56 results about "Morphological gradient" patented technology

The invention belongs to the technical field of image processing, and particularly relates to a cord fabric production defect identification and analysis system based on machine vision, which comprises a multispectral imaging module, an image preprocessing module, a target detection module, a result verification module, a weight correction module and a control response module, the method comprises the following steps: acquiring a multispectral image sequence of each partition of a cord fabric forming surface, combining motion compensation, geometric calibration and noise filtering to improve preprocessing precision, and generating a composite spectral image by adopting adaptive weight fusion on the basis of considering an image signal-to-noise ratio and texture definition; and detecting defects based on frequency division domain texture and morphological gradient parameter bimodal matching, further verifying detection credibility through defect sensitive spectral image feature enhancement degree, dynamically correcting a fusion weight and performing iterative optimization, and finally triggering a production control instruction according to a detection result. And closed-loop management of high-precision defect detection and production control linkage of the cord fabric is realized.

The invention discloses a morphological gradient region replacement method based on SAM semantic segmentation and user guidance, and relates to the technical field of computer vision and image processing, and the method comprises the steps: carrying out the semantic segmentation of a to-be-processed image through an SAM model, extracting a multi-level semantic feature, carrying out the standardization and dimension reduction, extracting a causal factor based on independent component analysis, and carrying out the user guidance. A directed causal factor association graph is generated through Granger causal relationship test, and a causal attribution probability graph is generated through reverse mapping; constructing a structured causal graph, and generating a causal mask through a graph convolutional network; encoding the original interaction signal into a guide thermodynamic diagram; constructing a diffusion equation, forming a gradual change control equation by dynamically fusing and guiding the intensity distribution of the thermodynamic diagram and an image semantic diffusion item, and iteratively solving the gradual change control equation; generating an anisotropic morphological operation kernel according to the geometric curvature characteristics of each region in the replacement mask; and fusing the optimized replacement mask with the target content based on a gradient domain optimization algorithm to generate a gradient replacement image.

The invention discloses a production quality evaluation method for a gold bonding wire, and relates to the technical field of microelectronic packaging, and the method comprises the following steps: obtaining a surface image of the gold bonding wire through a high-resolution image collection system; performing preprocessing on the image data based on the acquired image to eliminate noise and enhance contrast; through the image data, extracting surface topography gradient features by using morphological operation, and quantifying microstructure changes of the surface of the gold wire; extracting texture energy characteristics through wavelet multi-scale decomposition, and analyzing distribution characteristics of surface textures from different scales; self-adaptive fusion is carried out on the morphological gradient features and the texture energy features, and the feature expression of the defect area is enhanced by dynamically adjusting the weight; dividing the boundary of a defect region by adopting an adaptive threshold segmentation method, and separating a normal region from an abnormal region; and extracting geometric and textural features of the defect region obtained by segmentation, realizing automatic discrimination of defect types through a classification model, and outputting a quality evaluation result.

The invention relates to the technical field of medical treatment, and particularly discloses an intelligent wound measuring and recording system and method.The method comprises the steps that multi-dimensional image data and three-dimensional depth information of a wound area are synchronously collected, hierarchical space division and nonlinear feature coding are adopted for preprocessing the multi-dimensional image data, and wound basic information is generated; constructing a wound measurement model based on the basic information of the wound, extracting wound features with space-time relevance, and extracting a wound contour of multi-dimensional image data by using an improved U-Net architecture in combination with an HSV color space and morphological gradient method; according to the method, the multi-dimensional image data and the three-dimensional depth information are synchronously acquired, and the advanced image processing technology is combined, so that the accurate extraction of the wound contour is realized, and the problem that wound details cannot be accurately identified and measured in a traditional method is solved; based on an ICP registration algorithm and a three-dimensional point cloud reconstruction technology, submillimeter-level wound volume measurement is realized, and the wound volume evaluation precision is improved.

The invention provides a multi-scale morphological gradient and NSST-PCNN image fusion method, and relates to the technical field of multi-source remote sensing image fusion. The method comprises the following steps: performing Gaussian curvature filtering decomposition and LEE filtering decomposition on an SAR image to obtain SAR decomposition features; decomposing the optical image by using non-subsampled shearlet transform (NSST) to obtain an NSST coefficient of the optical image; performing feature extraction on the SAR image after LEE filtering through various structural elements to obtain a multi-scale morphological gradient feature map; and inputting the SAR decomposition features and the NSST coefficient of the optical image into a pulse coupled neural network (PCNN) to obtain a preliminary fusion image, and reconstructing the preliminary fusion image through NSST inverse transformation to obtain a final fusion image. The fused image retains the spectral characteristics of the original optical image, enhances the spatial resolution and edge details, and is suitable for scenes needing to retain high-precision edges, such as road slope deformation monitoring.

The invention discloses a rice field insect pest detection method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the following steps: receiving the soil humidity and environment temperature of a rice field, and determining a temperature and humidity abnormal region of the rice field based on the soil humidity and environment temperature of the rice field; acquiring field image data in the temperature and humidity abnormal area of the rice field; processing the field image data in the temperature and humidity abnormal region of the rice field based on morphological closed operation and morphological gradient operation to obtain processed field image data, inputting the processed field image data into a pre-established YOLOv5s + target detection algorithm model, and outputting to obtain a pest and disease damage identification result; and performing image splicing generation based on the field image data in the temperature and humidity abnormal region of the rice field and the pest and disease identification result to obtain a disaster grading map.

The invention provides an unmanned aerial vehicle real-time flight path dynamic adjustment system and method based on artificial intelligence, equipment and a medium. The system comprises a data acquisition module for acquiring target data; the image processing module performs image preprocessing and enhancement processing on the target image data; the image segmentation module can extract image edge features in the enhanced image according to a multi-scale multi-direction morphological gradient algorithm, segment the enhanced image according to the image edge features, and merge segmented images according to the same marks; a scene recognition and classification module performs scene recognition on the processed segmented image according to a convolutional neural network model to obtain a scene type; and the path dynamic adjustment module dynamically adjusts the flight path. The scene type can be accurately judged by using an image segmentation technology, and the flight control parameters corresponding to the scene are matched, so that the accuracy of path dynamic adjustment is improved.

The invention relates to the technical field of vision, in particular to an industrial intelligent spraying tracing method, which comprises the following steps of: firstly, acquiring an original depth map of a curved surface workpiece by utilizing a depth camera, and obtaining original data containing depth information; next, edge feature enhancement preprocessing is carried out on the original depth image, edge pixels are accurately extracted through morphological gradient operation and a user-defined anisotropic diamond kernel, and an edge mask image is constructed; then, curvature region division and self-adaptive interpolation restoration are carried out, quadratic polynomial curved surface fitting is carried out on a hole pixel neighborhood, a local curvature value is calculated, a curved surface region is divided according to the curvature value, and corresponding interpolation strategies are adopted for different regions; afterwards, edge detail sharpening post-processing is carried out on the repaired depth image, and the edge contrast is enhanced by using a non-sharpening mask algorithm; and the repaired and sharpened depth map is converted into a three-dimensional point cloud, workpiece pose information is determined based on three-dimensional point cloud data, and spraying track planning is carried out.

The invention discloses a bolt surface defect detection method and system based on computer vision. The method comprises the steps that 360-degree image collection and preprocessing are conducted on a bolt through an annularly-arranged multi-camera array; fourier transform is adopted to analyze thread periodic texture features, and an adaptive filter is designed to enhance the image; multi-scale morphological operation is adopted, and morphological gradient features are extracted by using structural elements for crack, pit and scratch type defects; a self-adaptive threshold segmentation method based on local statistical characteristics is combined with thread direction information to dynamically adjust a segmentation threshold, and defect candidate areas are segmented; and extracting geometric and textural features of the candidate region, and identifying the type, position and severity of the defect by adopting a random forest classifier. According to the invention, high-precision automatic detection and classification of various defects on the surface of the bolt are realized.

The invention belongs to the technical field of image processing, and particularly relates to a multi-source data-driven flood storage and detention area ecological toughness detection method and system, and the method comprises the steps: 1, carrying out the geometric correction and radiation correction of an optical satellite remote sensing image, and forming a time sequence basic data set; 2, a first axis and a second axis of the ecological state matrix correspond to space grids of the flood storage and detention area respectively, and a third axis corresponds to time; 3, mapping the ecological state matrix to a fractal space to form a time coupling matrix; 4, for each time slice of the time coupling matrix, setting a neighborhood with a fixed size around each pixel, and performing time window segmentation and fractal complexity estimation on the multi-scale morphological gradient map; and step 5, calculating the toughness value of the connected region according to the area and the toughness values of all positions in the connected region. According to the method, the accuracy and practicability of ecological toughness detection are remarkably enhanced.

The invention relates to the technical field of image analysis, in particular to an expressway green passing vehicle image feature extraction and rapid screening system. The expressway green passing vehicle image feature extraction and rapid screening system comprises an image acquisition unit, an image extraction unit and an image screening unit. According to the method, the image acquisition, extraction and screening processes are automated, the detection efficiency is greatly improved, the vehicle waiting time is greatly shortened, smooth passing of the expressway is guaranteed, the method based on morphological gradient and distance transformation is adopted, the complex scene that cargos are disorderly stacked and mutually shielded in a green passing vehicle container of the expressway can be dealt with, and the traffic safety of the expressway is guaranteed. In the image fusion process, the adjacent sub-regions with similar features are merged based on the feature similarity, redundant information in the image is reduced, in addition, non-green passing standard goods can be screened out, and subjectivity and uncertainty of manual judgment are avoided.

The invention provides an industrial anomaly classification method and system based on a multi-modal large language model, and belongs to the technical field of visual detection. The visual pre-screening module adopts a BYOL unsupervised feature extractor and a local block matching technology to calculate the minimum Euclidean distance between an image to be detected and a normal sample so as to generate an abnormal thermodynamic diagram, and efficient preliminary screening is realized; the coordinate set is converted into a binary mask through an abnormal contour marking module, and an accurate single-pixel contour is generated through morphological gradient and SIFT displacement correction; fusing a CLIP visual encoder and a GPT text encoder in a multi-modal decision engine, realizing visual-semantic alignment through cross-modal fusion and a multi-head attention mechanism, and supporting alpha hybrid dynamic rule update; and a dynamic classification executor calculates an abnormal region risk index based on the risk grading model, and generates a grading response action and a structured report. According to the invention, the classification precision is improved, so that the method can adapt to complex and diversified industrial abnormal scenes.

The invention discloses a remote sensing image zero watermarking method based on a multichannel pulse coupling neural network. According to the method, firstly, all wave bands of a remote sensing image are input into a multi-channel pulse coupling neural network model for parallel processing, and the model is optimized and excited by adopting a multi-scale morphological gradient and a differential box dimension, so that the accuracy and robustness of feature extraction are improved. And then, the extracted features are encrypted by using asymmetric Tent mapping, and the watermark image and the encrypted feature image are scrambled according to generalized Arnold transformation, so that the security of the algorithm is improved. And finally, performing XOR operation on the scrambled watermark image and the feature image to generate a zero watermark, and storing the zero watermark in an interstellar file system (IPFS). The method has relatively high robustness on non-geometric watermark attacks and geometric watermark attacks of common remote sensing images, and provides reliable guarantee for security management and copyright protection of the remote sensing images.

The invention discloses a ceramic microstructure defect intelligent identification system based on morphological image analysis, and relates to the technical field of ceramic material quality detection, and the system comprises a multi-scale morphological feature extraction engine, a defect morphological feature analysis module, a deep semantic enhancement module and a defect intelligent identification decision module. The multi-scale morphological feature extraction engine adopts an adaptive rotation morphological operator group to carry out multi-direction morphological operation, a multi-scale morphological gradient pyramid is generated and morphological feature vectors are extracted, the defect morphological feature analysis module carries out feature coding and classification through a morphological feature encoder to generate defect morphological descriptors, and the defect morphological descriptors are subjected to feature extraction; and the deep semantic enhancement module extracts semantic features under the guidance of morphological feature coding, performs feature fusion, generates operator optimization parameters and feeds the operator optimization parameters back to the feature extraction engine to form closed-loop optimization, and the system realizes deep fusion of morphological features and deep semantic features and adaptive operator optimization.

The invention provides a visual detection method and system and an edge calculation box, and the method comprises the steps: carrying out the multi-scale edge detection of a surface image of a target product, and generating edge feature maps of different scales; determining a defect area candidate frame on the surface of the target product according to all the edge feature maps, and determining gradient variances of pixels in all the edge feature maps; determining the positioning loss of the surface defects of the target product according to the defect area candidate frame and all gradient variances, and determining the contour form gradient of each type of defects on the surface of the target product based on the positioning loss and the fluctuation characteristics of the light intensity around the production line; and determining the influence grade of the surface defect of the target product on the target product through the contour form gradient of each type of defect, and outputting a grading alarm signal of the surface defect of the target product based on the influence grade. By adopting the scheme of the invention, dynamic grading alarm can be carried out on the product based on the defect geometric topology characteristics of the product.

The invention discloses a muzzle arc energy and mass center spatio-temporal evolution analysis method and system, and relates to the technical field of muzzle arc analysis, and the method comprises the steps: building an electromagnetic track launch muzzle arc simulation experiment platform, and obtaining a muzzle arc image through a high-speed photographing system; an arc contour image is obtained based on a local adaptive threshold segmentation and morphological gradient extraction algorithm, and the integrity and robustness of arc contour extraction are improved; based on a robust centroid tracking algorithm of multi-connected domain energy fusion and time sequence consistency constraint, a final arc centroid is obtained, and stable positioning and tracking of the arc energy centroid in the full evolution stage are achieved; and quantitatively analyzing the time-domain evolution characteristics of the highlighted area of the muzzle arc and the law of the spatial motion trail of the mass center of the highlighted area based on the muzzle arc images at different muzzle speed. An effective means is provided for quantitative characterization of a bore port arc, and an experimental basis and a theoretical reference are provided for optimization of a bore port structure, anti-ablation design, regulation and control of emission working conditions and the like of an electromagnetic emission system.

The present invention discloses a method for automatically distinguishing the results of multi-directional and multi-scale new coronavirus antigen detection. The method constructs a priori structural elements, performs a priori flexible morphological operator filtering on the original image of the antigen test kit test result, calculates the a priori flexible morphological gradient, extracts the edge of the antigen test kit, extracts the test paper area to be distinguished, eliminates interference such as reflective areas, and locates potential areas in sequence to perform pre-detection of letter discrimination areas; performs discrimination letter modeling in the angle dimension to form a training set, combines samples into a tensor form to form a sample feature matrix; converts the RGB color image to the HSV color space, expands the letter image block to extract features, and constructs a feature matching operator to classify the letters C and T; performs color histogram statistics on the test paper area to be distinguished, calculates the color segmentation threshold, and detects the red discrimination mark; and splices the letter classification result and the red discrimination mark detection result to give the discrimination result of the antigen test.

The application discloses a bolt surface defect detection method and system based on computer vision, comprising: 360-degree image acquisition and preprocessing of the bolt through a multi-camera array arranged in a ring; Fourier transform is used to analyze periodic texture features of the thread, and an adaptive filter is designed to enhance the image; multi-scale morphological operation is used to extract morphological gradient features using a structure element for crack, pit and scratch type defects; an adaptive threshold segmentation method based on local statistical features is used to dynamically adjust the segmentation threshold combined with thread direction information to segment out the defect candidate region; geometric and texture features are extracted for the candidate region, and a random forest classifier is used to identify the defect type, position and severity. The application realizes high-precision automatic detection and classification of various defects on the bolt surface.

The embodiments of the present application relate to the field of image recognition. The present application discloses a method, system and electronic device for detecting the surface area of ​​a supercapacitor electrode. The method includes: performing multi-layer discrete wavelet decomposition on the collected image, and nonlinearly enhancing the high-frequency coefficients of each layer according to different decomposition levels; reconstructing and denoising the enhanced image using a soft threshold method, and the threshold parameter is determined according to the decomposition level and the standard deviation of the noise; constructing circular structural elements of multiple scales, corresponding to different aperture feature extractions, and using a Gaussian kernel to calculate the morphological gradient for each scale; using a local window to calculate statistical features and determine the threshold, and weighted fusion of features of multiple scales, with the weight coefficient decreasing as the aperture increases. Through the scheme of the present application, the recognizability of microporous and mesoporous structures in the image can be improved, and accurate non-destructive detection of the electrode surface area can be achieved.

The invention relates to a carbonate rock reservoir pore structure parameter quantitative analysis method based on machine learning. The method comprises the following steps: acquiring a microscopic image of a casting body slice of a carbonate rock sample; analyzing the microscopic image according to the morphological characteristics of pores in the carbonate rock, respectively labeling pore and non-pore areas by adopting an interactive method, and generating a training data set; the method comprises the following steps of: extracting multi-scale Gaussian intensity features, LBP texture features and morphological gradient features of a microscopic image, and constructing a high-dimensional feature space by adopting a hierarchical fusion mode; training a random forest classifier to obtain a pore probability distribution diagram of the microscopic image; carrying out segmentation and morphological optimization on the pore probability distribution diagram of the microscopic image, and outputting a binary pore classification result; and calculating the surface porosity, the pore size, the pore size distribution and the pore connectivity index based on the binarized pore classification result. The method is high in accuracy and good in applicability; processing is very convenient, and manpower and material resource consumption is low.

The application relates to the technical field of image processing, in particular to a chemical intermediate purity detection method and system based on chromatographic features. The method comprises the following steps: performing baseline correction on a chromatographic detection signal of a chemical intermediate sample to obtain a one-dimensional discrete time sequence, converting the one-dimensional discrete time sequence into a gray-scale chromatographic image, determining a plurality of regions of interest from a morphological gradient image of the gray-scale chromatographic image, respectively determining scale energy attenuation rates and horizontal skewness of the regions of interest, determining confidence index of the regions of interest based on the scale energy attenuation rates and the horizontal skewness, removing the regions of interest with the confidence index lower than a preset threshold, and determining the purity of the chemical intermediate sample based on integral areas of the one-dimensional discrete time sequences corresponding to the retained regions of interest. Through the technical scheme, a more accurate detection result of the purity of the chemical intermediate sample can be obtained.

The invention discloses a prostate cancer image classification method based on axial consistency enhancement and threshold guide selection, and the method comprises the steps: designing an axial consistency context enhancement module ACE, combining strip pooling with axial self-attention, and enhancing the correlation expression between different spatial positions in a feature map through explicit modeling cross-row / cross-column long-distance structural dependence; a threshold guide semantic selection module TGSS is proposed to fuse horizontal, vertical and channel three-view weight information, differential processing of foreground focus enhancement and background tissue reconstruction is realized through hard threshold segmentation, and morphological gradient modeling is performed on a boundary transition region. Compared with the prior art, the method can further improve the precision of the algorithm model on related tasks of prostate cancer image classification.

The invention discloses an unmanned aerial vehicle light show three-dimensional image generation method based on dynamic path planning, and relates to the related technical field of image generation, and the method comprises the steps: constructing a dynamic accessibility tensor space which uniformly describes the space accessibility, the time propulsion and the light trace stability; according to the three-dimensional target morphological gradient, dividing an unmanned aerial vehicle group into three subgroups of configuration generation, light trace enhancement and morphological stability, and setting different path priorities, morphological sensitivities and energy consumption thresholds; and applying a three-dimensional image growth operator to the path field in the dynamic reachability tensor space to form a light trace density field, and performing double-coupling iterative analysis to generate a dynamic path planning result. The technical problem that unmanned aerial vehicle light show in the prior art is difficult to give consideration to space-time continuity, visual stability and cluster cooperation efficiency of three-dimensional light tracks is solved, and the technical effects of efficient, stable and high-quality three-dimensional light show image generation and improvement of planning efficiency, adaptability and three-dimensional image fidelity are achieved.

The application discloses a low-altitude target edge detection method based on PCNN and improved morphology, and belongs to the technical field of aerospace detection. The method is as follows: an image containing a low-altitude target is obtained by a photoelectric load on a low-altitude platform, and a filtering algorithm is used to pre-process the image; a PCNN model is used to perform image segmentation on the filtered image, and a region containing the low-altitude target is screened out; the region is extracted to obtain an image to be detected, and the influence of details on edge detection is eliminated; an improved morphological gradient is used to perform edge detection on the obtained image to be detected; the improved morphological gradient is used to suppress random noise, so that the precision of low-altitude target edge detection and the anti-noise capability are improved; a morphological thinning operator is used to thin the obtained low-altitude target edge, the connectivity of small parts of the image is maintained, and the precision of low-altitude target edge detection is improved. The application has the advantages of high detection precision and strong anti-interference capability.

The invention relates to the technical field of image processing, in particular to a chemical intermediate purity detection method and system based on chromatographic characteristics. The method comprises the following steps: performing baseline correction on a chromatographic detection signal of a chemical intermediate sample to obtain a one-dimensional discrete time sequence, converting the one-dimensional discrete time sequence into a gray-scale chromatographic image, and determining a plurality of regions of interest from a morphological gradient map of the gray-scale chromatographic image; respectively determining the scale energy attenuation rate and the horizontal skewness of the region of interest; determining a confidence index of the region of interest based on the scale energy attenuation rate and the horizontal skewness; and removing the region of interest with the confidence index lower than a preset threshold value, and determining the purity of the chemical intermediate sample based on the integral area of the one-dimensional discrete time sequence corresponding to the reserved region of interest. According to the technical scheme, a more accurate detection result of the purity of the chemical intermediate sample can be obtained.

The application provides a visual detection method, system and edge computing box. The surface image of a target product is subjected to multi-scale edge detection to generate edge feature maps of different scales. A defect area candidate frame of the surface of the target product is determined according to all the edge feature maps, and the gradient variance of pixels in each edge feature map is determined. The positioning loss of the surface defects of the target product is determined according to the defect area candidate frame and all the gradient variances, and the contour shape gradient of each type of defect on the surface of the target product is determined based on the positioning loss and the fluctuation characteristics of the light intensity around the production line. The influence level of the surface defects of the target product on the target product is determined through the contour shape gradient of each type of defect, and a graded alarm signal of the surface defects of the target product is output based on the influence level. The application can realize dynamic grading alarm of products based on the geometric topological characteristics of product defects.

The present invention discloses a method, medium, device and apparatus for extracting contour straight lines. The method includes: obtaining input information of a user and point cloud data to be processed, and determining an area of interest of the user according to the input information; preprocessing the point cloud data to be processed to obtain an initial binary image; performing morphological gradient processing on the initial binary image to obtain initial contour straight line data corresponding to the point cloud data to be processed; generating a final contour straight line based on the initial contour straight line data and the area of interest; and being capable of effectively extracting contour straight lines in the point cloud data when the quality of the point cloud data is poor, and improving the effect of contour straight line extraction.

The present application belongs to the technical field of image processing, and particularly relates to a multi-source data driven detection method and system for ecological resilience of a flood storage and detention area, which comprises the following steps: step 1: performing geometric correction and radiation correction on an optical satellite remote sensing image to form a time series basic data set; step 2: a first axis and a second axis of an ecological state matrix correspond to spatial grids of the flood storage and detention area respectively, and a third axis corresponds to time; step 3: mapping the ecological state matrix to a fractal space to form a time coupling matrix; step 4: for each time slice of the time coupling matrix, a fixed size neighborhood is set around each pixel, and a multi-scale morphological gradient graph is subjected to time window segmentation and fractal complexity estimation; and step 5: according to the area and the resilience values of all positions therein, the resilience value of the connected region is calculated. The present application significantly enhances the accuracy and practicability of ecological resilience detection.

The invention belongs to the technical field of asphalt pavement paving defect detection, and particularly discloses a video image-based asphalt pavement paving defect identification method, which comprises the following steps of: continuously acquiring video images of a paving working surface, respectively extracting an initial material area and a motion area, fusing the initial material area and the motion area, and combining morphological optimization and edge contour positioning to identify an asphalt pavement paving defect. According to the method, the paving operation area in the current frame is locked, the boundary positioning precision of the operation area is remarkably improved, a reliable analysis domain is provided for subsequent defect detection, and meanwhile, the paving operation area is identified by constructing a gridding local feature field based on image apparent features in the identified paving operation area and analyzing the consistency of grid unit features in the advancing direction of the paver. Therefore, a suspected defect area is identified, the suspected defect area is subjected to spatial position verification and morphological gradient feature verification, a real defect area is screened out, non-defect visual interference can be inhibited to the maximum extent, and the credibility of defect detection is improved.

The invention relates to the field of machine vision, in particular to a building roof defect detection method and system based on an unmanned aerial vehicle, and the method comprises the steps: obtaining a building roof gray image collected by the unmanned aerial vehicle; calculating morphological gradient maps in multiple directions of the roof grayscale image; calculating the variance of each pixel point in the gradient values in all directions, generating a gradient direction consistency graph, and carrying out the weighted fusion of the morphological gradient maps in all directions based on the gradient direction consistency graph, and generating an edge saliency map; in the neighborhood of each pixel point of the roof grayscale image, according to the Euclidean distance between a neighborhood pixel and a central pixel, performing attenuation weighting on the occurrence frequency of a grayscale pair, and based on the weighted symbiotic probability, calculating a spatial position sensitive entropy, and generating a texture complexity map; further obtaining a defect response diagram; calculating a local discrimination threshold value corresponding to the pixel point; and when the value of the pixel point in the defect response diagram is greater than the corresponding local discrimination threshold, determining that the pixel point is a defect point.