Child diet preference analysis method and system based on multi-modal large model

Abstract

The application discloses a child diet preference analysis method and system based on a multi-modal large model, relates to the mineral screening technical field, and comprises the following steps: S1, acquiring a color temperature drift fingerprint, establishing a reflection spectrum separation baseline, generating a space-time reference matrix through time-synchronized light mapping, and locking the light source intensity, incident angle and color temperature parameters corresponding to the high-reflectivity area by using the reference matrix; S2, extracting the multi-angle polarization features of the meal tray image area based on the space-time reference matrix, registering the light incident path in combination with a semantic timestamp rearrangement mechanism, positioning the mirror reflection artifact, and tracing the light source characteristics causing the artifact. Through light modeling and reflection correction, in combination with multi-dimensional feature extraction and energy level inversion, the application accurately recovers food nutrition texture; and by using light color coupling and light field regulation, the application realizes image color unification and reflection adaptive repair, and improves the accuracy and stability of child diet preference and nutrition analysis.

Description

Technical Field

[0001] This invention relates to the field of mineral screening technology, specifically to a method and system for analyzing children's dietary preferences based on a multimodal large model. Background Technology

[0002] Multimodal large-scale model-based analysis of children's dietary preferences refers to a method that integrates multimodal artificial intelligence technologies such as image recognition, speech understanding, text semantic analysis, machine learning modeling, and physiological data modeling to comprehensively analyze children's behavioral characteristics, language expression, and nutritional intake during the food selection process. Specifically, this method uses image recognition technology to identify the types, colors, portions, and combinations of food on a plate; simultaneously, it utilizes natural language processing technology to analyze children's language descriptions, emotional expressions, and selection preferences during meals, thereby constructing a semantic model of dietary intentions and psychological characteristics; furthermore, it combines machine learning algorithms to perform pattern mining on physiological monitoring data or historical intake records, establishing a dynamic correlation mapping between food selection and nutritional intake, and achieving intelligent analysis and prediction of children's dietary preferences, health risks, and nutritional balance. This method can provide data support and intelligent decision-making basis for personalized dietary guidance, nutritional intervention, and health management.

[0003] The existing technology has the following shortcomings: In existing technologies, image recognition-based analysis of children's diets typically relies on visible light images to identify and extract features from food on plates. However, under dynamic lighting or color temperature variations, the food surface can develop noticeably highly reflective areas due to the influence of external light source intensity, incident angle, and color temperature fluctuations. Since current image recognition models primarily rely on brightness, texture, and color distribution for feature determination, they lack adaptive correction mechanisms for reflectance spectral distortion. This makes it highly susceptible to misidentifying these highly reflective areas as surface features of foods with excessively high fat content or abnormal sugar density. Such misjudgments not only lead to biased identification of food nutritional properties but also introduce systematic errors in the energy structure calculation stage, severely distorting the model's estimation of children's energy intake ratios and thus affecting subsequent nutritional balance assessments and the development of health intervention strategies.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for analyzing children's dietary preferences based on a multimodal large model, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for analyzing children's dietary preferences based on a multimodal large model, comprising the following steps: S1, obtain color temperature drift fingerprint, establish reflection spectrum separation baseline, generate spatiotemporal reference matrix through time-synchronized illumination mapping, and use the reference matrix to lock the light source intensity, incident angle and color temperature parameters corresponding to the high reflectivity area; S2, based on the spatiotemporal reference matrix, extracts the multi-angle polarization features of the plate image region, combines the semantic timestamp rearrangement mechanism to register the light incident path, locates the specular reflection artifact, and traces the characteristics of the light source that causes the artifact; S3. Extract the material features, shape features and humidity features of food in the specular reflection artifact area, construct a three-dimensional coupled model, and separate the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. S4. Based on the restored nutrient texture, a cross-time frame light-color coupling tensor is generated. A white balance inverse estimation mechanism is applied to adaptively reconstruct the chromaticity values ​​of each frame of the image, correct the energy density mapping, and unify the image color space. S5 generates a dynamic reversible light field injection command in the reconstructed image color space, activates a microscale phase cancellation control mechanism, and performs pixel-level time-series closed-loop regulation of reflection intensity to achieve active compensation for illumination mismatch and stable extraction of food nutritional characteristics.

[0007] Preferably, step S1 includes: After obtaining the color temperature change sequence over a continuous time period, the brightness and contrast features of the plate area and edge area in the image sequence are extracted, and the time offset relationship between brightness drift and color temperature change is calculated to generate a color temperature drift fingerprint set. After constructing the color temperature drift fingerprint set, a reflection spectrum separation baseline is generated based on the evolution path of the main color tone in the image without obvious specular reflection areas. This baseline is used to separate the spectral coupling relationship between illumination interference factors and the real reflection signal. After establishing the reflectance spectrum separation baseline, the brightness, texture density and saturation gradient are labeled with temporal and spatial coordinates to generate a mapping reference table of the image in time series and spatial distribution, and a spatiotemporal reference matrix is ​​constructed. After constructing the spatiotemporal reference matrix, based on the pixel clustering results of the brightness response peak, and combined with the direction of color temperature change and the amplitude of light intensity change, the light source intensity, dominant color temperature parameters, and incident direction parameters of the high reflectivity area are locked.

[0008] Preferably, step S2 includes: After constructing the spatiotemporal reference matrix, image sequences of the plate image area at multiple polarization angles are collected, and the attenuation, enhancement, and phase features of light intensity response as polarization angle change are extracted to generate a multi-angle polarization feature map. After generating multi-angle polarization feature maps, the image frames are semantically time-rearranged by combining image timestamps and semantic tag information to establish the correspondence between polarization response peaks and semantic events. After semantic temporal reordering is completed, spatial trajectory modeling is performed on the specular reflection region, and ray path registration is performed in combination with illumination incident parameters to construct the direction vector of the ray in consecutive image frames and complete path backtracking. Based on path tracing, the peak reflection pixel set at the end of the light path is located. Combined with the light source position parameters and polarization response characteristics at the starting point of the path, the light source type, light intensity, color temperature value and spatial angle are identified.

[0009] Preferably, during the light path registration process, perspective projection correction is performed using the geometric contours of the plate boundary, tableware structure, and tabletop edge in the image to improve the spatial modeling accuracy of the reflection area location.

[0010] Preferably, step S3 includes: After the specular reflection artifacts are located, the material features, morphological features and humidity features of the reflection area in the image are extracted. Combined with the incident light path and the spatiotemporal reference matrix, a physical foundation layer with spatial consistency and physical response consistency is constructed. After constructing the physical foundation layer, the material features, morphology features and humidity features are jointly encoded to generate a three-dimensional coupled model, which records the response trajectory of pixels under multi-dimensional attributes over time. After constructing the three-dimensional coupled model, a reflection signal separation operation based on energy level inversion is performed to identify abnormal reflection signals in the energy domain, and to trace back and remove the excess energy introduced by the high brightness artifacts. After restoring the true reflection signal, the brightness distribution, texture structure and color level of the true reflection are reprojected into the image space. Combined with the three-dimensional coupling model and humidity characteristics, the nutritional texture of the food surface is restored.

[0011] Preferably, step S4 includes: After completing the nutritional texture restoration, the spectral change information of the texture restoration area in the image is extracted, and the stable pixels are selected to generate the light and color coupling tensor to record the color change pattern of each pixel in the time series. After generating the light-color coupling tensor, we analyze the regions of color abrupt change caused by illumination in the image sequence, and calibrate and mark the color drift regions by combining the time response trajectory and light source parameters. After calibration, a white balance inverse estimation mechanism is executed, and the chromaticity value is inverted with reference pixels in the texture area as the target to complete the adaptive color reconstruction of each pixel in the image. After chromaticity reconstruction, the energy density mapping is recalculated based on the corrected image data to construct a unified energy response matrix and eliminate systematic biases introduced by reflection artifacts and color errors. After energy density correction, the image is mapped to a standard color space to achieve color resolution consistency across multiple frames, which will then be used for subsequent dietary preference modeling and health risk assessment.

[0012] Preferably, the reference pixel selected in the white balance inverse estimation mechanism is the pixel in the nutrient texture region with the smallest change in color components across multiple frames, and the chromaticity of the entire image is weighted and adjusted based on the target color value of this pixel under standard lighting conditions.

[0013] Preferably, step S5 includes: After completing the image color space reconstruction, a dynamic reversible light field injection command is generated based on multi-frame image data. A light response mapping model is established based on the light source parameters, reflection feature distribution and time series changes, and a light field adjustment function is constructed to achieve pixel-level injection control. After generating the light field injection command, based on the current reflection intensity distribution of the image, the microscale phase cancellation control mechanism is activated. By analyzing the pixel brightness changes and the response of adjacent regions, the reflection phase is reversed to form an intervention response. After activating the control mechanism, a time-series driven pixel-level closed-loop control structure is established to track the reflection change trend and adjust the light field injection parameters in real time to maintain the dynamic synchronization of the intervention. After the closed-loop regulation is completed, the structural indicators of the nutrient texture region are extracted, the regulation effect is quantified, and the injection parameters of the deviation region are optimized to perform local light field compensation operation. After completing structural stability verification and optimization, the light-color coupling tensor, energy density mapping and light field response data are integrated to perform full-image pixel-level final state verification, and the output regulated image is used for dietary preference modeling and health intervention analysis.

[0014] Preferably, the generation of dynamic reversible light field injection commands uses the incident angle of the light source, the texture clarity of the reflection area, and the color saturation as input parameters, and combines the phase position and duration period information in the time series to complete the pixel-level light field injection control for abnormal reflection areas.

[0015] A children's dietary preference analysis system based on a multimodal large model includes modules for illumination modeling, reflection artifact localization, reflection correction and texture restoration, color reconstruction, and adaptive illumination control. The illumination modeling module acquires color temperature drift fingerprints, establishes a reflectance spectrum separation baseline, generates a spatiotemporal reference matrix through time-synchronized illumination mapping, and uses this reference matrix to lock the light source intensity, incident angle, and color temperature parameters corresponding to the high reflectivity area. The reflection artifact localization module extracts multi-angle polarization features of the plate image area based on the spatiotemporal reference matrix, and registers the light incident path by combining the semantic timestamp rearrangement mechanism to locate the specular reflection artifact and trace the characteristics of the light source that causes the artifact. The reflection correction and texture restoration module extracts the material features, shape features and humidity features of food in the specular reflection artifact area, constructs a three-dimensional coupled model, and separates the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. The color reconstruction module generates a light-color coupling tensor across time frames based on the restored nutrient texture, and applies a white balance inverse estimation mechanism to adaptively reconstruct the chromaticity values ​​of each frame of the image, correcting the energy density mapping and unifying the image color space. The adaptive illumination control module generates a dynamic reversible light field injection command in the reconstructed image color space, activates a microscale phase cancellation control mechanism, and performs pixel-level time-series closed-loop control of reflection intensity to achieve active compensation for illumination mismatch and stable extraction of food nutritional characteristics.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention establishes a color temperature drift fingerprint and a reflectance spectrum separation baseline to obtain the changing patterns of illumination behavior at the physical level. Combined with polarization features, semantic rearrangement, and ray registration techniques, it achieves accurate identification and source tracing of specular reflection artifacts. Simultaneously, a three-dimensional coupling model is constructed using multi-dimensional features such as material, morphology, and humidity, and an energy level inversion mechanism is introduced to effectively separate real food reflections from highlight interference, restoring nutritional texture information obscured by highlights. Furthermore, the introduction of light-color coupling tensors and inverse white balance estimation significantly enhances image color uniformity and corrects energy density mapping errors, further improving the stability and continuity of food nutrition analysis. Finally, through closed-loop control of dynamic light field injection and phase cancellation regulation mechanisms, adaptive repair of illumination mismatch is achieved, ensuring the stability of nutritional feature information in the image over time. This comprehensively enhances the system's intelligent analysis capabilities of nutritional intake, dietary preferences, and health risks in children's eating behavior, providing highly reliable data support and decision-making basis for personalized dietary guidance and precise nutritional intervention. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1This is a flowchart of the method for analyzing children's dietary preferences based on a multimodal large model, as described in this invention.

[0019] Figure 2 This is a schematic diagram of the modules of the children's dietary preference analysis system based on a multimodal large model according to the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 The method for analyzing children's dietary preferences based on a multimodal large model, as shown, includes the following steps: S1, obtain color temperature drift fingerprint, establish reflection spectrum separation baseline, generate spatiotemporal reference matrix through time-synchronized illumination mapping, and use the reference matrix to lock the light source intensity, incident angle and color temperature parameters corresponding to the high reflectivity area; To accurately analyze the mechanism of highly reflective areas in plates and provide an illumination reference for subsequent reflection recognition and misjudgment correction, temporal calibration and spectral decoupling are performed based on visible light image sequences and auxiliary illumination sensing information. The specific steps are as follows: The color temperature change trend within the target image acquisition area is used as a dynamic parameter input into the processing flow. At the beginning of image acquisition, adjustable optical components with color temperature sensing capabilities are deployed to sample the ambient light color temperature in real time. Color temperature sampling is not limited to a single moment but is continuously acquired at multiple time points through short-cycle inter-frame acquisition, forming a temporal color temperature sequence. This sequence reflects the color temperature fluctuation trend within the sampling period, including non-homogeneous changes caused by variations in natural light, indoor lighting adjustments, and light source obstruction and reflection. Simultaneously with acquiring the continuous color temperature sequence, image processing techniques are used to extract the average brightness and contrast features of the plate area and its edge areas in each frame. The temporal offset relationship between these features and color temperature changes is analyzed to calculate the temporal fingerprint of potential color temperature drift in each frame. This temporal fingerprint, with dimensions of color temperature change rate, color temperature change direction, and brightness drift amplitude, records the fundamental characteristics of the impact of ambient light changes on the image, thus forming a color temperature drift fingerprint set to characterize the image's response pattern during dynamic lighting processes.

[0022] After obtaining the color temperature drift fingerprint set, a spectral baseline for reflective region segmentation is constructed based on this set. During reflective characteristic modeling, a background image without obvious specular reflection areas is selected as a reference. The dominant color distribution in the image is separated using spectral channel deconstruction, and combined with the temporal features recorded in the color temperature drift fingerprint, the trajectory of this dominant color change at multiple time points is fitted into a continuous spectral evolution path. Subsequently, images containing obvious highlight areas are differentially mapped to this spectral evolution path to extract the deviation of reflective enhancement components in the spectral space. These deviation values, as spectral anomaly signals, are merged into a reflective spectral separation baseline after pixel-by-pixel analysis and multi-frame comparison processing. This baseline is used to separate the spectral coupling between ambient light and the actual food surface reflectance, further clarifying which regions' spectral characteristics belong to distortion signals caused by non-physical reflective enhancement.

[0023] Based on the spectral separation baseline and the time-series color temperature drift fingerprint, an image mapping operation is performed that synchronizes the temporal and spatial dimensions. The specific steps include: using the time series as the base axis, labeling the corresponding color temperature state and reflectance enhancement index value for each frame in the image sequence; using the spatial axis as the extended dimension, three-dimensional coordinate annotation of pixel brightness, texture density, and saturation gradient of the plate edge, background, and food body regions in each frame image, thereby forming a feature distribution model of each identifiable region in the image under both temporal and spatial dimensions. By mapping and matching the above features with the previously constructed spectral separation baseline, a spatiotemporal mapping reference table is formed, recording which pixels in the image exhibit spectral drift features related to reflectance enhancement within a specific time period and under specific illumination. Based on this, a spatiotemporal reference matrix is ​​further generated. This matrix binds the brightness fluctuation curve of each pixel at different time frames to the reflectance response of its spatial location, constructing a quantifiable basic structure for inverting abnormal illumination responses from the image sequence.

[0024] Supported by a spatiotemporal reference matrix, the core illumination feature parameters of highly reflective regions in the image are identified. Specifically, this includes: clustering pixels in the reference matrix whose brightness response exceeds the average value by a certain multiple to identify regions with peak light reflection in consecutive image frames; filtering the time-varying color temperature values ​​of these regions and calculating the intensity variation and incident angle trend of the dominant light source within that time period by combining the rate and direction of change recorded in the color temperature drift fingerprint; and further analyzing the texture attenuation rate of relevant regions in the image frames to determine whether the high reflectivity coincides with structural features. Based on these analytical results, the light source intensity, dominant color temperature parameter, and incident direction value corresponding to each reflection artifact region are finally identified, completing the quantitative extraction of reflection source characteristics and providing a traceable and retrospective physical reference for subsequent specular reflection identification and correction.

[0025] S2, based on the spatiotemporal reference matrix, extracts the multi-angle polarization features of the plate image region, combines the semantic timestamp rearrangement mechanism to register the light incident path, locates the specular reflection artifact, and traces the characteristics of the light source that causes the artifact; To further achieve accurate identification and source tracing analysis of highly reflective areas, based on the generated spatiotemporal reference matrix, a multi-angle polarization light feature extraction and semantic time series rearrangement mechanism are integrated to locate specular reflection artifacts and reconstruct the illumination path. The specific steps are as follows: Based on the illumination variation regions marked in the established spatiotemporal reference matrix, an optical filter with a variable polarization angle is introduced at the front end of the image acquisition device. The reflection image sequence of the plate image region under multiple polarization directions is acquired by sequentially adjusting the polarization angle. During the polarization angle change, the light intensity response in each frame of the image will exhibit attenuation, enhancement, or phase shift phenomena as the polarization angle is adjusted. By comparing these polarized image sequences frame-by-frame with the illumination feature points in the spatiotemporal reference matrix, regions exhibiting strong polarization responses are extracted. These regions often have high specular reflection components because specularly reflected light has obvious linear polarization characteristics. By measuring the brightness variation amplitude, phase stability, and reflection direction response of each pixel in the image under different polarization angles, a multi-angle polarization feature map is constructed, and this map is used as the input basis for reflection discrimination.

[0026] Based on pixel distribution regions with significant reflection responses in the polarization feature map, a semantic timestamp reordering mechanism is initiated to perform semantic hierarchical analysis on key time frames in the image acquisition sequence. Specifically, this includes aligning the timestamp information corresponding to each frame in the acquired image sequence and, in conjunction with semantic tags carried by the scene context (such as tableware placement, hand gestures, food handling behavior, etc.), reordering the image frames along the time axis. The core of this reordering lies in identifying the action nodes or object displacement trajectories corresponding to the time periods of strongest reflection, based on the coupling relationship between the peak distribution of polarization response and the illumination change pattern, and establishing a corresponding mapping between the reflection generation time and the image semantic events. During this process, the reordering operation also needs to incorporate the illumination drift trajectory recorded by the spatiotemporal reference matrix in the preceding steps, thereby ensuring the logical consistency between semantic events and illumination change curves and avoiding temporal mismatch problems in reflection artifact analysis.

[0027] After semantic timestamp rearrangement, the spatial trajectory of the identified high-response specular reflection regions is modeled across multiple image frames. Combined with the incident light direction information recorded by the acquisition device during image acquisition in each frame, a ray path registration operation is performed. This registration operation includes: extracting the edge structure of the reflection response region and performing perspective projection correction on the positional changes of the reflection region based on known geometric contours in the image (such as plate boundaries, tableware structures, and table edges); then, based on the initial parameters of the incident light direction in the acquired images, the projection direction of the bright areas in each frame is inverted in three dimensions, gradually tracing back the incident path of the reflected light. This tracing process relies on the spatiotemporal coupling information of the light source angle and pixel brightness in the spatiotemporal reference matrix constructed in the preceding steps. A direction vector of the reflection path is constructed in each image frame, and the spatial consistency principle of consecutive image frames is used to fit and verify the reflection path, ensuring the physical rationality and temporal logical consistency of the path tracing.

[0028] Based on the completed light path registration, the source attributes of specular reflection artifacts are analyzed. Specifically, the set of pixels exhibiting peak reflection in the image at the determined endpoint of the light path is located. Then, based on the light source position parameters associated with the path's starting point, the physical characteristics of the light source are identified, including light source type, light intensity, color temperature, and its spatial angular relationship relative to the reflection area. Simultaneously, by analyzing the distribution pattern of light intensity changes at different angles in the polarization feature map, it is further inferred whether the light source is a direct incident light source or a secondary light source reflected from surrounding surfaces or objects. Finally, based on the composite relationship between light source attributes and the geometric structure of the reflection area, a mapping chain from image reflection artifacts to the physical source of light is established, providing a complete foundation for subsequent operations such as reflection artifact removal, image reconstruction, and food texture restoration.

[0029] S3. Extract the material features, shape features and humidity features of food in the specular reflection artifact area, construct a three-dimensional coupled model, and separate the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. To further achieve accurate reconstruction of the actual food surface information in highly reflective areas, based on the completed localization of specular reflection artifacts and tracing of light paths, it is necessary to separate the actual physical reflections from the artifacts within the specular reflection area and reconstruct the nutritional texture of the food. The specific steps are as follows: In the image region after specular reflection artifact localization, material, morphological, and humidity characteristics of the food material under illumination conditions are extracted. Material characteristics include the diffuse reflectance of the food surface, microparticle density, and refractive index; morphological characteristics include the spatial unevenness of the food surface, edge undulation, local curvature, and geometric texture distribution; humidity characteristics are modeled through the asymmetry of image brightness distribution, surface gloss variations, and attenuation frequency of micro-reflection structures. These features are all based on the light incident path obtained in the previous steps and the relationship between the light source direction and brightness established in the spatiotemporal reference matrix. Combined with the continuous brightness change sequence of multiple frames, regions with spatial and physical response consistency are extracted from the image. Texture compensation and structural analysis are performed on these regions to construct the physical foundation layer of the original reflective region. In this way, the reflection artifact region and the non-reflective region can be distinguished in the physical property space, providing an initial parameter set for subsequent energy level inversion.

[0030] Above the base layer, the extracted material, morphology, and humidity features are uniformly modeled to construct a three-dimensional coupled model. This model jointly expresses the response state of each pixel in three physical dimensions, forming a multi-dimensional response matrix at the pixel level. Specifically, for each pixel, the refractive index change, local curvature configuration, and corresponding humidity-induced light reflection amplitude change at its location are coupled and encoded, and the trajectory of this coupled response across different frames is recorded in the image time series. This modeling process does not rely on any simplifying assumptions but is based on the physical illumination path and specular reflection artifact distribution map retrieved from previous images, achieving high-precision modeling of each pixel under multi-dimensional attributes. This three-dimensional coupled model not only describes the physical state of the food surface in the current image but also has the ability to predict the optical response under different illumination conditions, thus providing a highly distinguishable criterion for separating real reflection signals from artifact signals.

[0031] After the 3D coupled model is established, a reflection signal separation operation based on energy level inversion is performed. During energy level inversion, the pixel response state expressed by the 3D coupled model is considered as an energy transition behavior between the surface excited state and the ground state. By analyzing the standard energy absorption and reflection laws of each type of material under given illumination conditions, an energy shift map of the artifact region is constructed. In this map, the abnormal reflection peaks exhibited by the bright artifacts are considered abnormal signals exceeding the energy transition range of normal materials, thus forming a clear stratification with the real reflection signal in the energy domain. Based on this stratification, inversion processing is performed on the pixels in the artifact region of each frame image, gradually stripping away the excess energy caused by specular reflection, restoring the energy state it should present under no reflection enhancement effect, thereby achieving the retrospective restoration of the real reflection signal. This inversion is not only based on static image features but also integrates the reconstructed incident light path and reflection source attributes from previous steps, ensuring physical consistency and temporal continuity in the separation process, guaranteeing the authenticity and stability of the restoration results.

[0032] Based on the reconstruction of the true reflection signal, the nutritional texture of the food surface area in the image is restored. Specifically, this includes: reprojecting the brightness distribution, texture structure, and color levels presented by the true reflection signal into the original image space; reconstructing the fiber arrangement direction, microstructure patterns, and thermal processing traces in the surface texture using the calibrated material and morphology response features in the 3D coupled model; and compensating for the loss of gloss transition bands on the food surface due to highlight coverage by combining the dynamic attenuation performance of humidity features, further restoring the natural color transitions and the softening of shadow edges. This restoration process emphasizes starting from the physical response, performing pixel-by-pixel completion and reproduction of each texture detail, and working in conjunction with the preceding reflection artifact localization and energy level inversion processes. Ultimately, it restores the food nutritional texture appearance in the image that most closely resembles the actual consumption scenario, providing a high-fidelity image foundation for subsequent energy density recognition and food component analysis.

[0033] S4. Based on the restored nutrient texture, a cross-time frame light-color coupling tensor is generated. A white balance inverse estimation mechanism is applied to adaptively reconstruct the chromaticity values ​​of each frame of the image, correct the energy density mapping, and unify the image color space. To achieve consistency between image color representation and energy calculation, after restoring the nutritional texture, it is necessary to further construct a cross-timeframe light-color relationship structure, employ an inverse white balance mechanism for adaptive chromaticity value reconstruction, and ultimately complete the correction of image energy density mapping and color space unification. The specific steps are as follows: Building upon the restoration of nutrient textures, the spectral characteristic variation information of the nutrient texture regions is extracted based on the restored image pixel features, and a cross-time frame light-color coupling tensor is generated accordingly. This process relies on the separation results of the real reflection signal and the high-brightness artifact signal in the previous steps, identifying texture restoration regions in each frame and selecting pixels with stable texture responses as analysis benchmarks within these regions. For these benchmark pixels, the brightness distribution sequence, color component changes, and saturation gradients across multiple frames are extracted, and their temporal trends are modeled. By correlating these trends with the incident direction, intensity, and color temperature of the light source, a response function relationship between changes in illumination conditions and image color changes is established. Subsequently, this response relationship is tensor-expanded in the pixel dimension to form a three-dimensional structure, recording the color change pattern of each pixel in consecutive image frames. This structure is then used to construct a light-color coupling tensor containing temporal, spatial, and color components, providing structural support for subsequent chromaticity value reconstruction.

[0034] Based on the generated light-color coupling tensor, a holistic analysis of the image sequence is performed to identify color mismatch regions caused by changes in ambient lighting between frames. This identification process is not a simple color comparison, but rather analyzes the continuity, symmetry, and phase shift of color changes through the time response trajectory recorded in the coupling tensor, identifying regions of abrupt color changes that do not conform to physical laws. For example, if a region exhibits color temperature shifts and brightness jumps in consecutive frames without changes in texture, it can be determined that the region has experienced light-induced color shift. By calibrating these shifted regions and combining the light source parameters and shooting time of each frame, the mismatched regions are marked in the light-color coupling tensor, providing intervention positions and target parameter settings for the subsequent white balance inverse estimation mechanism.

[0035] A white balance inverse estimation mechanism based on the light-color coupling tensor is implemented to adaptively reconstruct the chromaticity values ​​of each frame of the image pixel by pixel. Specifically, pixels with stable colors in the nutritive texture region are used as white balance reference points. Based on the color drift of these points in different frames, the color values ​​they should present under standard lighting conditions are calculated, and these values ​​are used as targets for color correction in other areas. During correction, the time-dimensional change information provided by the light-color coupling tensor is used to inversely deduce the dominant hue of the light source at different time points, thereby reversing the type of color distortion caused by illumination interference in the current frame. After the inversion is completed, the current color value of each pixel is mapped to the target color space, and weighted adjustment between chromaticity channels is performed to correct brightness saturation or hue shifts caused by changes in the light source, making the overall image color representation tend towards a natural and balanced state.

[0036] After chromaticity reconstruction, the energy density mapping is corrected based on the reconstructed image data. This step combines the real reflection signal and nutrient texture features separated by energy level inversion in the previous step to estimate the light energy reflection density per unit area for each food region, and recalculates its light reflection efficiency under standard illumination conditions based on the white balance-adjusted color values. By re-estimating the corresponding energy values ​​of pixel brightness and color, and modeling the spatial distribution of energy density values ​​in each region within the image, a unified energy response matrix is ​​formed, eliminating systematic errors introduced by factors such as high-brightness artifacts and color mismatch. This energy response matrix not only reflects the energy distribution structure of food regions in the image but also provides accurate basic data for subsequent nutrient modeling and intake estimation.

[0037] The reconstructed color images are uniformly mapped to a standardized color space to achieve consistency in visual performance and numerical representation across different frames. Specifically, this involves: adjusting the color channels in each frame to the reference range of the standard red, green, and blue channels to ensure uniform color resolution across all images under the same parameters; linearly adjusting the contrast between color channels to eliminate local over-enhancement caused by white balance compensation; and fine-tuning the overall color curve to enhance the layering and texture clarity of food areas. After unifying the color space, the image data can proceed to the next stage of pixel-level reflectance control and nutritional characteristic steady-state inference, enabling more precise dietary preference modeling and health risk analysis.

[0038] S5 generates a dynamic reversible light field injection command in the reconstructed image color space, activates the microscale phase cancellation control mechanism, performs pixel-level time-series closed-loop regulation of reflection intensity, and realizes active compensation for illumination mismatch and stable extraction of food nutritional characteristics. To further enhance the preservation of nutritional features in images under complex lighting conditions, a reversible lighting intervention mechanism needs to be constructed, based on color space reconstruction and energy density correction, to achieve active control of reflection intensity and stable extraction of texture features. The specific steps are as follows: After reconstructing the image color space, a dynamic reversible light field injection command for reflection control is generated based on multi-frame image data under this unified color expression. This injection command simulates the actual light propagation path and establishes a light response mapping model for areas in the image where illumination mismatch may occur, based on known light source parameters, reflection feature distribution, and time-series variation patterns. In this mapping model, the reflection intensity change of each pixel is related to the texture sharpness, color saturation, and incident angle of the light source in its region. Based on these related parameters and the light-color coupling tensor formed in the previous steps, a reversible light field adjustment function is constructed, and the light field injection command is dynamically generated at the pixel level. This command not only covers the intensity distribution of the light field in the spatial direction but also includes precise control of the phase position and duration of the light field injection on the time axis, thus forming a controllable intervention mechanism for the image's illumination structure.

[0039] After the injection command is generated, the local reflection intensity distribution of the image at the current moment is calculated, and the light field injection structure is applied to the region of abnormal reflection intensity in the original image, activating the microscale phase cancellation control mechanism. This control mechanism relies on the synergistic relationship between the fine-grained brightness gradient and the time-series changes in the image. It analyzes the frequency, amplitude, and stability of brightness changes of each pixel in consecutive frames of the image, and combines this with the brightness change patterns of its adjacent pixel regions in the same time period to calculate the reflection phase offset of the pixel. Based on this, by reversely modulating the reflection phase, an light field adjustment that cancels out the original reflection waveform is implemented in the region of abnormal reflection, forming a phase intervention response. This microscale intervention process does not change the original structural information of the image, but rather generates a balancing effect of reflection energy in physical space through the principle of interference, achieving fine-tuning and correction of reflection intensity at the image level.

[0040] Based on the activation of the microscale phase cancellation control mechanism, a pixel-level closed-loop control structure with time as the core is constructed to continuously track the dynamic changes in the reflection response region of the image. In this structure, the brightness response of each pixel in consecutive frames is modeled over time and compared with its historical response trajectory to identify pixels with abnormal reflection fluctuations in the short term. For these abnormal points, the intensity component, injection angle, and duration of the aforementioned reversible light field injection command are adjusted in real time to ensure that the light field intervention always maintains a dynamic response capability synchronized with the reflection changes over time. In this way, a dynamically tracking and fast-responding reflection compensation network is constructed during image acquisition, realizing the ability to self-sensitize, self-repair, and self-close the reflection fluctuations in the image, so that the overall image maintains a stable and consistent state in terms of reflection intensity.

[0041] After completing the time-series closed-loop modulation, the structural stability of the nutrient feature regions in the image is checked. Specifically, this includes extracting the texture orientation field, edge continuity distribution, and color detail clarity parameters of the nutrient texture regions from the modulated image and comparing them with the data of the corresponding regions in the image before modulation. By comparing the degree of deviation in various indicators before and after modulation, the impact of modulation on the fidelity of the image's nutrient texture representation is quantified. For regions with significant impact, the phase matching error during the reversible light field injection process is further calculated, and the injection parameters are re-optimized based on this error, performing a local light field re-injection operation to improve the stability and consistency of the entire image in terms of nutrient feature representation.

[0042] After completing the stability verification and optimization of nutritional characteristics, a final-state unified verification of the illumination structure and energy distribution of the entire image is performed. This verification process integrates the previously constructed light-color coupling tensor, energy density mapping matrix, and light field injection response data, performing a pixel-level energy balance calculation and color consistency analysis across the entire image to ensure that all image regions exhibit consistent performance in illumination representation, color rendering, and energy density structure. Upon successful verification, the adjusted image data is output for subsequent dietary preference modeling, nutritional intake analysis, and health intervention strategy generation, ultimately achieving reliable extraction and stable identification of nutritional characteristics in children's dietary images under varying lighting conditions.

[0043] This invention establishes a color temperature drift fingerprint and a reflectance spectrum separation baseline to obtain the changing patterns of illumination behavior at the physical level. Combined with polarization features, semantic rearrangement, and ray registration techniques, it achieves accurate identification and source tracing of specular reflection artifacts. Simultaneously, a three-dimensional coupling model is constructed using multi-dimensional features such as material, morphology, and humidity, and an energy level inversion mechanism is introduced to effectively separate real food reflections from highlight interference, restoring nutritional texture information obscured by highlights. Furthermore, the introduction of light-color coupling tensors and inverse white balance estimation significantly enhances image color uniformity and corrects energy density mapping errors, further improving the stability and continuity of food nutrition analysis. Finally, through closed-loop control of dynamic light field injection and phase cancellation regulation mechanisms, adaptive repair of illumination mismatch is achieved, ensuring the stability of nutritional feature information in the image over time. This comprehensively enhances the system's intelligent analysis capabilities of nutritional intake, dietary preferences, and health risks in children's eating behavior, providing highly reliable data support and decision-making basis for personalized dietary guidance and precise nutritional intervention.

[0044] This invention provides, for example Figure 2 The children's dietary preference analysis system based on a multimodal large model shown includes a lighting modeling module, a reflection artifact localization module, a reflection correction and texture restoration module, a color reconstruction module, and a lighting adaptive control module. The illumination modeling module acquires color temperature drift fingerprints, establishes a reflectance spectrum separation baseline, generates a spatiotemporal reference matrix through time-synchronized illumination mapping, and uses this reference matrix to lock the light source intensity, incident angle, and color temperature parameters corresponding to the high reflectivity area. The reflection artifact localization module extracts multi-angle polarization features of the plate image area based on the spatiotemporal reference matrix, and registers the light incident path by combining the semantic timestamp rearrangement mechanism to locate the specular reflection artifact and trace the characteristics of the light source that causes the artifact. The reflection correction and texture restoration module extracts the material features, shape features and humidity features of food in the specular reflection artifact area, constructs a three-dimensional coupled model, and separates the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. The color reconstruction module generates a light-color coupling tensor across time frames based on the restored nutrient texture, and applies a white balance inverse estimation mechanism to adaptively reconstruct the chromaticity values ​​of each frame of the image, correcting the energy density mapping and unifying the image color space. The adaptive illumination control module generates a dynamic reversible light field injection command in the reconstructed image color space, activates a microscale phase cancellation control mechanism, and performs pixel-level time-series closed-loop control of reflection intensity to achieve active compensation for illumination mismatch and stable extraction of food nutritional characteristics.

[0045] The method for analyzing children's dietary preferences based on a multimodal large model provided in this invention is implemented through the aforementioned system for analyzing children's dietary preferences based on a multimodal large model. For details on the specific methods and processes of the system for analyzing children's dietary preferences based on a multimodal large model, please refer to the embodiments of the method for analyzing children's dietary preferences based on a multimodal large model, which will not be repeated here.

[0046] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for analyzing children's dietary preferences based on a multimodal large model, characterized in that, Includes the following steps: S1, obtain color temperature drift fingerprint, establish reflection spectrum separation baseline, generate spatiotemporal reference matrix through time-synchronized illumination mapping, and use the reference matrix to lock the light source intensity, incident angle and color temperature parameters corresponding to the high reflectivity area; S2, based on the spatiotemporal reference matrix, extracts the multi-angle polarization features of the plate image region, combines the semantic timestamp rearrangement mechanism to register the light incident path, locates the specular reflection artifact, and traces the characteristics of the light source that causes the artifact; S3. Extract the material features, shape features and humidity features of food in the specular reflection artifact area, construct a three-dimensional coupled model, and separate the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. S4. Based on the restored nutrient texture, a cross-time frame light-color coupling tensor is generated. A white balance inverse estimation mechanism is applied to adaptively reconstruct the chromaticity values ​​of each frame of the image, correct the energy density mapping, and unify the image color space. S5 generates a dynamic reversible light field injection command in the reconstructed image color space, activates the microscale phase cancellation control mechanism, and performs pixel-level time-series closed-loop control of the reflection intensity.

2. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 1, characterized in that, Step S1 includes: After obtaining the color temperature change sequence over a continuous time period, the brightness and contrast features of the plate area and edge area in the image sequence are extracted, and the time offset relationship between brightness drift and color temperature change is calculated to generate a color temperature drift fingerprint set. After constructing the color temperature drift fingerprint set, a reflection spectrum separation baseline is generated based on the evolution path of the main color tone in the image without obvious specular reflection areas. This baseline is used to separate the spectral coupling relationship between illumination interference factors and the real reflection signal. After establishing the reflectance spectrum separation baseline, the brightness, texture density and saturation gradient are labeled with temporal and spatial coordinates to generate a mapping reference table of the image in time series and spatial distribution, and a spatiotemporal reference matrix is ​​constructed. After constructing the spatiotemporal reference matrix, based on the pixel clustering results of the brightness response peak, and combined with the direction of color temperature change and the amplitude of light intensity change, the light source intensity, dominant color temperature parameters, and incident direction parameters of the high reflectivity area are locked.

3. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 1, characterized in that, Step S2 includes: After constructing the spatiotemporal reference matrix, image sequences of the plate image area at multiple polarization angles are collected, and the attenuation, enhancement and phase features of light intensity response as polarization angle change are extracted to generate a multi-angle polarization feature map. After generating multi-angle polarization feature maps, the image frames are semantically time-rearranged by combining image timestamps and semantic tag information to establish the correspondence between polarization response peaks and semantic events. After semantic temporal reordering is completed, spatial trajectory modeling is performed on the specular reflection region, and ray path registration is performed in combination with illumination incident parameters to construct the direction vector of the ray in consecutive image frames and complete path backtracking. Based on path tracing, the peak reflection pixel set at the end of the light path is located. Combined with the light source position parameters and polarization response characteristics at the starting point of the path, the light source type, light intensity, color temperature value and spatial angle are identified.

4. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 3, characterized in that, During the light path registration process, perspective projection correction is performed using the geometric contours of the plate boundary, tableware structure, and tabletop edge in the image to improve the spatial modeling accuracy of the reflection area location.

5. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 1, characterized in that, Step S3 includes: After the specular reflection artifacts are located, the material features, morphological features and humidity features of the reflection area in the image are extracted. Combined with the incident light path and the spatiotemporal reference matrix, a physical foundation layer with spatial consistency and physical response consistency is constructed. After constructing the physical foundation layer, the material features, morphology features and humidity features are jointly encoded to generate a three-dimensional coupled model, which records the response trajectory of pixels under multi-dimensional attributes over time. After constructing the three-dimensional coupled model, a reflection signal separation operation based on energy level inversion is performed to identify abnormal reflection signals in the energy domain, and to trace back and remove the excess energy introduced by the high brightness artifacts. After restoring the true reflection signal, the brightness distribution, texture structure and color level of the true reflection are reprojected into the image space. Combined with the three-dimensional coupling model and humidity characteristics, the nutritional texture of the food surface is restored.

6. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 1, characterized in that, Step S4 includes: After completing the nutritional texture restoration, the spectral change information of the texture restoration area in the image is extracted, and the stable pixels are selected to generate the light and color coupling tensor to record the color change pattern of each pixel in the time series. After generating the light-color coupling tensor, we analyze the regions of color abrupt changes caused by illumination in the image sequence, and calibrate and mark the color drift regions by combining the time response trajectory and the light source parameters. After calibration, a white balance inverse estimation mechanism is executed, and the chromaticity value is inverted with reference pixels in the texture area as the target to complete the adaptive color reconstruction of each pixel in the image. After chromaticity reconstruction, the energy density mapping is recalculated based on the corrected image data to construct a unified energy response matrix and eliminate systematic biases introduced by reflection artifacts and color errors. After energy density correction, the image is mapped to a standard color space to achieve color resolution consistency across multiple frames, which will then be used for subsequent dietary preference modeling and health risk assessment.

7. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 6, characterized in that, The reference pixel selected in the white balance inverse estimation mechanism is the pixel in the nutrient texture region with the smallest change in color components across multiple frames, and the color of the entire image is weighted and adjusted based on the target color value of this pixel under standard lighting conditions.

8. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 1, characterized in that, Step S5 includes: After completing the image color space reconstruction, a dynamic reversible light field injection command is generated based on multi-frame image data. A light response mapping model is established based on the light source parameters, reflection feature distribution and time series changes, and a light field adjustment function is constructed to achieve pixel-level injection control. After generating the light field injection command, based on the current reflection intensity distribution of the image, the microscale phase cancellation control mechanism is activated. By analyzing the pixel brightness changes and the response of adjacent regions, the reflection phase is reversed to form an intervention response. After activating the control mechanism, a time-series driven pixel-level closed-loop control structure is established to track the reflection change trend and adjust the light field injection parameters in real time to maintain the dynamic synchronization of the intervention. After the closed-loop regulation is completed, the structural indicators of the nutrient texture region are extracted, the regulation effect is quantified, and the injection parameters of the deviation region are optimized to perform local light field compensation operation. After completing structural stability verification and optimization, the light-color coupling tensor, energy density mapping and light field response data are integrated to perform full-image pixel-level final state verification, and the output regulated image is used for dietary preference modeling and health intervention analysis.

9. The method for analyzing children's dietary preferences based on a multimodal large model according to claim 8, characterized in that, The generation of dynamic reversible light field injection commands takes the incident angle of the light source, the texture clarity of the reflection area, and the color saturation as input parameters, and combines the phase position and duration period information in the time series to complete the pixel-level light field injection control for abnormal reflection areas.

10. A children's dietary preference analysis system based on a multimodal large model, used to implement the children's dietary preference analysis method based on a multimodal large model as described in any one of claims 1-9, characterized in that, It includes a lighting modeling module, a reflection artifact localization module, a reflection correction and texture restoration module, a color reconstruction module, and a lighting adaptive adjustment module: The illumination modeling module acquires color temperature drift fingerprints, establishes a reflectance spectrum separation baseline, generates a spatiotemporal reference matrix through time-synchronized illumination mapping, and uses this reference matrix to lock the light source intensity, incident angle, and color temperature parameters corresponding to the high reflectivity area. The reflection artifact localization module extracts multi-angle polarization features of the plate image area based on the spatiotemporal reference matrix, and registers the light incident path by combining the semantic timestamp rearrangement mechanism to locate the specular reflection artifact and trace the characteristics of the light source that causes the artifact. The reflection correction and texture restoration module extracts the material features, shape features and humidity features of food in the specular reflection artifact area, constructs a three-dimensional coupled model, and separates the real reflection signal and the high brightness artifact signal through energy level inversion calculation to restore the nutritional texture features of the food surface. The color reconstruction module generates a light-color coupling tensor across time frames based on the restored nutrient texture, and applies a white balance inverse estimation mechanism to adaptively reconstruct the chromaticity values ​​of each frame of the image, correcting the energy density mapping and unifying the image color space. The illumination adaptive control module generates a dynamic reversible light field injection command in the reconstructed image color space, activates the microscale phase cancellation control mechanism, and performs pixel-level time-series closed-loop control of the reflection intensity.

Images