62 results about "Non-local means" patented technology

The invention discloses a cross-modal eye fundus image generation method and system based on a generative adversarial network, relates to the technical field of medical image processing, and constructs an eye fundus focus perception and edge consistency generative adversarial network by taking a cyclic consistency generative adversarial network as a baseline. The core of the method is that a lesion perception mixed attention module is embedded in a bottleneck layer of a generator so as to strengthen the extraction capability of fine features of a lesion area; an edge information extraction module is designed, and key edge features are accurately extracted in combination with Roberts edge detection, wavelet transform and non-local mean denoising; and a joint loss function containing edge consistency loss is constructed, and the semantic consistency of a focus structure during cross-modal generation is ensured by minimizing the feature difference between the source image and the generated image. According to the method, the problems of disordered content, inconsistent structure and unstable training of the generated image in the prior art are effectively solved, and the simulation degree and clinical availability of the generated image are remarkably improved.

The invention relates to the technical field of computer vision, and discloses a jade defect intelligent detection method and system based on machine vision and deep learning, and the method comprises the following steps: S1, based on a high-resolution industrial camera and a laser three-dimensional scanner, adopting a multi-mode synchronous collection strategy, and rotating a jade sample through a precise motion control system, a jade surface high-resolution two-dimensional color image and high-precision three-dimensional point cloud data are respectively obtained, and a jade multi-mode original data set is generated. A high-resolution two-dimensional color image and high-precision three-dimensional point cloud data are integrated through a multi-modal synchronous acquisition strategy, and multi-dimensional feature expression under unified coordinates is constructed, so that the limitation of a single data source is effectively overcome; an image registration algorithm and a feature pyramid network are combined with a point cloud network to perform multi-modal feature fusion, complementarity of color texture and geometric morphology information is enhanced, and image quality is optimized based on adaptive histogram equalization and non-local mean filtering.

The invention discloses an infrared image denoising method and system based on NSST decomposition, and the method comprises the steps: S1, decomposing a transformer infrared image containing noise based on an NSST algorithm, and outputting a high-frequency sub-band coefficient, a low-frequency sub-band coefficient and a high-frequency sub-band coefficient; s2, based on the low-frequency sub-band in the S1, filtering and denoising the low-frequency sub-band by adopting an improved non-local mean algorithm, and outputting the denoised low-frequency sub-band; s3, based on the high-frequency sub-band coefficients in the S1, introducing local entropies for different high-frequency sub-band coefficients, and outputting each high-frequency sub-band after the local entropies are denoised; and S4, based on the low-frequency sub-band in S2 and each high-frequency sub-band in S3, obtaining a denoised image by adopting NSST inverse transformation, and completing infrared image denoising. According to the method, the high-frequency sub-band coefficients containing different frequency noise are denoised by adopting the local entropy threshold method, noise filtering parameters do not need to be manually set, denoising threshold parameters are established according to entropy information of the high-frequency sub-band coefficients, and the method has good adaptability.

The invention discloses a stone surface flaw automatic identification method and system based on an intelligent algorithm, and belongs to the technical field of machine vision and stone processing. The invention provides an automatic detection scheme for solving the problems that in the prior art, manual stone slab defect detection is low in efficiency and high in subjectivity, and a traditional machine vision method is insufficient in detection precision under the conditions of complex stone slab textures and mixed noise. The method comprises the following steps: constructing an image acquisition system and carrying out camera calibration and image correction; the method comprises the following steps: preprocessing a stone plate image by adopting a denoising method combining median filtering and non-local mean (NLM) filtering, and extracting a stone plate contour by combining an improved Canny algorithm; and then, solving the maximum inscribed rectangle in the contour through a histogram area method, performing histogram equalization enhancement on the rectangular region, and finally, segmenting and identifying defects such as color spots and color lines by adopting a region splitting and merging algorithm combined with morphology.

The invention relates to the technical field of industrial image quality detection, and discloses an image recognition-based toothbrush quality detection method, which comprises the following steps of: acquiring a to-be-detected image of a toothbrush, and performing non-local mean denoising and CLAHE contrast enhancement processing on the to-be-detected image to obtain a preprocessed image; pixel-level segmentation is carried out on the preprocessed image through an Attention-UNet + + model, and a toothbrush head area mask M1 and a toothbrush handle area mask M2 are generated; according to the detection method, the quality conditions of the toothbrush head and the toothbrush handle are independently and clearly known through regional evaluation, whether quality defects exist in the two regions or not and the severity of the defects can be judged respectively, and on the basis, the quality of the toothbrush head and the quality of the toothbrush handle can be judged by calculating a weighted total score Stotal and comparing the weighted total score Stotal with a dynamic threshold value T; according to the method, organic combination of independent evaluation and comprehensive decision is realized, and the combination mode not only considers quality characteristics of different areas of the toothbrush, but also comprehensively evaluates the quality of the toothbrush on the whole.

The application provides an intelligent image processing system and method for a car based on a perception algorithm model, and the method comprises the following steps: S1. spatio-temporal reference double anchoring and dynamic intrinsic extrinsic parameter calibration of a vehicle-mounted image acquisition node; S2. multi-node image heterogeneous domain normalization and adaptive preprocessing based on a self-adaptive kernel regression non-local mean denoising algorithm; S3. hierarchical image feature extraction and semantic anchoring based on a graph neural network dynamic feature interaction network; S4. cross-node and cross-frame feature mutual checking and pseudo-feature elimination; S5. full-scene semantic completion and dynamic target trajectory prediction based on a variational autoencoder trajectory prediction model; S6. dynamic lightweight adaptation and algorithm power adaptive scheduling of the perception algorithm model; and S7. risk scene grading identification and image targeted enhancement output based on semantics and trajectories. The application provides stable, accurate and efficient vehicle-mounted image perception support for intelligent driving of a car, and improves the safety and adaptability of environmental perception of intelligent driving.

The application discloses a non-local mean-based adaptive SAR interferogram filtering method, and relates to the technical field of synthetic aperture radar. Firstly, two input SAR images are used to generate an interferogram and calculate the correlation coefficient of each pixel to form a correlation coefficient matrix. Phase information of the interferogram is extracted. Pixels that do not need filtering are screened. For pixels that need filtering, different search window radii are determined according to the correlation coefficient values, and the search window radius corresponding to the pixels that do not need filtering is set to 0. For each pixel that needs filtering, the weight of each pixel in the corresponding search window is determined, and the filtering result of each pixel is obtained by using a weighted average method to form a filtered SAR interferogram.

The application discloses a kind of automatic control method and equipment for sweet potato seedling seedling pipeline operation, it utilizes differential pulse modulation to collect the excited state and background state image of vine, through space self-adaptive weighted difference modeling and non-local mean regularization enhancement technology, effectively suppress environmental light noise and retain weak biological fluorescence signal, to generate the high confidence vine node distribution map.On this basis, abandon traditional fixed-length cutting mode, construct dynamic programming model based on agronomic constraints, solve the optimal segmentation strategy of maximum output rate in the whole vine range.Finally, through the space-time fusion of visual coordinates and conveyor encoder, drive multi-axis servo system to track and accurately work the real-time virtual cutting point of dynamic movement, realize the efficient, low-loss automatic production of sweet potato seedling.

The invention relates to the technical field of intelligent image detection, in particular to a Graph mamba-based two-dimensional wavelet multi-scale high-resolution image anomaly detection system, which comprises a two-dimensional wavelet transform multi-scale high-frequency automatic self-encoding model, a high-frequency detail feature extraction model, a high-frequency detail feature extraction model and a multi-scale high-frequency automatic self-encoding model, and is used for carrying out multi-scale decomposition on an input high-resolution image to extract high-frequency detail features; generating corresponding multi-scale sub-image embedded data; and the input end of the dynamic characteristic calibration module is connected with the coding output end of the two-dimensional wavelet transform multi-scale high-frequency automatic self-coding model. According to the method, the two-dimensional wavelet transform multi-scale high-frequency automatic self-encoding model is combined with the Butterworth high-pass filtering and Marr kernel function, so that the extraction of abnormal related edge detail features is enhanced, and the detection accuracy is improved; the dynamic feature calibration module rejects inconsistent feature points through cosine similarity verification, filters redundant information in cooperation with a non-local mean de-duplication algorithm, reduces the data processing amount, and optimizes the calculation efficiency.

The application provides a RAW domain non-local mean image denoising method, comprising the following steps: S1, collecting a RAW image by a sensor; S2, dividing the image into N channels, such as R, Gr, Gb and B channels according to an actual Bayer format; S3, performing non-local mean filtering on each channel; further comprising: S301, calculating the similarity between each channel pixel point and the center pixel point of a neighborhood window; S302, calculating the weight according to the similarity information; S303, calculating the filtering result of each pixel point in each channel according to the weight; S4, merging the results of each channel after filtering into a RAW image, and outputting the result. The application utilizes the original characteristics of noise to perform image filtering in the RAW domain, reduces the difficulty of noise suppression, and has smaller side effects under the premise of achieving the same filtering effect; the non-local mean filtering is adopted, so that the problems of pattern noise after filtering, and false color and color cast caused by unbalanced channel filtering results are avoided.

The present application relates to the technical field of machine vision, in particular to a kind of robust visual detection system of ton bag goods based on multi-cluster feature fusion and posture correction drive, the system includes non-local mean enhancement module, multi-feature clustering fusion module, geometric feature extraction module, dominant posture clustering module, posture correction transformation module and goods segmentation determination module, non-local mean filtering is carried out to the original image of ton bag goods to obtain enhanced image, the present application is equipped with non-local mean enhancement module and multi-feature clustering fusion module, can realize the efficient enhancement of ton bag goods image and multi-dimensional feature fusion, the space alignment weighted superposition of color cluster and texture cluster is generated Fusion segmentation chart, let the feature representation of ton bag goods homogeneous region be more comprehensive, accurate, effectively improve the integrity and effectiveness of geometric feature extraction, so that the feature analysis ability of system to ton bag goods image is significantly improved.

The present application relates to the technical field of image processing, in particular to a kind of cardiovascular medicine OCT image filtering enhancement method, the method comprises: by obtaining optical coherence tomography OCT image, and the classification distribution parameter and tissue distribution parameter of image, wherein classification distribution parameter is based on the light and dark change of original image and is obtained by classifying original image, and tissue distribution parameter is based on the tissue distribution of original image and is obtained by segmenting original image.Then based on the difference of classification distribution parameter and tissue distribution parameter, determine the noise influence factor of original image, and according to noise influence factor, original image is carried out non-local mean filtering, clear target image that inhibits speckle noise is obtained, finally, target image is enhanced, it is ensured that the OCT image after enhancement retains important structure details, and unnecessary noise interference is reduced.

The invention relates to a welding seam laser point cloud data acquisition and preprocessing technology, and provides a method for integrating multi-source data fusion and an AI enhancement algorithm, which comprises the following steps: S1, laser triangulation ranging: projecting a light band to the surface of a welding seam through a line laser, capturing a distortion image by combining a high-resolution camera, and carrying out laser triangulation; displacement sensor compensation and adaptive focal length adjustment functions are integrated, and the surface three-dimensional morphology is dynamically reconstructed; s2, multi-modal denoising: matching an optimal filtering strategy based on frequency domain analysis, fusing a deep learning model and improved bilateral filtering, and improving the image quality in a complex noise scene; s3, sub-pixel-level track extraction: a non-local mean-Hessian matrix fusion algorithm is combined with multi-scale eigenvalue decomposition to realize high-precision positioning of the center of the light band; s4, curvature-driven outliers are removed, specifically, a local curvature self-adaptive threshold value method is combined with a graph optimization algorithm, and weld edge feature points are prevented from being mistakenly deleted; and S5, layered resampling: preferentially reserving high-precision region points based on curvature importance, and optimizing space coverage uniformity. According to the method, the defects in the aspects of environmental adaptability, noise suppression and feature integrity in the prior art are overcome, high-speed, high-precision and high-robustness welding seam point cloud collection and preprocessing are achieved, and the method is suitable for special-shaped welding seam detection under complex working conditions.

The invention provides a wafer defect detection method and system based on YOLO-Label comparison, and the method comprises the steps: obtaining a to-be-detected wafer image, carrying out the adaptive acceleration non-local mean filtering denoising and adaptive multi-scale gradient enhancement Canny edge detection of the to-be-detected wafer image, and obtaining a binary edge image; inputting the image into a pre-trained target detection model to obtain target positioning information in a YOLO-Label format, wherein the model is obtained by training a YOLOv8n network which is subjected to Ghost convolution lightweight and single-channel adaptive transformation by adopting a wafer binary image sample set; and comparing the target positioning information of the to-be-detected image with the standard positioning information of the defect-free image, extracting bounding box feature points, calculating an Euclidean distance matrix, and judging missing defects or redundant defects based on threshold matching. According to the invention, high-precision, high-efficiency and high-robustness wafer defect automatic detection can be realized with less labeled data under a complex imaging condition, and the detection speed and integrity are significantly improved.

The invention discloses a cold region concrete dam dangerous case intelligent identification method based on image information, and the method comprises the following steps: constructing a cold region concrete dam dangerous case image noise reduction method through the comprehensive application of wavelet transform, improved non-local mean noise reduction and a bilateral filtering algorithm; through improvement of a multi-scale Retinex algorithm and fusion of gamma transformation and a contrast-limited adaptive histogram equalization algorithm, a cold region concrete dam dangerous case image quality enhancement method is provided. Based on the pre-processed high-quality image, combining with an improved YOLOv8 network structure, introducing an attention mechanism to enhance the feature extraction capability, and constructing an intelligent identification model for the dangerous case of the concrete dam in the cold region; according to the method, the recognition precision of typical diseases such as cracks and freeze-thaw damage is remarkably improved.

The present application relates to the technical field of vacuum gauge detection, and particularly relates to a vacuum degree monitoring method and system for refrigerator vacuumizing. The method comprises the following steps: constructing a selectable range of data points with each data point in a time sequence of refrigerator vacuum degree as the center; obtaining a noise index of the data point through the difference between the data point in the selectable range and the average value of the selectable range; calculating an abnormal factor of each data point in the selectable range of the target data point; obtaining the number of single-sided data points of the target data point through the proportion of the abnormal factor in the single-sided selectable range of the target data point, so as to construct a reference data segment of the target data point; and using the reference data segment of each data point in the non-local mean filtering algorithm to denoise the time sequence of the vacuum degree, so as to realize the vacuum degree monitoring of the refrigerator vacuumizing and effectively improve the accuracy of the refrigerator vacuum degree data monitoring.

The application discloses an endoscope image enhancement method and device, equipment and medium, and relates to the technical field of endoscopes. The method comprises the following steps: performing guided filtering on an original image returned by an endoscope to obtain an original base layer image of the original image, and performing image subtraction processing on the original image and the original base layer image to obtain an original detail layer image of the original image; performing non-local mean filtering on the original detail layer image to obtain a filtered detail layer image, and performing gain control on the filtered detail layer image to obtain a target detail layer image; processing the original base layer image by using a limited contrast adaptive histogram equalization algorithm, a non-uniform illumination correction algorithm and a linear weighted fusion algorithm to obtain a target base layer image; and performing image fusion on the target detail layer image and the target base layer image to obtain an enhanced target image of the original image. The endoscope image can be reasonably enhanced to improve the image quality.

The invention discloses an image sequence flicker elimination method and system based on clustering and multi-scale histogram matching, and relates to the technical field of image processing, and the method comprises the steps: carrying out the denoising processing of an image through employing a non-local mean value denoising algorithm, obtaining a denoised image sequence, calculating the histogram features, and carrying out the calculation of the histogram features; grouping the images in the de-noised image sequence by adopting a K-Means clustering algorithm so as to determine a category label of each frame of image; for the de-noised image sequence, taking an image of which the category label is a standard category as a reference, performing multi-scale histogram matching on an image of which the category label is an abnormal category, and performing weighted fusion on optimized images under all scales through a Gaussian kernel function to obtain an optimized image sequence; and for the optimized image sequence, carrying out time sequence smooth transition processing at the image junction of different types of labels by adopting a cross gradual change technology so as to obtain a final image sequence. The flicker of the image sequence can be effectively eliminated, and the visual experience is improved.

The application discloses a kind of image noise suppression method and system for laser speckle, it is related to image noise suppression technical field, first, two laser images of same position are acquired, and two-dimensional discrete wavelet transform is carried out, and high, low frequency image is obtained;High, low frequency image is fused, and high-frequency fusion image and low-frequency fusion image are obtained;Optimized non-local mean filtering algorithm is used to filter high-frequency fusion image, and filtered high-frequency fusion image is obtained;By wavelet inverse transform, low-frequency fusion image and filtered high-frequency fusion image are restored, and restoration image is obtained;Optimized non-local mean filtering algorithm is used to filter restoration image, and final result image is obtained.The application is aimed at laser speckle high-frequency noise, introduces wavelet transform and improved optimized non-local mean filtering combination, not only eliminates high-frequency speckle noise in laser image, but also completely retains low-frequency information part, and the obtained image effect is best.

The invention discloses a mesoscale eddy detection method and system based on an improved RT-DETR network structure. According to the method, firstly, a multi-source high-resolution SAR satellite image is preprocessed through a fast non-local mean denoising algorithm, so that the data quality is effectively improved; secondly, aiming at the problem that the boundary of a scale vortex is fuzzy in an SAR image, a feature extraction network, a feature fusion network and a loss function of an end-to-end target detection model RT-DETR are improved, and based on the improved RT-DETR network structure, the capability of capturing vortex features of different scales is enhanced through a multi-scale feature fusion module; the innovatively designed Inner-MPDIOU loss function remarkably improves the bounding box regression precision, the problem of detection of the fuzzy vortex boundary in the SAR image is solved, and high-precision and high-efficiency mesoscale vortex detection can be achieved through the method.

This invention provides an adaptive window clustering method for shale MAPs based on a fuzzy neural network, relating to the fields of oil and gas development and image segmentation. Specifically, it includes the following steps: performing domain decomposition, nonlocal mean filtering, and gray-level normalization on the shale MAPs; dynamically adjusting the sliding window based on local gray-level variance; characterizing the shale micro-features within the sliding window into gray-level features and structural features using the gray-level co-occurrence matrix and structural parameters; and combining the gray-level features and structural features as a feature vector and inputting it into a fuzzy neural network (FNN) for micro-feature clustering. The technical solution of this invention overcomes the problem in existing technologies that cannot simultaneously consider the multi-scale micro-features and boundary uncertainties in shale MAPs.

This invention discloses an adaptive filtering method for phase-sensitive optical time-domain reflectometry (OTDR) signals. The method evaluates the quality of non-local mean filtering results by calculating correlation coefficients, and the filtering quality can be used to optimize the filtering parameters of the image filtering method. These parameters are first verified using a hypothetical standard signal and then applied to denoising a real signal from a phase-sensitive OTD instrument. Comparison of the correlation-optimized filtering parameters with those optimized using traditional peak signal-to-noise ratio (PSNR) and structural similarity methods shows that, except for slight differences in the neighborhood window size, the optimized filtering parameters are almost identical. This method is also applicable to image filtering of other two-dimensional periodic signals.

The invention relates to the technical field of signal detection, and discloses a pipeline leakage detection method based on an acoustic emission signal and DBN-GA-LSSVM hybrid architecture, computer equipment, a computer readable storage medium and a computer program product in order to solve the problem of low detection precision of a traditional pipeline leakage detection method. The method comprises the following steps: collecting an acoustic emission signal during pipeline operation; converting the acoustic emission signal from a one-dimensional time-domain signal into a two-dimensional time-frequency scale map by using continuous wavelet transform; non-local mean filtering and self-adaptive histogram equalization are adopted to process the scale image, and an enhanced leakage induction scale image is obtained; inputting the enhanced leakage induction scale map into a deep belief network, and outputting a feature vector; optimizing the feature vector by using a genetic algorithm; and inputting the optimized feature vector into a least square support vector machine, classifying the health state of the pipeline, and realizing leakage state identification or leakage-free state identification. By adopting the method, the accuracy of pipeline leakage detection can be improved.

The application discloses a single image rain removal method based on rain map decomposition. The application inputs a rain map into a rain streak feature prediction network to obtain a rain streak estimation map; then inputs the rain streak estimation map into a rain density perception classifier to obtain a rain density level label of the rain map; then inputs the rain streak estimation map and the rain density level label into a rain streak pixel kernel positioning network to obtain a pixel-by-pixel rain streak pixel kernel position vector; then inputs the rain streak pixel kernel position vector into an SNIP scale adaptive range predictor to generate a corresponding rain streak coverage area, and superimposes and synthesizes to obtain a rain streak overall coverage area; finally, uses a non-local mean filter to perform rain removal processing on the rain streak overall coverage area to obtain a rain-removed image. The application accurately decomposes the rain map into a rain streak area and a background area, can better solve the problem of loss of background details caused by excessive rain removal in the rain removal process, and effectively removes the rain streak and restores the background texture detail information.

The invention relates to the technical field of property management, in particular to a property management monitoring management method and system, and the method comprises the steps: collecting a face image of a target community, and constructing an image library; de-noising the images in the image library by using a non-local mean correction de-noising algorithm; collecting route tracks of pedestrians in a target community, and constructing a route database; inputting the de-noised face image and the route database into a resident safety identification model, and outputting a pedestrian safety coefficient of the target community; the resident safety identification model comprises a face identification module for calculating face similarity by using a D-Faster R-CNN model and a route matching degree module for calculating route familiarity by using a route matching degree model; a safety threshold value is set, and when the safety coefficient is larger than the safety threshold value, pedestrians are allowed to enter; otherwise, the access is not allowed, and early warning is carried out on the security guard; two key factors of face recognition and route matching degree are considered at the same time, the safety of pedestrians in a community is analyzed from double angles, and the safety guarantee level of property management is effectively improved.

The invention discloses a CLAHE-unsupervised learning InSAR geological disaster hidden danger intelligent identification method and system, and the method comprises the steps: obtaining multi-temporal SAR data of a research region, and carrying out the processing of MT-InSAR to generate an average deformation rate field; noise suppression and weak signal enhancement are carried out through non-local mean filtering and wavelet domain multi-scale processing; a contrast-limited adaptive histogram equalization algorithm is adopted to improve the local deformation contrast, and an enhanced deformation intensity graph is generated; constructing a feature space in combination with the enhanced strength graph and an original deformation value, and preliminarily identifying a hidden danger region by using an unsupervised clustering algorithm and space-deformation double-constraint region growth; optimizing a boundary topological relation through triangulation and a minimum convex hull algorithm; and finally, calculating a comprehensive quality index based on geometric consistency, signal intensity and spatial aggregation degree, realizing risk grading, and outputting a vector boundary, a deformation parameter and a grade result.

The invention provides a method and system for diagnosing a cerebral hemorrhage period based on a large model, and the method comprises the steps: firstly, designing a method based on the large model and non-local mean filtering to enrich a cerebral hemorrhage CT image data set based on the characteristics that the number of cerebral hemorrhage CT images is small, and the images have complex noise, and then carrying out the diagnosis of the cerebral hemorrhage period based on the large model. A light-weight network module is fused to replace an encoder module in U-net to facilitate deployment of a designed detection model, finally, an equivalent volume calculation formula is designed to calculate the amount of cerebral hemorrhage of a patient, and a detection result of a cerebral hemorrhage period is given under an overall detection architecture. And timely and effective diagnosis information is provided for rapid and accurate detection of cerebral hemorrhage clinically.

The invention relates to the technical field of machine vision, in particular to a ton bag cargo robust visual detection system based on multi-cluster feature fusion and attitude correction driving. The system comprises a non-local mean value enhancement module, a multi-feature clustering fusion module, a geometric feature extraction module, a dominant attitude clustering module, an attitude correction transformation module and a cargo segmentation judgment module, and non-local mean value filtering is carried out on an original image of the ton bag cargo to obtain an enhanced image; according to the method, the non-local mean value enhancement module and the multi-feature clustering fusion module are carried, efficient enhancement and multi-dimensional feature fusion of the ton bag cargo image can be realized, the fusion segmentation image is generated through spatial alignment weighted stacking of the color clustering cluster and the texture clustering cluster, feature representation of the homogeneous region of the ton bag cargo is more comprehensive and accurate, and the method is more suitable for mass production. The integrity and effectiveness of geometric feature extraction are effectively improved, and the feature analysis capability of the system on the ton bag cargo image is remarkably improved.

The invention discloses an intelligent slow-release drug carrier screening method based on material microscopic image recognition, which integrates nano hyperspectrum, cryoelectron microscope and Raman spectrum data through a multi-modal data synchronous acquisition system to realize nanoscale space alignment of microscopic morphology and component distribution. A non-local mean denoising algorithm is adopted to enhance pore structure recognition, and chemical bond state changes are quantified in combination with adaptive baseline correction and multi-peak fitting technologies. A characteristic spectrum library is constructed, a chemical distribution diagram is generated through spectrum angle mapping, and the co-localization of the carrier and the medicine is analyzed in combination with spatial mutual information entropy. A space-spectrum combined encoder is innovatively developed, multi-modal features are fused by using a cross attention mechanism, and quantitative indexes such as a pore filling rate, crystallinity change and an interaction stability score are output. According to the method, the static and single-mode limitation of a traditional method is broken through, and a complete screening system from microstructure analysis to release dynamics prediction is established.

The present application relates to a multi-center mental image data enhancement method and system based on federated learning, receiving image data of patient mental image and modal identifier of image data, converting to uppercase form and matching; performing wavelet denoising processing on CT image for CT modal, performing non-local mean denoising processing on MRI image for MRI modal, and returning denoised image; parameter initialization, training sample number adaptive adjustment, noise level adaptive adjustment and model consistency adaptive adjustment are performed on the federated oriented enhancer, and the federated oriented enhancer is trained; based on the final parameter combination, the trained federated oriented enhancer is used for image self-adaptation of CT image or MRI image. Using adaptive enhancement selector, traditional deformation and pathological simulation of conditional generative adversarial network are fused to generate samples reflecting real diversity; a privacy-performance linkage mechanism is designed; a light security communication layer is constructed, which reduces the communication overhead while effectively defending against inversion attacks.