56 results about "Gradient operators" patented technology

The invention relates to the technical field of geophysical exploration, in particular to a multi-scale three-dimensional gravitational-seismic joint frequency domain inversion method based on hybrid constraints. The method comprises the following steps: determining observation gravity anomaly on the ground by adopting discrete Fourier transform, and discretizing the observation gravity anomaly by adopting a method of weighted average of values at a plurality of Gaussian points; calculating spherical harmonics and spherical harmonic coefficients of two-dimensional density distribution of each layer in the inversion range, and calculating gravity anomalies on the observation surface based on the spherical harmonics and the spherical harmonic coefficients; converting the inversion range data into a three-dimensional gravitational field forward kernel matrix; constructing a three-dimensional gravitational field hybrid constraint inversion objective function; inputting pre-collected earthquake three-dimensional velocity model data, and converting the earthquake three-dimensional velocity model data into a reference density model; and calculating a gradient operator in the inversion objective function, simplifying the inversion objective function based on the gradient operator obtained through calculation, and then solving the simplified inversion objective function by using an alternating direction multiplier method.

The invention relates to a cloth flaw detection method, device and equipment and a storage medium, and the method comprises the following steps: carrying out the multi-angle light source irradiation of the surface of cloth, obtaining a texture image, and carrying out the grid segmentation of the texture image into detection units; and extracting a gray value sequence of each unit, calculating an edge difference by using a gradient operator, generating a texture edge change curve to identify an abnormal wave crest, and determining a defect candidate region. And carrying out local amplification imaging based on the candidate region coordinates, and extracting defect shape feature parameters. The defect types are determined according to the parameters, quality grade division is performed on the cloth according to the parameters, and a quality grade report is finally generated, so that the technical problems of low defect identification precision and easy missing detection of fine defects caused by uneven illumination and texture interference in the existing cloth defect detection method are solved.

The invention discloses an ancient building repair effect evaluation method based on image enhancement. The method comprises the steps of image acquisition, image enhancement, ancient building repair effect evaluation model establishment and ancient building repair effect evaluation. The invention belongs to the field of image processing, and particularly relates to an ancient building repair effect evaluation method based on image enhancement, which is characterized in that an evaluation area is positioned based on block average gray scale, detail reinforcement is performed on the premise of retaining a structure, and local defects are highlighted to the maximum extent through priority ranking and matching block measurement definition while the consistency of the overall contour is guaranteed; and the final repair evaluation effect is improved. The method comprises the following steps: defining a repair activation function, introducing a channel adaptive threshold, carrying out adaptive noise suppression and significant texture reservation, introducing a fractional gradient operator, considering local smoothness and long-range dependence, and amplifying gradient response of small cracks and texture edges; and the repair effect evaluation accuracy is improved.

The invention discloses a background schlieren method flow field flow velocity measurement method based on pseudo-schlieren images. The method comprises the following steps: constructing a background schlieren experiment measurement system; obtaining a speckle plate image sequence without a flow field background; obtaining a background distortion image sequence; obtaining a pseudo schlieren image sequence; and obtaining the final flow velocity of the flow field to be measured. The method has the advantages that a background plate distortion image sequence under the action of a flow field is collected through a high-speed camera, and pseudo-schlieren transformation processing is carried out through an improved mixed gradient operator and a multi-scale feature fusion technology; and a flow velocity field is solved in combination with a PIV-optical flow fusion algorithm, and real-time processing is realized through GPU parallel computing. The method breaks through the limitation that the traditional PIV technology depends on tracer particles, has the advantages of non-contact, full-field measurement, high temporal-spatial resolution and the like, and is particularly suitable for compressible flow field and turbulent flow field measurement. Experiments show that the method can realize high-resolution and high-frame-rate image processing, the uncertainty of flow velocity measurement is less than 3.2%, and a new technical means is provided for complex flow field diagnosis.

The invention discloses an image quality optimization method and system based on multi-scale feature decoupling and dynamic fusion, and the method comprises the steps: carrying out the multi-scale decomposition of an input degraded image, and extracting the feature components of illumination-color, texture-noise and edge-structure; illumination normalization and color fidelity enhancement are realized through illumination estimation and color space transformation; the base layer and the detail layer are separated through edge preserving filtering, and self-adaptive contrast enhancement and noise suppression are carried out on the detail layer; extracting edge information by using a multi-directional gradient operator, and strengthening significant structural features through nonlinear mapping; constructing a lightweight weight learning network, and generating a spatial self-adaptive dynamic fusion weight map according to the multi-scale features; executing progressive three-level fusion according to the dynamic weight, and reconstructing to obtain a high-quality image; according to the method, the image is decomposed into feature components with different physical meanings, targeted optimization and adaptive fusion are carried out, and more accurate and robust image quality improvement is realized.

The invention provides a numerical control machining tool wear multispectral imaging detection method and system. The method comprises the steps that a three-dimensional data cube is constructed through a multispectral image sequence of a tool machining area; defining a spectrum detection window of each pixel point in the three-dimensional data cube, dividing the spectrum detection window into a main detection sub-block and auxiliary detection sub-blocks, and generating a spectrum difference feature map based on a difference relationship between a spectrum curve of each pixel in the main detection sub-block and an average spectrum curve of all the auxiliary detection sub-blocks; determining a projection vector according to a target spectral feature in the center detection sub-block and a background spectral feature in the local detection window; and determining a spectral feature map according to the projection vector and the three-dimensional data cube, and recognizing an abnormal wear area in the tool machining area according to a multidirectional gradient feature map constructed by a spatial gradient operator corresponding to each pixel in the spectral feature map and a spectral difference feature map. According to the technical scheme provided by the invention, the subtle difference between the wear region and the background region can be distinguished in a multi-dimensional interference state.

The invention relates to the technical field of target detection, in particular to a blueberry fruit focusing detection method and system based on a lightweight YOLO model, and the method comprises the steps: obtaining blueberry images to form a multi-dimensional data set, and carrying out the marking of the obtained multi-dimensional data set; performing data enhancement based on the acquired multi-dimensional data set, including performing geometric transformation on the acquired original data set, and performing optical simulation and occlusion simulation by adding Gaussian noise and gradient operators; and constructing a detection model by taking the preprocessed multi-dimensional data set as input, introducing a dynamic attention mechanism on the basis of the detection model, carrying out model training and optimization on the basis of the constructed detection model, and inputting a test set for detection by utilizing an optimal weight obtained by training to generate a final detection result. According to the method, dual optimization of semantic understanding and accurate positioning is realized through feature fusion, and the detection precision of overlapped fruits and small targets is effectively improved.

The invention relates to the technical field of infrared image processing, in particular to an infrared image non-uniformity vignetting correction method based on gradient prior, which comprises the following steps: acquiring a uniformly radiated infrared noise image of a target imaging system, and acquiring a gradient trend of infrared non-uniformity vignetting based on the infrared noise image; obtaining a real-time infrared image of the target imaging system, and obtaining a real-time gradient operator and a real-time gradient trend of the real-time infrared image based on the gradient trend; constructing an objective function based on the gradient trend, the real-time gradient operator and the real-time gradient trend, and solving the objective function to obtain an optimal value of the real-time gradient operator; and obtaining an optimal real-time gradient trend based on the real-time gradient operator optimal value, and correcting the real-time infrared image by using the optimal real-time gradient trend to obtain a target infrared image after non-uniformity vignetting correction. By using the method provided by the invention, the correction of the infrared non-uniform vignetting can be better realized.

The invention provides a paver intelligent welt control method and system, and particularly relates to the technical field of road construction machinery intelligentization, the method comprises the following steps: firstly, using industrial cameras installed on two sides of a paver to collect an operation area image in real time, then extracting image edge features, and calculating a definition score; according to a scoring result, selecting different paths to generate target boundary position coordinates: if the score is greater than a set threshold value, adopting a gradient operator and a traditional edge detection algorithm of Hough transformation; and otherwise, applying a target detection model based on YOLO series deep learning. Afterwards, bird's-eye view transformation is carried out based on the target boundary position coordinates to eliminate perspective distortion, and transverse offset is obtained. And finally, generating a screed stretching control instruction according to the transverse offset, transmitting the screed stretching control instruction to a paver controller through a CAN (Controller Area Network) bus, and driving a telescopic oil cylinder to execute screed welting action so as to realize high-precision self-adaptive welting.

The invention belongs to the technical field of unmanned aerial vehicle gate maintenance scheduling, and provides an unmanned aerial vehicle gate maintenance scheduling method and system based on multi-sensor physical correction. Texture, depth and thickness are re-projected to the same graph plane, and material codes, a sand content mean value and a coating thickness form a scene vector; mapping the scene vectors into residual bias and adding the residual bias with the primary features to form fusion features containing physical semantics; according to the fusion features, determining a corrosion-crack confidence map; aligning the corrosion-crack confidence map with a preset determined correction depth and recess map to obtain a pixel-level corrosion level and spray gun accessibility; according to the defect semantic graph, the corrosion grade and the spray gun accessibility; a segmented adaptive corrosion gradient operator is adopted to complete block segmentation, reordering is carried out according to block material requirements and residual material vectors, and a nozzle wheel disc parameter library is inquired to determine a nozzle model, a target pipe diameter and a target flow velocity; and the consistency of thickness measurement and visual texture can be considered.

The application discloses a multi-directional excitation magneto-optical image registration method under hybrid driving, and the method comprises the following steps: collecting the defect magnetic field distribution information of the surface of a ferromagnetic material under multi-directional excitation through a magneto-optical imaging device to obtain a reference image and a floating image; extracting the defect contour of the two images according to a gradient operator; performing spatial transformation on the defect contour of the floating image, and then performing or operation and morphological closing operation on the defect contour of the floating image and the defect contour of the reference image to obtain a defect contour map with closed shape; obtaining the magnetic leakage field distribution map of the defect contour map through defect magnetic leakage field model forward modeling; calculating the similarity measure of the reference image and the magnetic leakage field distribution map generated by the model; and finally continuously updating the spatial transformation parameters through an optimization search algorithm, and the registration is successful when the similarity measure reaches the maximum.

The invention discloses an auxiliary positioning method and system for anesthesia puncture, and the method comprises the steps: determining the number of adaptive blocks according to the variable coefficients of a target image gray histogram and an LBP histogram of a collected ultrasonic image, and achieving the intelligent response to the image content. And the problem of information loss or excessive processing possibly caused by traditional fixed partitioning is avoided. And meanwhile, a gradient operator is utilized to calculate a pixel point gradient magnitude, and a target pixel point is verified in combination with eight-neighborhood analysis, so that the pixel classification accuracy is improved, and texture regions and edge details in the image can be effectively distinguished. And finally, the enhancement degree is determined based on the confidence and gradient magnitude of the pixel points, and enhancement factors are obtained by combining local display comparative analysis, so that the problem of excessive enhancement or insufficient enhancement possibly brought by global unified enhancement is avoided, and the efficiency and effect of image processing are improved. And a diagnosis basis with higher quality is provided for doctors.

The invention relates to the field of aggregate morphology detection and image processing, and discloses an aggregate particle morphology quantification method and system based on Fourier spectrum decoupling, and the system comprises a three-dimensional point cloud collection module which is used for obtaining the three-dimensional point cloud data of aggregate particles; the point cloud data processing module is used for processing the three-dimensional point cloud data to obtain the two-dimensional contour data; the morphological feature analysis module is used for processing the two-dimensional contour data to obtain aggregate decoupling feature data; and the detection and analysis module is used for associating the modules. According to the method, the RGB-D depth point cloud system is used for collecting multi-view point cloud, and the complete three-dimensional morphology point cloud is obtained through noise reduction and registration. Through Monte Carlo simulation sectioning, two-dimensional equal circumference mapping, contour Fourier series fitting, gradient operator definition and the like, a method for performing decoupling calculation on the shape, the corner angle and the texture of the aggregate through aggregate form parameters is established, and the accuracy and the accuracy of aggregate three-dimensional form test analysis are improved.

The present invention discloses an infrared image non-uniformity correction method based on image gray gradient, which includes the following steps: acquiring two frames of images, performing two-point non-uniformity correction on the two frames of images; extracting the image gray gradient features of the corrected images by using a gradient operator; finding the brightest point in the feature map, expanding N pixels up, down, left, and right centered on the brightest point to obtain an image block of a specified size, and performing a binarization operation on the image block; performing row and column projection matching on the two frames of binarized images to obtain offset information; according to the offset information, subtracting the overlapping parts of the two frames of images, calculating the compensation amount required to update the background, and correcting the background; using the corrected background to perform non-uniformity correction on the current frame of image. The beneficial effects of the present invention are: effectively eliminating fixed noise, high-frequency noise, and low-frequency noise, improving the efficiency and accuracy of non-uniformity correction, saving computing resources, and not requiring a shutter as a reference.

The invention relates to the technical field of image data processing, in particular to a screen support defect positioning method and system based on image processing. The method comprises the following steps of: obtaining local change intensity of each pixel point through gradient amplitude difference of the pixel points in a screen bracket surface image under different scale gradient operators; obtaining suspected pixel points in the screen support surface image through the local change intensity of each pixel point, and calculating the target neighborhood radius of the suspected pixel points and the change rate of the suspected pixel points; calculating the defect degree of the suspected pixel point, wherein the defect degree is positively correlated with the local change intensity and the change rate of the suspected pixel point and the LBP positive value extraction value and the LBP negative value extraction value in the target neighborhood radius of the suspected pixel point; in response to the comparison result of the suspected pixel points in the screen support surface image and the threshold value, the defect area in the screen support surface image is positioned, and the accuracy of screen support defect positioning can be effectively improved.

The application discloses a kind of medical image cardiothoracic ratio measurement method and system based on artificial intelligence, method includes: obtaining the first picture that user stands in front of radiographic mainboard and is photographed by top camera, according to first picture, obtain the initial azimuth of user relative to radiographic mainboard;Judge whether the absolute value of initial azimuth exceeds first preset angle;Start chest radiography image device, collect the first chest medical image of user standing in front of radiographic mainboard;According to initial azimuth adjustment binary threshold, according to binary threshold to first chest medical image is carried out binary processing and obtains binary image;Gradient operator is used to extract the contour line of lung in binary image;The cardiothoracic ratio obtained by calculating heart transverse diameter divided by thoracic transverse diameter.The binary threshold of lung edge is adjusted in the application to improve the position accuracy of lung contour line, and then accurate cardiothoracic ratio is obtained.

An image edge detection method for processing an image including multiple pixels includes performing a convolution operation on the image by a gradient operator in a first direction and a gradient operator in a second direction to obtain a first-direction gradient data and a second-direction gradient data, wherein the first direction is perpendicular to the second direction, performing a gradient statistical calculation within a neighboring area of a target pixel of the image according to the first-direction gradient data and the second-direction gradient data to obtain a gradient statistic, and determining an edge significance corresponding to the target pixel according to the gradient statistic.

The present invention discloses an intelligent vision positioning method, including: obtaining a color image feature sequence and a depth image feature sequence through a feature sequence extraction module and using them as inputs; a multi-level depth-embedded Transformer module consists of multiple depth-embedded units and Transformer layers, and each depth-embedded unit takes the feature sequence as an input, aiming to output a feature sequence with enhanced spatial perception; sending the scene feature representation obtained by the multi-level depth-embedded Transformer module into a prediction head to obtain a scene coordinate prediction result, and respectively applying a horizontal gradient operator and a vertical gradient operator to the scene coordinate prediction result to generate a horizontal gradient and a vertical gradient of the scene coordinate prediction result; the horizontal gradient operator and the vertical gradient operator also respectively act on the depth image to generate a horizontal gradient and a vertical gradient of the depth image; training an intelligent vision positioning network using depth-guided smooth constraints, regression loss, and reprojection loss, and constructing a pose solver for pose sampling and pose refinement.

The application discloses a U-Net-based optical and SAR remote sensing image optical flow registration method and relates to the technical field of heterogeneous image registration, and comprises the construction of an image data set, image preprocessing and the labeling of a region of interest; two U-Net network models are used to train an optical remote sensing image segmentation model and a SAR remote sensing image segmentation model respectively; U-Net network segmentation results of a to-be-registered image pair are acquired, and pixel point sets marked in specified channels of two segmentation images are recorded respectively; gradient operators are used to construct class GLOH descriptors of region-of-interest feature points and pixel points in specified neighborhoods of the region-of-interest feature points in the to-be-registered image pair; and the region-of-interest feature points in the to-be-registered image pair are registered by using a Gaussian pyramid LK optical flow method. The application can realize the registration of heterogeneous images with higher precision, stronger purpose and better timeliness.

A fast binocular method applicable to active obstacle avoidance of a mobile robot is provided. The method includes: S110) obtaining a first disparity map and a depth and point cloud map having the same first resolution as the first disparity map based on an initial left image and an initial right image in a binocular camera; S120) using a predetermined number of disparity grid gradient operators (e.g., 32) to generate a disparity grid gradient texture map of a predetermined number of bits (e.g., 32 bits) from the first disparity map with a super-grid composed of m rows * m columns of grids of the first disparity map as a unit; S130) obtaining the positions of obstacle foreground objects and the ground from the disparity grid gradient texture map; S140) determining, based on the obtained positions of the obstacle foreground objects and the ground, the obstacle foreground objects and the ground for which the mobile robot needs to actively avoid obstacles from the depth and point cloud map, where each grid is composed of n rows * n columns of pixels, and each of m and n is an integer greater than or equal to 1 and m is an odd number.

The present application relates to the field of aggregate morphology detection and image processing, and discloses a method and system for quantifying aggregate particle morphology based on Fourier spectrum decoupling, including: a three-dimensional point cloud acquisition module for acquiring three-dimensional point cloud data of aggregate particles; a point cloud data processing module for processing the three-dimensional point cloud data to obtain the two-dimensional contour data; a morphological feature analysis module for processing the two-dimensional contour data to obtain aggregate decoupling feature data; and a detection and analysis module for associating various modules. The present invention uses an RGB‑D deep point cloud system to collect multi-view point clouds, and obtains a complete three-dimensional morphological point cloud through noise reduction and registration. Through Monte Carlo simulation sectioning, two-dimensional isoperimetric circle mapping, contour Fourier series fitting and gradient operator definition, a calculation method for decoupling the shape, edges and texture of aggregates using aggregate morphological parameters is established, thereby improving the precision and accuracy of aggregate three-dimensional morphological testing and analysis.

The invention relates to the technical field of ultrasonic tomography, and discloses a three-dimensional ultrasonic tomography sound velocity imaging method based on a multi-template fast marching method, which comprises the following steps: acquiring parameters of a three-dimensional ultrasonic tomography system and initial sound velocity distribution of biological tissues, and establishing a curved ray model based on an eikonal equation; discretizing the propagation time gradient in the eikonal equation by adopting second-order finite difference approximation, and covering 18 neighborhoods of the grid by adopting templates in four directions to carry out combined solution, so as to obtain ultrasonic propagation time distribution of the whole grid; and iteratively reconstructing a three-dimensional sound velocity distribution image through a maximum posterior probability estimation algorithm based on the ultrasonic propagation time distribution. According to the method, the propagation time gradient in the second-order finite difference approximation eikonal equation is adopted to reduce discretization errors, a plurality of direction templates are used, more accurate difference approximation is carried out on gradient operators, and the non-orthogonal propagation path of the ultrasonic waves in the three-dimensional space is completely captured.

The application discloses a metal artifact removal method of a CT image, and the method comprises the following steps: constructing an initial adaptive iterative learning model based on wavelet transform, decomposing a CT image by using a target adaptive iterative learning model obtained by optimizing an optimization objective function, calculating the area of a metal artifact, removing the metal artifact in the CT image, and obtaining a CT image with reduced metal artifacts; the initial adaptive iterative learning model based on wavelet transform is used for artifact removal, so that the spatial distribution characteristics of the metal artifact under different domains and resolutions can be fully utilized, the artifact removal is better, and the interpretability is high; in addition, when a first optimization objective function is solved by combining a proximal gradient descent algorithm and a Taylor formula, a proximal gradient operator obtained can be replaced by a simple network module, so that the network can be more easily constructed, and the adaptability of the network is enhanced. The application can be widely applied to the technical field of CT image processing.

The application discloses a laser stripe center line extraction method based on Chebyshev moments, and belongs to the technical field of line structured light vision detection. The Scharr gradient operator is used to detect the laser stripe edge, and then searching is performed along the normal direction of the edge point, so that the laser stripe cross-section gray distribution information is obtained. On the basis of analyzing the laser stripe cross-section gray distribution characteristics, a laser stripe cross-section gray distribution model is constructed, and the laser stripe cross-section center sub-pixel coordinates are solved by using Chebyshev moments, and the laser stripe center line is obtained by connecting the laser stripe cross-section centers. Compared with the Steger method, the gray gravity center method, the spatial moment method and the Legendre moment method, when the gray saturation laser stripe is processed, the center line detection precision can be improved while the algorithm real-time is ensured, the balance between the algorithm precision and speed is realized, and the measurement requirements of the line structured light measurement system can be met in the actual use.

This invention discloses a solution to the joint path and destination planning problem based on a distributed generalized Nash equilibrium algorithm. First, the joint path and destination planning problem is modeled, transforming it into a non-cooperative game model. This model includes the objective function of each electric vehicle, global coupling constraints, and local constraints. Second, pseudo-gradients are used to transform the game model into a VI problem, introducing edge-based consistency constraints and a heterogeneous step size mechanism. Based on fixed-point iteration and proximal gradient operator theory, a distributed solution algorithm under complete information is proposed. Then, a global estimate of the plans of other users is introduced, proposing a distributed solution algorithm under partial information. This invention avoids the construction of double random matrices through edge-based consistency constraints, and can maintain solution accuracy while meeting low computational requirements when more users participate in the game model.

The application relates to a gradient distribution-based video image denoising method and system, which comprises the following steps: separating each channel of a to-be-processed image; setting a plurality of direction gradient operators, and calculating the gradient values of each direction of a single-channel image current frame; setting a plurality of pixel value screening directions with a current pixel point as the center, judging whether the gradient value difference and the gray value difference between the current pixel point and the neighborhood pixel point meet the screening condition in each pixel value screening direction in turn; if the screening condition is met, the gray value and the pixel number of the neighborhood pixel point are accumulated; if the screening condition is not met, the pixel point screening in the direction is terminated; the average value of the accumulated pixel gray value and the pixel number in each pixel value screening direction is calculated as the filtered gray value of the current pixel point; the gradient distribution-based spatial filtering and the time domain filtering based on adjacent frame difference constraint are combined, so that the real-time video image denoising demand is met.

The invention discloses a physical field gradient calculation method and system based on a graph neural network. The method comprises the following steps: acquiring physical field data; obtaining an original gradient operator based on the grid topological relation; updating the physical field data based on the graph neural network, and obtaining a first gradient operator based on the grid topological relation of the updated physical field data and the original gradient operator; decoupling different directions of the physical field to set corresponding direction training weights; obtaining a second gradient operator based on the original gradient operator, the first gradient operator and the direction training weight; and performing gradient calculation of the physical field based on the second gradient operator. The system corresponds to the method. According to the method, the problem of large physical field gradient calculation error caused by grid defects in the prior art is solved, and the precision and robustness of internal force calculation in finite element analysis are improved.

The invention provides a landslide multi-scale boundary sensing and recognition network based on remote sensing images, and belongs to the technical field of geological disaster prevention. A coder-decoder structure is adopted, a multi-scale cross interaction convolution module is arranged in each stage of a coder, a boundary sensitive refining attention module is arranged in a network bottleneck layer, and self-adaptive cross fusion of multi-branch features is realized through the multi-scale cross interaction convolution module. The boundary sensitive refining attention module is used for realizing dual-path collaborative refining of traditional gradient operator boundary prior and learning boundary, so that the identification recall rate of landslides with different scales and the positioning precision of landslide boundaries are improved at the same time. The technical problems of insufficient multi-scale feature interaction, boundary gradient prior deficiency and large feature fusion information loss in the prior art are solved. All indexes are obviously improved, the overall recognition precision and the boundary positioning precision are higher, and the method has higher practical application value.

The application provides a multi-focus image fusion method combining a deep residual network and a variational method, and relates to the technical field of digital image processing. The application firstly acquires a first source image and a second source image which are complementary to a focusing area, determines a first-order gradient and a first-order gradient module of the first source image and the second source image; outputs a focus point image and a gradient module score image through an MResNet deep residual network after training, generates an initial fusion image and a fusion gradient guide term; constructs a light-weighted total variation model which only adopts a two-direction first-order gradient operator, and obtains a total variation fusion result at a junction; and performs block fusion based on a weight map to obtain a final fusion image. The application can retain clear area information of source images and improve the transition continuity at a focusing and defocusing junction.

The invention relates to the technical field of computer vision and artificial intelligence image processing, and discloses an explosive object target identification method based on multi-dimensional data optimization and feature fusion, and the method comprises the following steps: carrying out the HSV space brightness adjustment, random cutting and gray conversion of an original image, and generating a single-channel gray image; extracting a multi-level feature map by using a convolutional neural network; utilizing a fixed gradient operator to generate a structure attention mask enhanced superficial layer feature map; re-standardizing each feature map to generate a multi-scale feature map with aligned distribution; a bidirectional cross-scale transmission path is constructed and weighted fusion is carried out; and performing category prediction and bounding box regression through the decoupling prediction network, and outputting an explosive object detection result. According to the method, random disturbance and single-channel graying preprocessing logic based on an HSV brightness channel is constructed, non-essential color interference can be actively stripped, a model is forced to focus on inherent geometric texture features of an object, and therefore the illumination invariance basis is established at the data input end.