive processing flow. When a viewer moves, the system must sequentially complete a series of steps, including position capture, data processing, image rendering, and optical adjustment. This entire processing chain inevitably introduces significant "motion-photon" latency. This latency causes the displayed image to always lag behind the viewer's actual position. When the viewer's head moves even slightly quickly, they will immediately perceive a brief collapse, distortion, or ghosting (crosstalk) in the 3D image. This not only severely damages the immersive experience of the 3D scene but also easily induces visual fatigue.

[0004] To reduce this latency, existing technologies are often forced to compromise in the image rendering stage, generating new views within a very short time budget. This extreme demand for speed often comes at the cost of rendering quality, especially when dealing with complex 3D scenes. The rendered images may exhibit visual artifacts such as obvious geometric distortion, inconsistent lighting, or loss of material details. These rendering-level flaws, combined with inherent imperfections in the optical system such as crosstalk, further exacerbate the decline in the final viewing experience, making it difficult for existing technologies to continuously provide stable, high-quality 3D visual effects while ensuring the viewer's freedom of movement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for dynamically adjusting the image based on viewing conditions in adaptive glasses-free 3D. This aims to solve the problems of discontinuous 3D viewing experience and degraded image quality when the viewer moves, caused by system latency and rendering quality bottlenecks in existing adaptive glasses-free 3D technologies.

[0006] To achieve the above objectives, this invention provides the following technical solution: a method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D. This method organically combines viewer intent prediction, high-fidelity scene rendering, and physical compensation of the display terminal through a collaborative closed-loop system. The method includes:

[0007] First, in step a, the system acquires the viewer's current viewer state. In an optional implementation, this state information is multimodal and may include the precise coordinates of the viewer's eyes in three-dimensional physical space acquired through a three-dimensional sensing device such as a depth camera, and the viewer's current visual focus direction vector acquired through an eye-tracking device. This information together constitutes a complete description of the viewer's current behavior.

[0008] Next, in step b, the system predicts the viewpoint for the next moment based on the acquired current viewer state and a key feedback information—the rendering confidence map of the previous cycle. This step is the core of this invention, transforming passive tracking into active prediction. In a specific embodiment, this step is further refined as follows: First, the system analyzes the 3D scene content to be displayed and generates a scene saliency map representing the attractiveness of the content. However, directly using this saliency map may guide the user's visual focus to areas that the system struggles to render with high quality. Therefore, this invention introduces the quantitative evaluation result of the rendering quality of the previous cycle, i.e., the rendering confidence map of the previous cycle, and performs a weighted correction on the scene saliency map to generate a corrected saliency map. This correction process can follow the following relationship:

[0009] S mad (p)=S map (p)·W(C map_prev (p))

[0010] Among them, S mad (p) is the corrected saliency map, representing the visual guidance direction that comprehensively considers scene attractiveness and rendering feasibility; S mad (p) is the original scene saliency map; The system uses the rendering confidence map fed back from previous rendering cycles; W(·) is a weighting function used to reduce the guiding weights of potentially low-quality rendering regions; p is the pixel coordinate of the image. Finally, the system inputs the current viewer state and this corrected saliency map into a temporal prediction model, thereby outputting a more forward-looking and smoother predicted viewpoint.

[0011] Subsequently, in step c, the system renders left-eye and right-eye images for viewing by the left and right eyes, based on the predicted viewpoint generated in the previous step. In a preferred embodiment, this rendering process employs a neural light field rendering method. This method utilizes a pre-trained continuous light field function that characterizes the complete 3D scene, enabling the direct synthesis of geometrically consistent, high-quality images free of hole artifacts from any viewpoint. A key innovation of this step is that, while generating the image, a current-cycle rendering confidence map characterizing the quality of this rendering is generated in parallel. This map quantifies the rendering engine's "confidence" in the results for each pixel region in its generated image. In a specific implementation, the generation of this confidence map is closely related to the volume rendering mechanism in the rendering process, and its value can be determined by calculating the cumulative opacity of the rendering rays in the light field function, specifically as follows:

[0012] C map_curr (p)=1-T(t f )

[0013] Among them, C map_curr (p) is the value of the confidence map rendered in the current cycle at pixel coordinates pp; T(t f This represents the total transmittance of the light rays rendering that pixel at the end of its integral path. This step not only outputs an image for viewing, but more importantly, generates crucial self-evaluation information for the entire system to use in the next cycle of decision-making.

[0014] Finally, in step d, the system performs adaptive optics compensation and display based on the generated left-eye and right-eye images and the current periodic rendering confidence map. In a preferred embodiment, this step is implemented through an active microlens electrically controlled beam shaping display panel. This panel has a dual compensation and display function: firstly, it precisely controls the direction of the emitted beam from each microlens according to the predicted viewpoint, projecting the beams corresponding to the left and right eye images to the viewer's eyes separately and without crosstalk; secondly, it compensates for the physical shape of the beam in real time based on the current periodic rendering confidence map. Specifically, for areas with high confidence values ​​in the confidence map, the system controls the microlenses to generate highly focused beams to present sharp image details; conversely, for areas with low confidence values, it actively and controllably increases the beam divergence angle to achieve a softening effect at the optical level, thereby masking potential rendering flaws in human perception.

[0015] Crucially, the method of this invention forms a feedback loop. The current-cycle rendering confidence map generated in step c will be used as the previous-cycle rendering confidence map when making predictions in the next display cycle in step b. In this way, the system can continuously and dynamically adjust its prediction behavior based on its rendering performance, forming an intelligent, adaptive, and continuously optimized workflow.

[0016] This invention provides a method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D. It has the following beneficial effects:

[0017] 1. This invention introduces a viewing intention prediction engine, transforming the system's operating mode from a passive response to the viewer's historical position to an active prediction of their future viewpoint. This proactive approach effectively compensates for the system's inherent processing latency, ensuring that the final optical image is precisely aligned with the viewer's actual eye position when the viewer moves. This fundamentally eliminates image jitter and 3D effect collapse caused by latency, significantly improving the continuity and smoothness of the viewing experience.

[0018] 2. This invention utilizes the rendering confidence map from the previous cycle to weight and correct the scene saliency map. The system can proactively reduce the visual attention-grabbing appeal of potentially poorly rendered areas. This subtly guides the predicted viewpoint towards areas of the scene that can be rendered with high quality, thereby minimizing visual artifacts and improving the overall image's perceived purity and stability without forcibly interfering with viewer behavior.

[0019] 3. This invention adaptively controls the divergence angle of the outgoing beam using an active microlens array in real time, based on the rendering confidence map of the current period. This allows the system to perform controllable optical softening in low-confidence regions. This physical-level compensation effectively masks unavoidable minor imperfections during digital rendering while maintaining sharp details in high-confidence regions, achieving synergistic optimization between digital rendering and physical optics.

[0020] 4. This invention generates a quantitative assessment of its own work quality simultaneously with the image generation by the neural rendering module. This assessment result directly serves as a key basis for the prediction decision in the next cycle. This self-assessment and continuous optimization mechanism makes the entire system highly robust, dynamically adapting to different scene complexities and viewer behavior patterns, and always maintaining optimal working conditions.

[0021] 5. This invention employs neural light fields as a continuous representation of 3D scenes, enabling the synthesis of geometrically consistent and hole-free high-fidelity images from any viewpoint, laying the foundation for high-quality display. More importantly, this invention directly correlates rendering confidence with the cumulative opacity during volume rendering, giving it clear physical meaning. This transforms the feedback signal from a vague evaluation into a precise and reliable quantitative indicator, greatly improving the accuracy and efficiency of the entire closed-loop system's decision-making. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0023] The technical solutions in 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.

[0024] Please see the appendix Figure 1 This invention provides a method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D. This method is implemented through a collaborative system that integrates prediction, rendering and optical compensation, aiming to provide a stable, smooth and high-quality naked-eye 3D visual experience for moving viewers.

[0025] The adaptive 3D display system may include a viewer state acquisition unit, an image processing unit, and a display assembly consisting of a display panel and an active optical modulation unit in terms of hardware configuration.

[0026] The viewer state acquisition unit is responsible for capturing the viewer's physical state information in real time. In one specific embodiment, this unit may consist of a depth camera and an infrared eye tracker. The depth camera is used to simultaneously acquire color images and depth information of the viewing area, providing a data basis for accurately calculating the three-dimensional spatial position of the viewer's eyes. The infrared eye tracker is used to detect the viewer's current gaze direction.

[0027] The image processing unit is the computational core of the entire system, responsible for executing the core algorithms in the method of this invention. This unit can be implemented by one or more graphics processing units, field-programmable gate arrays, or dedicated integrated circuits to meet the requirements of high-throughput real-time computing.

[0028] The display assembly is responsible for the final image presentation. The image source, based on the display panel, can be a high-resolution, high-refresh-rate self-emissive display panel. An active optical modulation unit, closely covering the display panel, is an electrically controllable optical array, such as an active microlens array. Its function is to independently and in real-time modulate the direction and shape of each beam of light emitted from the display panel according to instructions from the image processing unit.

[0029] The software architecture running on the image processing unit can be divided into multiple cooperating software modules, including: a state acquisition module, a viewing intent prediction engine, a neural rendering module, and an optical control module.

[0030] A complete workflow of the method of the present invention, namely the closed-loop flow of data between the aforementioned hardware and software modules, is as follows:

[0031] At the start of a work cycle, the viewer state acquisition unit captures raw data. The state acquisition module then processes this raw data to calculate the current three-dimensional coordinates E of the viewer's eyes. L E R and the current visual focus vector G vec This precise information about the current viewer state is then passed to the viewing intent prediction engine.

[0032] The viewing intent prediction engine receives the current viewer state. Simultaneously, it reads feedback information left over from the previous cycle—the rendering confidence map C from the previous cycle—from the memory of the image processing unit. map_prev The engine integrates these two inputs and combines them with scene saliency analysis of the content to be displayed, executes a prediction algorithm, and ultimately generates the predicted viewpoint V for the next moment. L and V R° .

[0033] This predicted viewpoint is then sent to the neural rendering module. The neural rendering module, based on V... L and V R° From a pre-trained neural light field model representing a 3D scene, the corresponding left-eye image I is rendered and generated. L And right eye image I R Simultaneously with rendering, this module also generates a current-cycle rendering confidence map C, which is closely related to the current rendering process. map_prev .

[0034] After rendering, the generated left eye image I L Right eye image I R And the current cycle rendering confidence map C map_currThis is also passed to the optical control module. Simultaneously, Cmap_curr is written to memory to overwrite C. map_curr This provides the latest feedback basis for predicting the intentions of the next work cycle, thus forming the core feedback loop of the system.

[0035] Based on the received image information and confidence map, the optical control module calculates precise control signals for each microlens in the active optical modulation unit. These signals drive the active optical modulation unit to precisely modulate the light emitted from the display panel, ultimately projecting compensated, high-quality 3D image content into the viewer's eyes, completing one full adaptive display cycle. This process cycles continuously at an extremely high frequency, enabling real-time and smooth response to changes in viewing conditions.

[0036] In one specific embodiment, the detailed implementation of the viewer state acquisition module is described. This module is responsible for converting the raw sensor data from the viewer state acquisition unit into accurate and structured viewer state information required for subsequent processing steps.

[0037] The process of acquiring 3D eye coordinates in this module begins with the synchronized alignment of a color image and a depth map captured by a depth camera. First, the state acquisition module processes the color image using a cascaded computer vision algorithm. An initial face detector locates the viewer's face region within a large image area. Subsequently, a more refined facial keypoint detection model runs within this facial region to accurately identify multiple feature points, including the centers of both pupils, thereby obtaining the pixel coordinates of the left and right pupils on the 2D image plane, denoted as p. L,img and p R,img .

[0038] After obtaining the two-dimensional pixel coordinates of the pupil, the module will query the depth map aligned with it. This is done by reading p... L,img and p R,img The depth value at the location gives the straight-line distance from the center point of each pupil to the camera plane, denoted as z. L and z R .

[0039] Next, the module performs a coordinate system transformation process, combining the two-dimensional image coordinates with depth information to reconstruct the three-dimensional coordinates in the camera coordinate system. This process utilizes the pre-calibrated camera intrinsic parameter matrix K of the depth camera. Taking the left eye as an example, its three-dimensional coordinates P in the camera coordinate system are... L,cam It can be calculated using the following formula:

[0040]

[0041] in, It is a two-dimensional pixel coordinate p L,img Homogeneous coordinate representation, K -1 It is the inverse of the intrinsic parameter matrix K. The intrinsic parameter matrix KK characterizes the internal optical properties of the camera.

[0042] To unify the coordinates in the camera coordinate system to the world coordinate system used by the entire system (e.g., a coordinate system with the center of the display screen as the origin), the module further utilizes pre-calibrated camera extrinsic parameters, namely the rotation matrix. Perform a rigid body transformation using the translation vector T. The final three-dimensional coordinates of the left eye in the world coordinate system after the transformation are E. L The calculation is as follows:

[0043] E L =R·P L,cam +T

[0044] Similarly, the three-dimensional coordinates E of the right eye can be calculated. R To improve the stability of the output coordinates and reduce coordinate jitter caused by minor body movements or sensor noise, the state acquisition module can process the calculated coordinate sequence (E... L E R Apply a time-dimensional filter, such as a low-pass filter or a one-dimensional Kalman filter, to output a smoother, more fluid coordinate flow.

[0045] Meanwhile, the state acquisition module processes data from the infrared eye tracker to calculate the viewer's visual focus vector. The raw output of the eye tracker is typically the relative position of the pupil center with respect to the corneal reflective spot (i.e., the Purchiner image). Internally, the module contains a user-calibrated geometric model or mapping function that converts this raw two-dimensional relative position data into a unit vector, denoted as G, representing the gaze direction in a three-dimensional world coordinate system. vec This vector, together with the aforementioned three-dimensional eye coordinates, constitutes a complete description of the current viewer's state and is output to the viewing intent prediction engine.

[0046] In a specific embodiment of the present invention, the viewing intention prediction engine is described in detail. This engine is the core decision-making unit of the method of the present invention, and its function is to integrate multi-source information to transform the system from a passive response to the viewer's historical position to an active and forward-looking prediction of the viewer's future visual focus.

[0047] The engine's operation begins with the analysis of the 3D scene content to be displayed. In one embodiment, the engine integrates a pre-trained, lightweight fully convolutional neural network model, such as a variant of the U-Net architecture. This model receives one or more rendered views of the scene to be displayed as input, extracts and upsamples features at multiple scales, and finally outputs a two-dimensional scene saliency map S. map (p). Each pixel in this image The grayscale values ​​are normalized, and their magnitudes directly correspond to the probability that the content at that location in the scene will visually attract human attention.

[0048] Obtain the scene saliency map S map (p) Following this, the viewing intent prediction engine performs its crucial feedback correction step. It reads the rendering confidence map from the shared memory of the image processing unit, generated by the neural rendering module in the previous display cycle. To proactively avoid directing the viewer's attention to potentially poorly rendered areas, the engine uses this confidence map to perform a weighted correction on the scene saliency map. This correction process follows the following defined mathematical relationship:

[0049]

[0050] Among them, S mod (p) is the final guide image obtained after revision, which combines content appeal and rendering feasibility; S map (p) is the original scene saliency map; It represents the feedback rendering confidence from the previous cycle; while W(·) is a non-linear weight function. In one specific implementation, W(·) can be designed as a parameterized sigmoid function to provide a smooth weight transition:

[0051]

[0052] In this function, c is the input confidence value; c0 is a preset confidence threshold, representing an acceptable benchmark for rendering quality; and k is a gain coefficient used to control the steepness of the weight curve near the threshold. This function causes the weights in high-confidence regions to approach 1, while the weights in low-confidence regions are suppressed and approach 0.

[0053] After generating the corrected significance map Next, the engine enters the final predictive viewpoint generation stage. This stage uses a time-series prediction model, such as an extended Kalman filter, to model and predict the viewer's head movements. The state vector x of this filter... tAt time t, it is defined as a combination of the viewer's three-dimensional eye position, linear velocity, and linear acceleration to comprehensively describe its kinematic state.

[0054] A key aspect of this invention is that it integrates visual guidance information as an external control input into the filter's prediction update process. Specifically, the engine first starts from the corrected saliency map... The region with the highest probability is identified, and its centroid is calculated. This centroid location is then combined with the current visual focus vector G provided by the state acquisition module. vec Together, these constitute a target attraction vector. This attraction vector is used as an external control input to the Kalman filter, thus considering not only the inertial motion trend based on historical data during state prediction, but also introducing an active guiding force determined by scene content and rendering quality. Through the filter's prediction step, the engine finally outputs a smoothed and forward-looking predicted binocular viewpoint V for the next moment. L and V R They are then sent to the neural rendering module.

[0055] In a specific embodiment of the present invention, the confidence-aware neural rendering module is described in detail. This module receives a predicted viewpoint from a viewing intent prediction engine, and its core task is to efficiently synthesize a high-quality two-dimensional image from a continuous representation of the scene, and simultaneously generate a quantitative evaluation result of the quality of this synthesis, i.e., the confidence map for the current period of rendering.

[0056] In a preferred embodiment, the module uses a neural light field as a representation of the three-dimensional scene. The neural light field model F... NeRF The system was built through an offline training process before operation. The training data consisted of a set of two-dimensional images of the target scene taken from different known camera positions. The training process aimed to optimize the weight parameters Θ of a multilayer perceptron (MLP) network, enabling the network to learn a continuous mapping from arbitrary 3D spatial coordinates x = (x, y, z) and 2D observation direction d = (θ, φ) to a color value c = (r, g, b) and a scalar volume density value σ.

[0057] (c,σ)=F NeRF (x,d;Θ)

[0058] After training, this compact neural network F NeRF This results in a high-fidelity, continuous digital representation of the entire 3D scene.

[0059] During the real-time rendering phase, after the neural rendering module receives the predicted viewpoint, it generates the left-eye image I to be generated. L For each pixel in the image, the rendering process is performed. For any pixel p in the image, the system renders from the viewpoint V.L Starting from this point, construct a ray r(t) = V that passes through the center of the pixel and has a direction of d. L +td, where t is the step distance along the direction of the light ray.

[0060] To ensure computational feasibility, the continuous integral rendering process is discretized. The system is divided into near and far cutoff planes t. n and t f Between, sample N discrete spatial points x along the ray r(t). i =r(t) i For each sampling point x i Query the pre-trained neural light field model F NeRF To obtain its corresponding color value and volume density value σ i The final color of this pixel. An approximate calculation is performed by combining the color and transparency of these sampling points using alpha.

[0061]

[0062] Where, α i It is the transparency of the i-th sampling point interval, which is determined by the volume density σ. i The distance δ between adjacent sampling points i =t i+1 One t i Joint decision: a i =1-exp(-σ i δ i T i It is the cumulative transmittance before the light reaches the i-th sampling point, representing the probability that the light is not blocked by the object in front of it:

[0063]

[0064] A key step in this invention is to simultaneously generate the current period rendering confidence map C while calculating the pixel color. map_curr (p). The confidence plot value is precisely defined as the cumulative opacity of the rendered light rays, that is, the total probability that light rays are absorbed or scattered after passing through the entire scene. This value can be directly calculated using the cumulative transmittance during the volume rendering process, and its discrete form is calculated as follows:

[0065]

[0066] This formula intuitively reflects the physical process of rendering: if the cumulative volume density along a ray path is high, resulting in a high final transmittance T... N+1 If the value approaches 0, then the confidence level C of that pixel is... mIf ap_curr(p) approaches 1, it indicates that the color is determined by one or more opaque surfaces encountered along the path, and the result is reliable. Conversely, if light passes through the entire scene with almost no attenuation, its final transmittance is close to 1, and the confidence level is close to 0, indicating that the color of the pixel is mainly determined by the background, and the rendering result contains very little information about the internal geometry of the scene, resulting in high uncertainty.

[0067] After completing the calculations for all pixels, the neural rendering module outputs the final left eye image I. L And the corresponding pixel-aligned current-cycle rendering confidence map C map_curr They are then transmitted together to the subsequent optical control module.

[0068] In a specific embodiment of the present invention, the adaptive optics compensation and display process is described in detail. This process is completed collaboratively by the optical control module and the display assembly. This process is the final execution stage of the method of the present invention, responsible for converting the rendered high-quality digital image into a physically compensated, high-quality optical image that is precisely projected into the viewer's eye.

[0069] In one specific embodiment, the active optical modulation unit is an active microlens array based on liquid crystal technology. The arrangement of liquid crystal molecules within each microlens unit of the array can be precisely controlled by applying different voltage signals. By changing the effective refractive index distribution of the liquid crystal molecules, the optical characteristics of each microlens unit, such as its focal length and principal axis direction, can be dynamically and independently altered, thereby achieving real-time shaping and deflection of the light beam passing through it.

[0070] When the optical control module receives the left eye image I generated by the neural rendering module L Right eye image I R And the current cycle rendering confidence map C map_curr Then, it first puts I L and I R Based on the sub-pixel arrangement of the display panel, a final image to be displayed is formed. Subsequently, the optical control module generates a composite control signal for each microlens unit in the active optical modulation unit. The generation logic of this signal is coordinated, and it includes both precise beam direction delivery and adaptive compensation for beam shape.

[0071] To control the beam direction, the module first calculates the target direction of the emitted beam for each microlens unit. Taking a microlens responsible for projecting the image content for the left eye as an example, its target direction vector... The three-dimensional position pos of the microlens in the world coordinate system m And left eye predictive viewpoint V LJointly determined:

[0072]

[0073] unit vector It is converted into a specific control voltage and applied to the corresponding microlens unit, so that the main optical axis of its emitted beam is precisely pointed to the center of the viewer's moving left pupil.

[0074] Another key aspect of this invention is the adaptive compensation of the beam shape. The optical control module renders a confidence map C based on the current period aligned with image pixel p. map_curr (p) dynamically adjusts the divergence angle θ of the emitted beam. d iv(p). This adjustment process is performed through a preset control function f. div The function maps the input confidence value to a target divergence angle. In one embodiment, the function can be designed as a nonlinear function inversely proportional to the confidence value.

[0075] θ div (p)=θ min +(θ max -θ min )·(1-C map_curr (p)) γ

[0076] Where, θ min It is the minimum divergence angle allowed by the system, corresponding to the sharpest focusing state; θ max It is the maximum allowable divergence angle, corresponding to the smoothest diffusion state; γ is an exponential coefficient greater than 0, used to adjust the degree of nonlinearity of the compensation effect.

[0077] According to this function, when the rendering confidence of a pixel... When it is very high, the corresponding beam divergence angle θ div (p) will approach θ m The beam of light is focused to create a highly focused beam, resulting in clear, sharp image details on the viewer's retina. Conversely, when a pixel's rendering confidence is low, its divergence angle will approach θ. max This creates a relatively diffuse and soft light spot. This controllable softening, achieved at the optical physics level, can effectively utilize the visual integral effect of the human eye to mask and blur potential visual flaws caused by uncertainties in the rendering model, thereby improving the overall perceived quality of the image without sacrificing the clarity of high-confidence areas.

[0078] Finally, a composite control signal containing both directional and morphological information is applied to each microlens of the active optical modulation unit. The display panel illuminates the interwoven image, and the light emitted is precisely shaped and projected by the modulation unit to form two separate and clear viewing windows in the viewing space. This delivers the image content for the left and right eyes to the viewer's eyes separately and precisely, completing a high-quality, crosstalk-free, and intelligently optically compensated naked-eye 3D display.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D, characterized in that, Includes the following steps: a. Obtain the current viewer status of the viewer; b. Based on the current viewer state obtained in step a and the rendering confidence map of the previous period, predict and generate the prediction viewpoint for the next moment. c. Based on the predicted viewpoint obtained in step b, render and generate left-eye and right-eye images, and generate a current-cycle rendering confidence map that characterizes the quality of this rendering. d. Based on the left-eye and right-eye images obtained in step c and the current periodic rendering confidence map, perform adaptive optics compensation and display.

2. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 1, characterized in that, In step b, based on the current viewer state and the rendering confidence map of the previous period, the predicted viewpoint for the next moment is predicted, including: analyzing the content of the scene to be displayed to generate a scene saliency map, using the rendering confidence map of the previous period to perform weighted correction on the scene saliency map to generate a corrected saliency map, and predicting the predicted viewpoint based on the current viewer state and the corrected saliency map.

3. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 2, characterized in that, The weighted correction of the scene saliency map using the rendering confidence map from the previous cycle follows the following formula: S mod (p)=S map (p)·W(C map_prev (p)); Among them, S mad (p) is the corrected significance plot, S map (p) is the saliency map of the scene, C map_prev (p) is the rendering confidence map of the previous cycle, W(·) is the weighting function for weighting the rendering confidence value, and p is the image pixel coordinate.

4. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 1, characterized in that, In step c, rendering based on the predicted viewpoint is performed by generating the left-eye and right-eye images from a continuous light field function that characterizes the three-dimensional scene using a neural light field rendering method.

5. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 4, characterized in that, In step c, the current cycle rendering confidence map is generated by calculating the cumulative opacity of the rendering rays in the light field function, and the relationship is as follows: C map_curr (p)=1-T(t f ); Among them, C map_curr (p) represents the value of the current periodic rendering confidence map at pixel coordinate p, T(t) f The transmittance of the light rays that render this pixel at the end of their path.

6. The method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D according to claim 1, characterized in that, The adaptive optics compensation and display in step d are achieved through an active microlens electronically controlled beam shaping display panel.

7. The method for dynamically adjusting the image based on viewing conditions in adaptive naked-eye 3D according to claim 6, characterized in that, The adaptive optics compensation includes: controlling the outgoing beam direction of each microlens in the active microlens array based on the predicted viewpoint, so as to project the light from the left-eye image and the right-eye image to the viewer's left and right eyes respectively; and adjusting the divergence angle of the outgoing beam based on the current periodic rendering confidence map to perform optical compensation for the rendering quality.

8. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 7, characterized in that, The adjustment of the divergence angle of the emitted beam specifically involves: for areas with high confidence values ​​in the current period rendering confidence map, reducing the divergence angle of the corresponding beam to present sharp details; and for areas with low confidence values ​​in the current period rendering confidence map, increasing the divergence angle of the corresponding beam to achieve a softening effect.

9. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 1, characterized in that, In step a, the current viewer state obtained includes the three-dimensional spatial coordinates of the viewer's eyes and the viewer's visual focus direction vector.

10. The method for dynamically adjusting the image based on viewing conditions according to adaptive naked-eye 3D as described in claim 1, characterized in that, This method forms a feedback loop by using the current cycle rendering confidence map generated in step c as the rendering confidence map of the previous cycle to be used in step b when processing in the next display cycle.

Images