93 results about "Image gradient" patented technology

The invention discloses a reversible information hiding method and system for enhancing image smoothness, and relates to the technical field of information security, and the system comprises a preprocessing module, an information embedding module and an optimization recovery module. According to the method, the Sobel operator is utilized to calculate the image gradient map, the region is divided based on the threshold T, the smooth region and the texture region can be accurately distinguished, an accurate basis is provided for differential processing of different regions, region characteristics are fitted, large-size blocks and sub-blocks of the smooth region are divided to facilitate subsequent refinement processing based on the gradient mean value, and the processing efficiency is improved. The small-size blocks of the texture area are beneficial to operation aiming at texture features, the rationality and effectiveness of overall processing are improved, meanwhile, the pixel mean value in the large-size blocks of the smooth area is calculated, the pixel and mean value difference value is recorded, reversible low-pass filtering is carried out, and the reversibility of the image in the whole information hiding and recovering process is ensured.

The invention relates to the technical field of defect detection, in particular to a corrugated paper defect detection method and system based on machine vision. The method comprises the following steps: acquiring a corrugated paper image and converting the corrugated paper image into a frequency domain; identifying a periodic texture peak point based on amplitude spectrum angle energy distribution, generating a texture suppression filter, and combining with a multi-scale Gaussian high-pass filter to obtain a composite filter bank; the filter bank is used for filtering the frequency domain signals and performing energy weighted fusion; performing inverse transformation on the fused signal to obtain an enhanced image; and segmenting the enhanced image to obtain candidate defects, and verifying by combining the gradient information of the original image to determine a final defect. According to the scheme, the texture background can be adaptively inhibited, the multi-scale defect is enhanced, the false alarm rate is effectively reduced in combination with gradient check, and the detection accuracy and robustness are improved.

The invention discloses a visual calibration method, a terminal and a computer storage medium. The method comprises the following steps: generating theoretical grid coordinates; controlling a motion platform carrying a visual imaging unit to move to each target moving position in sequence, wherein the positions are obtained by superposing random two-dimensional offsets in a preset range; at each target moving position, a laser galvanometer and a visual imaging unit are synchronously triggered through the same hardware trigger source, and single-point mark engraving and image acquisition are executed; calculating a weighted centroid of the mark based on an image gradient magnitude, and extracting a sub-pixel center coordinate; combining the visual internal reference, the fixed coordinate transformation relation and the geometric constraint of the bearing plane, and resolving actually measured physical coordinates; carrying out space alignment, deviation calculation and invalid point interpolation on theoretical and actually measured coordinates, and generating a distortion data file which can be identified by a controller; the method further comprises a retry and fault-tolerant processing mechanism when coordinate calculation fails. According to the scheme, the calibration precision, the robustness and the automation level are effectively improved.

The invention relates to the technical field of image recognition control, in particular to a linear motion module accurate control method and system based on AI vision, and the method comprises the following steps: obtaining an interference fringe image and extracting a gray path, analyzing central point displacement to generate a coordinate sequence, repairing an interference feature position, and aligning the image and a control time sequence to generate a synchronization time table. The mapping structure response forms a set of control output configurations. According to the method, a peak path is constructed through interference fringe gray scale information in an image frame, central point pixel displacement is extracted, a coordinate migration sequence is constructed, fine-grained capture and tracking of an execution state are realized, an external interference area is identified according to image gradient change, vector repair is carried out, and image data stability and feature reliability are guaranteed. And aligning the repaired image response with a control instruction time sequence, constructing an impulse response lag statistical mechanism, realizing accurate mapping adjustment of the control frequency, and improving the control precision and response consistency in high-frequency displacement execution.

The invention discloses an image binarization method based on a dynamic window. The method comprises the steps of 1, target image noise suppression and gray level conversion; 2, carrying out compression processing on the dynamic range of the bright part of the image through a logarithmic function; 3, calculating a global threshold value of the grayscale image by utilizing a maximum between-class variance method; 4, dividing the grayscale image into image blocks according to the image gradient, and calculating a local threshold value for each block by using a Sauvola algorithm; 5, according to the global threshold value and the local threshold value of each image block, the threshold value of each image block is obtained through calculation according to the weight proportion; and 6, segmenting each image block by using a corresponding threshold value to obtain a final segmented image. According to the invention, the size of a local window is adaptively changed when any pixel point is binarized; and the image binarization capability under the condition of non-uniform illumination in a complex illumination environment is remarkably improved.

The present application belongs to the technical field of mobile robot navigation, and particularly relates to a SLAM navigation method and system for a mobile intelligent cabinet, which comprises the following steps: acquiring a laser point cloud sequence and a grayscale image of the mobile intelligent cabinet and performing data cleaning; obtaining a geometric feature index based on the spatial jump distribution of the laser point cloud sequence; obtaining a visual texture index based on the local dispersion of the image gradient; calculating laser dynamic weight and visual dynamic weight by using the geometric feature index and the visual texture index, weighting and fusing the pose change quantity calculated by the single-line laser radar odometry and the visual odometry to obtain a fused pose quantity; and updating the global state based on the fused pose quantity and driving autonomous navigation. The present application can adjust the sensor weight in real time according to the environmental characteristics, solves the problem of positioning divergence in the long corridor of a shopping mall or a high-reflectivity environment, and improves the robustness of navigation.

The invention discloses a multi-angle template matching method based on gradient topology sparse coding. The multi-angle template matching method comprises the following steps of S1, extracting a gradient topology skeleton of a template image; s2, constructing and coding a skeleton sparse dictionary; s3, coefficient domain rotation mapping construction is carried out; s4, performing multi-angle matching and similarity calculation; and S5, result fusion and optimization. The invention relates to the technical field of computer vision and image processing, and has the beneficial effects that the features based on the gradient topology skeleton are insensitive to illumination variation, gray shift and noise interference can be effectively overcome, and the matching stability in a complex environment is improved. By constructing the coefficient domain rotation mapping relation, physical rotation reconstruction does not need to be carried out on the image, the calculation overhead of multi-angle matching is remarkably reduced, the search efficiency is improved, sparse coding is used for carrying out compact representation on the image structure, and the adaptive capacity to local deformation and partial shielding is enhanced.

The invention discloses an endoscope video frame specular reflection highlight detection method and system, and belongs to the field of endoscope optical detection. The method comprises the following steps: firstly, converting an endoscope video frame into a grey-scale map; a Sobel operator is utilized to extract edge gradient to generate a binary image gradient mask, and meanwhile, a binary image highlight mask is generated by obtaining an area with the brightness obviously higher than the average level of a grey-scale map; dividing the binary image gradient mask and the binary image highlight mask into image blocks, adaptively adjusting the sizes of the blocks according to the highlight ratio of the image blocks so as to fuse the binary image gradient mask and the binary image highlight mask, and performing morphological processing and region screening to obtain a final binary image highlight mask of the grey-scale image; and carrying out time domain compensation on the final binary image highlight mask to obtain a specular reflection highlight detection result of the endoscope video frame. The accuracy, integrity and robustness of specular reflection highlight detection of the endoscope video frame are effectively improved.

The application relates to the technical field of image recognition, and discloses a hydraulic plunger pump intelligent fault diagnosis method and system based on pressure signals, which comprises the following steps: arranging pressure signals of a hydraulic plunger pump according to a spiral structure, and performing interpolation processing on the arranged signals to obtain an optimized pressure image; performing visual enhancement processing on the optimized pressure image to obtain an enhanced pressure image; analyzing multi-scale features in the enhanced pressure image to obtain global feature parameters of image brightness statistical features; performing local texture feature analysis on the enhanced pressure image to obtain local feature parameters of image gradient changes; taking the global feature parameters as an upper layer framework and taking the local feature parameters as a bottom layer benchmark to construct overall feature parameters of a running state; comparing the overall feature parameters with preset standard feature threshold values in multiple levels, and determining a fault state according to a comparison result; and the application can improve the efficiency of the hydraulic plunger pump intelligent fault diagnosis based on the pressure signals.

The application discloses a kind of intelligent image compression fast code rate control methods based on block level.The method includes the following steps: S1, the input picture is divided into non-overlapping same size block;S2, sample part image block;S3, the image block of twice coding and decoding is decoded to the sample, according to the relationship between the coding and decoding result and the quality factor λ of image block rate R, and the relationship between the quality factor λ of image block distortion D;S4, the average image gradient of each image block is calculated;S5, using linear relationship fitting each sampled image block average image gradient and the coefficient of R-λ relationship and D-λ relationship;S6, the R-λ relationship and D-λ relationship of not being sampled block are predicted;S7, according to the R-λ relationship and D-λ relationship of each block, the code rate control of image is carried out.The application compared with other code rate control methods, has more accurate code rate control precision, faster speed and lower memory consumption.

Images are reconstructed from data acquired using an ultra-fast-pitch acquisition with a CT system. As an example, an ultra-fast-pitch acquisition mode in single-source helical CT (≥1.5) can be used to acquire data. A trained machine learning algorithm, such as a neural network, is used to reconstruct images in which artifacts associated with insufficient data acquired in the ultra-fast-pitch mode are reduced. An example neural network can include customized functional modules using both local and non-local operators, as well as the z-coordinate of each image, to effectively suppress the location- and structure-dependent artifacts induced by the ultra-fast-pitch mode. The machine learning algorithm can be trained using a customized loss function that involves image-gradient-correlation loss and feature reconstruction loss.

The application discloses a micro-vision threshold segmentation algorithm based on OTSU improvement for micro-nano object images, and comprises the following steps: S1, combining image gradient information, calculating a new gray weight factor based on an exponential function and a logarithmic function, and performing bilateral filtering processing on the image; S2, calculating a traditional gray threshold of the filtered image by using a traditional inter-class variance formula, and calculating an improved gray threshold by using an improved inter-class variance formula based on a natural constant function; and S3, performing adaptive iterative operation on pixels of the image by using the traditional gray threshold and the improved gray threshold, so that three-region threshold segmentation is completed. The micro-vision threshold segmentation algorithm based on OTSU improvement disclosed by the application overcomes the problem that the calculation result is inaccurate due to the imbalance of the pixel quantity ratio during threshold segmentation, and improves the limitation of poor local analysis capability of the global threshold segmentation algorithm.

This invention belongs to the field of image generation, specifically relating to a method and system for generating 360° images of subway vehicles based on image stitching. The method includes: acquiring source images collected along the subway vehicle; obtaining effective matching point pairs through feature extraction matching and spatial topological similarity filtering; selecting the camera pair with the strongest geometric constraints as a benchmark; using incremental triangulation to recover the camera pose and 3D point cloud; in global bundle adjustment, optimization is performed by combining reprojection error and linear regularization terms based on the geometric principal axes of principal component analysis, with the regularization term weights adjusted according to the average distance from the point cloud to the principal axes; after optimization, the source images are projected onto a cylindrical canvas, and the optimal image gradient is selected based on gradient information entropy to construct a gradient field; a seamless 360° panoramic image is generated by solving the Poisson equation. This invention can effectively eliminate mismatches, suppress 3D reconstruction offset, and generate high-quality seamless panoramic images.

The invention belongs to the technical field of unmanned aerial vehicle visual guidance, and particularly relates to a visual guidance autonomous docking method and system for an intelligent cabinet by an unmanned aerial vehicle, and the method comprises the steps: obtaining a motion blur index according to the distribution characteristics of image gradient amplitudes; determining an adaptive sharpening coefficient based on the motion blur index, extracting a high-frequency detail layer and performing weighted enhancement to obtain an enhanced docking image; extracting a contour feature and a texture feature in the enhanced docking image, and obtaining a confidence coefficient weight of the texture feature according to a scale proportion factor of the contour feature; and calculating poses according to the contour features and the texture features, performing weighted fusion on translation vectors and rotation vectors of the two groups of poses by using a confidence coefficient weight, and performing unmanned aerial vehicle docking according to the fused poses. Through adaptive sharpening and feature fusion, the problem of inaccurate positioning caused by motion blur and scale change is effectively solved, and smooth docking of the unmanned aerial vehicle is realized.

This invention relates to the field of thread inspection technology, and discloses a thread inspection method, apparatus, system, and device based on image recognition and deep learning (DL). The method includes: acquiring thread images from multiple angles; filtering the thread images by combining image gradients and information entropy; measuring the thread's inner and outer diameters, thread pitch, and thread angle using image recognition methods based on the filtered thread images; detecting whether there are damage or folding defects in the thread teeth; detecting whether there are crack defects in the thread as a whole; and using a deep learning model combining classification and segmentation networks to detect whether there are crack defects at the top, waist, and bottom of the thread; comparing the measured thread inner and outer diameters, thread pitch, and thread angle with standard specifications; and determining whether the thread is qualified based on the defect detection results of the thread teeth, the thread as a whole, and the top, waist, and bottom of the thread. This invention can reduce inspection costs and improve inspection efficiency and accuracy while achieving overall thread inspection.

The present application relates to the technical field of medical image analysis, in particular to a burn area automatic evaluation method based on AI image recognition, comprising the following steps: analyzing image gradient sequence, screening for inflection points and encoding texture order, analyzing brightness difference to obtain attenuation trend to generate degradation structure, comparing tension mutation and convergence direction, screening for bridging points, identifying color difference density inversion to construct boundary point set, analyzing curvature difference to perform area expansion correction, in the present application, the texture order identifier is constructed by analyzing the pixel gradient sequence in the original image, the regional damage intensity judgment mechanism is established by combining the brightness change and the thermal damage signal difference, the wound edge point structure is derived by the tension direction trend, the continuous boundary range is identified by combining the color difference density, the area expansion comparison and correction are carried out by fusing the curvature difference, the systematic evaluation logic from boundary recognition to area mapping is established, the damage range is accurately delimited and the body surface characteristics are matched, and the precision evaluation and regional expansion of the burn area are realized.

The present application relates to lightning arrester intelligent monitoring technical field, specifically to a kind of lightning arrester lightning fracture risk intelligent monitoring system, system includes collecting lightning arrester tension electric field displacement disturbance data and rearranging section according to time, screening multi-source disturbance node and locating paragraph position, according to node number matching unmanned aerial vehicle continuous navigation section, extract corresponding image analysis brightness gradient and texture change, combine image and paragraph number cross relationship output lightning fracture risk section number.The present application is through the time series reorganization of tension, electric field, displacement disturbance data and paragraph number, constructs continuous disturbance structure, selects trigger node by the relationship of multiple physical quantity change, matches planning path according to node paragraph, compresses the range of inspection, identifies fracture suspect point by image gradient and texture change, combines the cross relationship of coordinate and outputs risk number section, enhances the coverage and traceability of risk identification.

The invention discloses an optical index testing method and device, wearable display equipment and a storage medium, and the method comprises the steps: carrying out the shooting of a test graphic card at different visual angles, and obtaining at least two to-be-spliced images; aligning the images to be spliced to form an overlapping region; the image gradual change direction of each repeated image in the overlapping area is determined, and the image gradual change direction is the direction from the to-be-spliced image side from which the repeated image comes to the other adjacent to-be-spliced image side; the pixel values of the corresponding repeated images are gradually reduced in the overlapping area according to the image gradual change direction, the pixel values of the repeated images are overlapped, a spliced image is generated, and the overlapped pixel values of the repeated images in the overlapping area are consistent with the pixel values of the to-be-spliced images; the optical index is tested according to the spliced image, so that an obvious boundary trace is prevented from appearing at the splicing line of the overlapped region, and the accuracy of the optical index test is effectively improved.

The application discloses a blood oxygen distribution imaging method and system based on double-band diffuse reflection space mapping, relates to the field of non-contact physiological parameter detection for biomedical optical application, and comprises the following steps: acquiring a double-band diffuse reflection original image sequence of a to-be-detected tissue region under two specific bands, obtaining normalized double-band images and a region of interest through preprocessing and region segmentation; estimating an effective optical path ratio between the two specific bands based on image gradient information in a local neighborhood of each pixel, and constructing a local optical path correction matrix; according to a double-band reflection attenuation ratio, a corresponding effective optical path ratio, and extinction coefficients of oxygenated hemoglobin and deoxygenated hemoglobin under the two specific bands, a double-band inversion model based on an improved Lambert-Beer law is used to calculate a preliminary blood oxygen saturation, and a two-dimensional blood oxygen saturation distribution map is generated through spatial regularization optimization. Therefore, the spatial representation capability, inversion stability and imaging accuracy of non-contact blood oxygen imaging can be improved.

The invention relates to the technical field of pipeline electromagnetic detection, in particular to a pipeline inner wall detection method and system based on a variable-diameter pipeline robot, and the method comprises the steps: synchronously collecting a pipeline inner wall image and a magnetic flux leakage signal through a camera and a magnetic flux leakage sensor carried by the variable-diameter pipeline robot, and preliminarily determining a first contour of a defect area through the magnetic flux leakage signal; then searching a most matched target contour in an image domain based on the contour, and performing adaptive filtering on the magnetic flux leakage signal by using image gradient information in the target contour; and iteratively optimizing edge detection parameters to enable the correlation between the second contour determined by the filtered magnetic flux leakage signal and the image contour to be optimal. According to the method, vision and magnetic flux leakage data are fused, the limitation of single sensor detection is overcome, the accuracy and robustness of defect contour recognition are remarkably improved, and more accurate and reliable detection of the defects of the inner wall of the pipeline is achieved.

The application discloses an image preprocessing method and system for laser light imaging, and relates to the technical field of image processing. The method comprises the following steps: under consistent laser light and shooting conditions, collecting multiple images of a target area; generating a reference background image according to the multiple images by using a preset algorithm; determining a corresponding difference image according to the reference background image, generating a guide image based on image gradient amplitudes, fusing the guide image into the difference image after normalization to obtain a target difference image; performing two-dimensional Fourier transform on the target difference image to obtain a frequency spectrum image; constructing a two-dimensional mask and performing inverse Fourier transform on the frequency spectrum image after the two-dimensional mask is applied to the frequency spectrum image to obtain a target image. The application solves the problem of poor image quality in the prior art.

The invention discloses a notebook computer shell injection molding part size monitoring method and system based on image vision. According to the method, a unified world coordinate system is established through multi-camera joint calibration, multi-view images of injection molding parts are synchronously collected, and key measurement feature points are extracted; an error compensation model is introduced in the three-dimensional coordinate reconstruction process, the model calculates a weight coefficient according to the distance from a measurement point to the optical center of a camera, the visual angle deviation degree and the image gradient intensity, weighted fusion is carried out on multi-visual-angle observation results, and corrected high-precision three-dimensional coordinates are obtained; and calculating key dimensions such as aperture, hole pitch, side length and the like based on the corrected coordinates, and comparing the key dimensions with standard values to judge the product quality. The system comprises an image acquisition device, a calibration module, a feature extraction module, a coordinate conversion module, a size calculation module and a quality judgment module. The size measurement precision, efficiency and stability of the notebook computer shell injection molding part are effectively improved, and the high-precision online detection requirement of an automatic production line is met.

The application discloses an industrial defect detection method based on causal feature enhancement, relates to the technical field of industrial visual detection, and constructs a structural causal model of industrial defect formation and obtains a causal diagram; collects an industrial component surface image and performs pretreatment to obtain an input image; the input image is input into a backbone network of a target detection network to extract multi-scale backbone features; a causal attention map is calculated according to causal variables, i.e., environmental variables and image gradient information; the backbone features are subjected to causal feature enhancement based on the causal attention map to obtain causal enhanced features; the causal enhanced features are input into a neck network and a detection head of the target detection network, and a defect bounding box, a defect category and a confidence are output. Through the introduction of the causal reasoning and feature enhancement mechanism, the application realizes high-precision and high-robustness industrial defect detection.

The invention discloses an enhanced imaging method for a microwave ESM sparse scanning system, and the method comprises the steps: obtaining the amplitude and phase information of a part of radiation sources on a far-field sampling plane, and obtaining sparse sampling data; on the basis, an image reconstruction problem is converted into a sparse optimization problem on a gradient domain by using total variation regularization, the gradient of the image in a smooth region approaches to zero by applying gradient constraint, and a relatively large gradient value is only reserved at a significant feature position. And an alternating direction multiplier method is further adopted to carry out iterative solution on the total variation minimization problem, and complete amplitude and phase information is gradually reconstructed by decomposing and optimizing a target and alternately updating variables. And dynamically adjusting the sampling rate according to the SSIM difference of the reconstructed image between adjacent iterations, and stopping sampling in advance when the SSIM value is higher than the quality threshold and the change tends to be stable. And finally, transforming the reconstructed amplitude and phase information through a far and near field inversion algorithm to obtain a near field diffraction pattern of the radiation source, and completing high-precision positioning and imaging of the radiation source.

The invention provides a wood board surface defect detection method and system based on machine vision and a medium, and the method comprises the steps: controlling the angle of a light source array based on a light source control instruction, and obtaining board images of multiple angles; carrying out fusion processing on the plate images of the multiple angles based on a maximum value fusion method, calculating an image gradient, generating a gradient tensor of the image, and carrying out eigenvalue decomposition on the gradient tensor to obtain texture direction distribution information; performing region segmentation on the enhanced image, constructing a texture prediction model, and generating a model prediction result of each sub-region based on the texture prediction model; analyzing a model prediction difference value, and generating a residual response diagram; analyzing the residual response value of each sub-region to obtain surface defect information of the plate; according to the method, the plate images at multiple angles are collected, fusion processing and eigenvalue decomposition are carried out, so that texture direction distribution is accurately analyzed, defects on the surface of the plate are accurately analyzed according to the residual response diagram, and the detection precision is improved.

The invention belongs to the technical field of displacement field monitoring, and particularly relates to a dropper wind-induced vibration displacement field monitoring method and system based on an optical flow method.The dropper wind-induced vibration displacement field monitoring method comprises the following steps that S1, local direction parameters of multi-scale Gabor wavelet transformation are set based on the dropper global geometric morphology at the previous moment, and texture response of pixels of a current frame is extracted; a weighted evaluation function combining an image gradient covariance matrix and texture response is constructed, and sparse feature points are determined by solving feature values of the weighted evaluation function; and S2, establishing a Kalman filtering model of which the state vector contains the position and the speed, and predicting a target search region of the sparse feature points in the current frame based on the updating state of the previous frame by using the Kalman filtering model. According to the invention, by carrying out frequency domain filtering processing on the displacement time-history signal, main vibration components can be extracted from measurement data, and full-field monitoring of a dropper wind-induced vibration displacement field is realized.

This invention provides a method for generating a seabed DEM based on large-scale nautical chart depth points and high-resolution remote sensing surfaces, belonging to the field of marine surveying technology. The method includes: acquiring discrete depth points from remote sensing images and nautical charts; extracting a relative depth trend surface based on a dual-band logarithmic ratio model; calculating the image gradient field and pixel-level structure tensor, solving for the principal eigenvectors, and constructing an anisotropic diffusion tensor by combining local gradient magnitudes; using the nautical chart depth points to perform regression calibration on the trend surface to generate sparse residuals, and establishing a residual calculation grid; using the anisotropic diffusion tensor as a guiding constraint, performing non-uniform diffusion calculation on the sparse residuals to generate a global continuous residual correction surface; superimposing the correction surface onto the trend surface, and obtaining the final seabed DEM through anomaly removal and edge smoothing. This invention solves the problems of low accuracy and loss of detail in complex shallow sea topography inversion.

Embodiments of the present application disclose a kind of image gradient determination method, device, equipment and storage medium.The method comprises: obtaining the pixel value of each pixel point in digital image that has been loaded into the global memory of graphics processor;According to the storage capacity of shared memory in graphics processor, determine target pixel value from each pixel value, load each target pixel value to the shared memory;Read each target pixel value from shared memory, and determine the image gradient of digital image on the pixel position of each target pixel value based on the reading result respectively.The technical scheme of the embodiment of the present application guarantees the locality of target pixel value in two-dimensional space by loading target pixel value in digital image to shared memory with two-dimensional physical structure, which can avoid cache miss in subsequent data reading process, thereby achieving the effect of improving the determination speed of image gradient in computer implementation process.

A surface optimization method for an optical lens or reflector. The method comprises: given a unit illumination direction, a target image and a receiving plane, using an original mesh to represent the surface of an optical lens or reflector to be optimized (1); on the basis of the alignment degree at each position, the variation degree of luminous flux, the smoothness of the original mesh and a barrier term, establishing an optimization problem to obtain a deformed mesh (3); recalculating in the deformed mesh the positions of all basic geometric units on the receiving plane, and performing superposition and rendering to obtain an image (4); on the basis of the image difference between a rendered image and the target image, an image gradient difference, a penalty term for exceeding an image boundary, the smoothness of the deformed mesh and the barrier term, establishing an optimization problem to obtain a twice-deformed mesh (5); and refining the mesh level by level until the last iteration at the maximum resolution terminates, so as to obtain a final optimized mesh (6). The method can optimize the surface shape of an optical lens or reflector on the basis of light source information, the position of a receiving plane, and a target image.

The application discloses a defect detection method based on physical rendering image adaptive restoration applied to the field of industrial production defect detection, comprising S100, an image parameter acquisition step, S200, an image restoration step and S300, a defect prediction step; S100 comprises: S110, acquiring an industrial field image; S120, calculating an image gradient and splicing the original image to obtain a multi-channel input set; S130, decomposing the multi-channel input set frequency domain feature and then reconstructing the frequency domain to obtain a reconstructed feature map; S140, inputting the reconstructed feature map into an image parameter prediction model to output an interference parameter; S200 comprises: S210, inputting the industrial field image into an image restoration model and performing defogging enhancement based on the interference parameter; S220, performing brightness enhancement on the defogging image to obtain a final restoration image J; S300: inputting the final restoration image J into a defect detection network to output defect recognition information. The application can perform online casting surface defect detection based on the real-time acquired industrial field production monitoring image of the ingot production system.