254 results about "Confidence map" patented technology

The invention discloses a pump shell welding seam quality detection method based on image segmentation. The method comprises the following steps: generating a steady-state pump shell welding seam image flow under the driving of motion compensation; obtaining a domain adaptive DINOv2 visual embedded feature map; performing adaptive pyramid fusion and cross-scale attention operation on the domain adaptive DINOv2 visual embedded feature map to generate a semantic form segmentation map; generating a semantic-texture fusion mask; performing uncertainty weighted optimization on the semantic-texture fusion mask in combination with the pixel-level confidence map to obtain a weld defect instance map; generating an interpretable texture anomaly heat map; and through multi-view supplementary shooting or manual auditing, supplementary pump shell welding seam image data is obtained, and the steady-state pump shell welding seam image flow is updated. According to the method, the system can continuously keep accurate positioning of the pixel-level segmentation boundary in a weak-label or even non-label migration scene, and boundary drift and area missing detection of a segmentation result are effectively avoided.

The invention provides an intelligent eddy current nondestructive testing method and system based on a flexible GMR sensor array, and relates to the technical field of nondestructive testing. According to the invention, a GMR sensor array and a micro excitation coil array are integrated on a flexible substrate, and synchronous acquisition under multi-frequency and multi-phase excitation is realized by establishing a probe-workpiece unified space coordinate system and a time synchronization reference; performing temperature drift compensation, attitude correction and gap normalization processing on the original response tensor to generate a corrected response tensor; the intelligent inversion model based on the prior constraint of the electromagnetic field jointly identifies the conductivity, the thickness and the crack parameters, and outputs a defect parameter map and a confidence map; and a detection report containing the position, the size, the depth and the confidence coefficient is generated through spatial clustering and connectivity analysis, so that high-sensitivity and high-reliability nondestructive detection under a complex curved surface structure is realized.

The invention relates to a three-dimensional point cloud registration method and system based on a two-dimensional visual large model. A source point cloud and a target point cloud are decoupled into two-dimensional projection representation; constructing a structure consistency restoration domain, restoring geometric structure fracture and semantic deficiency in the projection image, and generating a dense and continuous restoration image; inputting the repaired image into a pre-trained visual geometry large model, carrying out cross-modal feature matching and pose resolving, and synchronously outputting a pixel-level matching confidence map; and inversely mapping a two-dimensional pose calculation result to a three-dimensional space to complete coarse registration, constructing a high-credibility anchor point set in the three-dimensional space by using a confidence map, and guiding a point cloud fine registration algorithm to converge to global optimum. According to the method, the registration large model for natural image training can be seamlessly migrated to the point cloud data, and retraining is not needed; in a weak texture and partially overlapped point cloud scene, a robust rough registration initial value can still be provided; the convergence speed and the anti-noise performance of fine registration are remarkably improved, and adaptive improvement from coarse to fine is achieved.

The invention discloses a real-time monitoring system based on highway traffic flow monitoring, and the system specifically comprises a video collection and self-calibration module which obtains calibration video data with an aligned visual angle; the edge consistency adaptive enhancement module outputs an enhanced video frame, an edge intensity graph and a noise risk graph, and constructs an edge confidence graph; the improved segmentation SAM module is used for carrying out traffic scene domain adaptation training on the improved segmentation SAM model by adopting Adapter to generate an initial multi-target segmentation mask; the deformation optimization module is used for calculating a corresponding boundary stability index; the traffic flow index extraction module is used for generating a corrected tracking result; the abnormal event candidate recognition module is used for generating an event confidence score for each traffic flow abnormal event candidate; and the abnormal event output module is used for outputting graded abnormal event alarm information. According to the method, the problems of boundary edge eating, missing detection and inter-frame drifting are remarkably reduced, and high-time-space-consistency segmentation is realized.

The invention relates to the technical field of detection, in particular to a complex shell detection method and system based on structured light scanning. The method comprises the following steps: calculating a wrapped phase value and a phase modulation amplitude of each pixel point under each frequency, calculating discrete probability distribution containing candidate stripe series and a corresponding probability for each pixel point based on a multi-frequency wrapped phase difference, generating an initial pixel confidence map, and constructing an energy function; the penalty weight of the smooth item is determined according to the phase modulation amplitude and the wrapped phase gradient local variance, obtaining an initial fringe order distribution diagram, updating discrete probability distribution by using an initial pixel confidence map, re-minimizing an energy function to obtain a corrected fringe order distribution diagram, and reconstructing a final three-dimensional shape. And iteratively optimizing a projector principal point and a radial distortion coefficient in system calibration parameters. According to the scheme, the geometric details of the surface can be reserved, and the processing effect and the measurement precision of the complex curved surface and the high-curvature area are improved.

The invention relates to the technical field of intelligent detection, and discloses a high-temperature alloy casting size on-line detection system based on machine vision, and the system comprises an imaging fusion module which collects image data of a high-temperature alloy casting, carries out the pixel-level registration of infrared image data and visible light image data, and generates a multispectral fusion image. The feature extraction module performs temperature distribution analysis and brightness feature extraction based on the multispectral fusion image, generates an edge confidence map and extracts a feature point set. And the calculation module performs weighted fitting on the feature points according to the feature point set to obtain thermal state size parameters. And the thermal compensation module performs thermal compensation and geometric correction on the thermal-state size parameters to obtain cold-state size parameters. And the judgment module carries out comparison to judge whether the casting is a qualified casting. According to the invention, stable imaging and accurate registration in high-temperature radiation and strong reflection environments are realized, and the detection efficiency and the size control level of the high-temperature alloy casting are improved.

In various examples, a deep neural network(s) (e.g., a convolutional neural network) may be trained to detect moving and stationary obstacles from RADAR data of a three-dimensional (3D) space, in both highway and urban scenarios. RADAR detections may be accumulated, ego-motion-compensated, orthographically projected, and fed into a neural network(s). The neural network(s) may include a common trunk with a feature extractor and several heads that predict different outputs such as a class confidence head that predicts a confidence map and an instance regression head that predicts object instance data for detected objects. The outputs may be decoded, filtered, and / or clustered to form bounding shapes identifying the location, size, and / or orientation of detected object instances. The detected object instances may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.

The invention provides an electroencephalogram signal artifact removing method and device, equipment and a medium, and relates to the technical field of biological signal processing, collected electroencephalogram signal data is processed to obtain an optimal modal number and an optimal bandwidth parameter, and the optimal modal number and the optimal bandwidth parameter are used for conducting self-adaptive variational mode decomposition on the electroencephalogram signal data to obtain a plurality of modal components; then carrying out multi-dimensional feature analysis to obtain a plurality of feature indexes for judging an artifact suspicion mode and an effective mode component; performing short-time Fourier transform on the artifact suspicion mode, and constructing a time-frequency confidence map to guide weighted time-frequency independent component separation on the artifact suspicion mode to obtain an artifact component; suppressing the artifact component to obtain a suppressed independent component, and performing weighted reconstruction and multi-component fusion on the suppressed independent component and the effective modal component based on the time-frequency confidence map to obtain an artifact-removed electroencephalogram signal; and the fidelity, the robustness and the real-time performance of the electroencephalogram signal are improved.

The invention discloses a three-dimensional scene reconstruction method and system based on monocular depth estimation, and belongs to the technical field of computer vision and three-dimensional reconstruction, and the method comprises the steps: carrying out the multi-scale feature coding of a monocular RGB image through a mixed attention depth coding module, and obtaining the hierarchical depth feature representation; carrying out autoregression depth decoding through a self-adaptive edge perception depth decoding module to generate an initial depth map; a depth confidence map is calculated through a geometric consistency constraint optimization module and is fed back to a coding module for iterative optimization, and a refined depth map is output; and three-dimensional Gaussian ellipsoid scene representation is constructed through the Gaussian ellipsoid scene reconstruction module. According to the invention, high-precision depth estimation and high-quality three-dimensional reconstruction are realized by constructing a depth-coupled closed-loop cooperative system.

The invention discloses a method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction, and the method comprises the steps: cutting an image in real time, obtaining an image block which takes a component as a target main body, and synchronously recording a homography matrix for geometric mapping; selecting an amplification strategy to improve the resolution, and recording a scale mapping relation; inputting the enhanced image block into a double-flow network; adaptive fusion and reconstruction are carried out on the two branch features, and a high-resolution texture image is output; generating a geometrically corrected ortho-image, fusing the geometrically corrected ortho-image with original illumination information, and outputting a corrected image with a known pixel size; identifying cracks, spalling and honeycomb diseases in parallel; generating a unified defect confidence map; calculating real geometric parameters of the BIM in a BIM global coordinate system through coordinate back projection; and generating quantitative defect reports and maintenance suggestions. The method has the advantage that seamless connection between the detection result and the BIM global coordinates is realized.

The invention provides a water body water level monitoring method of an image segmentation model, and the method comprises the steps: obtaining multi-modal water flow data of a target water area, carrying out the feature extraction of the multi-modal water flow data, and obtaining a turbulence feature map, the multi-modal water flow data comprising water flow motion data and water body image data; performing physical constraint segmentation processing on the turbulence feature map and the water body image data to obtain a water body region segmentation mask and a corresponding pixel-level confidence map; carrying out water gauge geometric correction processing on the water body region segmentation mask to obtain a corrected water line; and performing water level value conversion processing on the corrected water level line and the pixel-level confidence map to obtain a water level measurement value and uncertainty data. By adopting the method, the real water line and the dynamic water flow artifact can be effectively distinguished, and the monitoring precision under the complex hydrological condition is ensured.

The invention relates to the technical field of communication, and discloses a three-dimensional radio environment map construction method and system, a terminal and a storage medium, and the core is to construct a'base station unmanned aerial vehicle 'bidirectional interaction closed-loop optimization framework to realize efficient construction of a high-precision map. Self-adaptively fusing sparse radio measurement data and environmental building structure features; introducing a confidence evaluation mechanism based on adversarial learning and position weighting, and generating a pixel-by-pixel confidence map; an intelligent planning method based on a trajectory diffusion model is designed, local perception constraint and long-term information gain are cooperated with a classifier-free guide mechanism, and an optimal trajectory considering both safety and sampling efficiency is generated; and a continuously self-optimized closed-loop system is formed through newly acquired data of the unmanned aerial vehicle and periodical updating of the model. According to the invention, a high-reliability technical basis is provided for applications such as urban air communication and spectrum resource management.

The invention relates to a solid-state laser radar ranging method and system based on neural network processing, and belongs to the technical field of laser radar ranging, and the method comprises the steps: obtaining a laser echo signal when laser radar detection is carried out on a surrounding target object, and obtaining a photon counting histogram based on the laser echo signal; preprocessing the photon counting histogram, and inputting the preprocessed photon counting histogram into a pre-trained neural network model to obtain a peak confidence map containing an echo peak value and a peak time offset map; on the basis of the peak confidence map and the peak time offset map, obtaining one or more echo peak values of which the confidence exceeds a predetermined threshold and time position information of the one or more echo peak values of which the confidence exceeds the predetermined threshold; and based on the time position information, the distance information with one or more target objects is determined, so that the time position information measurement accuracy of one or more echo peak values is improved, and the laser radar ranging accuracy is improved.

The invention provides a three-dimensional reconstruction method based on guided filtering and Mama geometric feature fusion in the technical field of three-dimensional reconstruction, and the method comprises the steps: S1, collecting a multi-view RGB image, carrying out the calibration of a camera, obtaining a projection matrix and an internal reference matrix, and obtaining a reference image; s2, extracting image features from each RGB image; s3, sampling on the reference image to obtain a depth hypothesis plane, performing homography transformation on each image feature, and aggregating with the depth hypothesis plane to obtain a cost body; s4, sequentially performing filtering, regularization and exponential normalization on the cost body to obtain a probability body, and constructing a depth map and a confidence map based on the probability body; s5, performing frequency domain filtering and up-sampling on the depth map and the confidence map, and then performing feature fusion to obtain a visual angle depth; and S6, executing three-dimensional reconstruction operation based on the visual angle depth, the projection matrix and the internal reference matrix. The method has the advantage that the precision and robustness of three-dimensional reconstruction are greatly improved.

The invention relates to an underwater single target tracking method based on a wavelet token and a space-time Transform. The method comprises the following steps: firstly, constructing a reference frame sequence, a search frame and a previous frame historical token into a space-time input sequence, and extracting cross-frame features through a Transform encoder; then, Haar wavelet decomposition is carried out on the historical token, and a low-frequency component representing a target structure and a high-frequency component capturing motion details are separated out; then, adaptively fusing the global features and the historical components of the current search frame by using a gating mechanism, and generating a wavelet token; and finally, inputting the wavelet token and the global feature into a prediction head, and outputting a target classification confidence map and a bounding box regression map to determine the position and the scale of the target. According to the technical scheme of the invention, the interference of underwater low-illumination noise can be effectively suppressed through the wavelet token, and the space-time continuity of target motion modeling is maintained in combination with a gating strategy, so that the tracking robustness of an underwater complex scene is effectively improved.

The invention relates to the technical field of geological survey, in particular to an unmanned aerial vehicle multispectral geological survey method and system, and the method comprises the steps: fusing a multispectral image, a digital elevation model, geophysics and historical geological data, systematically constructing a geological feature priori knowledge model, including lithology, construction and alteration feature libraries, and mapping with multispectral data; multi-scale geologic features are extracted through adaptive wavelet transform and morphological analysis, and feature weight adaptive adjustment is achieved; geological units are accurately divided by adopting geological scene perception superpixel segmentation and combining geological boundary constraint and similarity recursion combination; identifying an interference mode, generating an adaptive filtering matrix, and enhancing image quality; cooperatively interpreting multi-source information by using a deep auto-encoder network to generate a high-precision geological interpretation map and a confidence map; geological professional knowledge is introduced, so that the geologic body recognition accuracy is remarkably improved; the adaptive flight control strategy ensures the consistency of complex terrain data, and improves the precision and efficiency of geological survey.

The invention relates to a transient electromagnetic and seismic wave multi-mode joint inversion imaging method based on physical information Transform, and belongs to the crossing field of geophysical exploration and artificial intelligence. Comprising the following steps: carrying out anomaly detection, interpolation, filtering and normalization preprocessing on transient electromagnetic and seismic wave original data; extracting features representing electrical property, elasticity and cross physical significance; serializing the spatial data through a gridding and alternating fusion strategy, and constructing an enhanced code fusing absolute and relative positions and physical attributes; a physically constrained encoder-decoder architecture is designed to carry out multi-scale forward modeling-inversion; and quantizing the uncertainty of an inversion result by adopting a Bayesian Monte Carlo method, and generating a confidence map. According to the method, through physical rule driving and multi-modal depth complementary fusion, the fine recognition capability and interpretation reliability of hidden disaster-causing geologic bodies such as underground goaf and collapse columns are effectively improved while the physical consistency of data is kept.

The invention discloses an unmanned aerial vehicle electric power inspection image segmentation method and system, and relates to the technical field of image segmentation. By means of a labeled sample image set, positive and negative information is fully fused in a confidence map cooperation module to carry out confidence modeling, so that a cooperation confidence map can accurately reflect the difference between a target and a background; the negative sample distribution data effectively depicts the features of background interference, then high-discrimination positive and negative points are screened out in the point selection module, accurate prompt is provided for segmentation, iterative optimization is performed through the noise sensing and refining module, and therefore the segmentation efficiency is improved under the condition that only few labeled samples are needed. According to the method, high-precision, high-robustness and training-free segmentation of the target component in the electric power inspection image is realized, the influence of complex backgrounds and similar interferents is effectively resisted, the method has high robustness, the technical problems of low segmentation precision and insufficient robustness under the condition of few labeled samples in the prior art are fully solved, and the requirements of electric power inspection on automation and precision are met.

The invention discloses an operation interaction method and system applied to camera image editing, and relates to the technical field of image processing, and the method comprises the steps: obtaining a voice semantic heat map, a pointing intensity map, a touch confidence map and a gazing confidence map based on a multi-modal interaction data packet, and calculating an image feature matrix at the same time; fusing into a multi-modal evidence graph through a normalized scale; performing semantic segmentation according to the image feature matrix to obtain a semantic segmentation first draft and a pixel-by-pixel category confidence coefficient, and performing position correlation weighting on the pixel-by-pixel category confidence coefficient by taking the multi-modal evidence graph as a confidence coefficient modulation factor to generate a candidate object mask sequence; and performing highlight display on the candidate object mask sequence, and performing conflict resolution and priority rearrangement in combination with the multi-mode evidence graph to generate a target object mask. According to the method, deep fusion of the interaction intention and image segmentation is realized, the precision and consistency of candidate region detection are improved, and the stability of real-time rendering and the reliability of an editing result are improved.

The invention discloses an extremely low bit rate image compression coding and decoding method of a stream matching diffusion model. The coding method comprises the following steps: acquiring an original image, and coding the original image into a compression latent variable and a hierarchical feature; based on the compression latent variable, generating an anchor point set and a mask parameter in a continuous domain; generating a visible index, a stream matching scheduling parameter and a confidence graph according to the compression latent variable, the hierarchical feature and the target bit rate; and combining the visible index, the anchor point set, the mask parameter, the stream matching scheduling parameter and the confidence map to form a code stream. According to the invention, through continuous domain mask and confidence gating reasoning of code rate perception, the method is suitable for image compression with an extremely low bit rate.

An autonomous warehouse robotics system integrates a multi-sensor platform, adaptive payload handling, dynamic task reallocation, advanced navigation, energy management, and comprehensive safety features into one mobile robot chassis. The system utilizes LiDAR, stereo vision, ultrasonic sensors, mmWave radar, thermal cameras, and event cameras to perform complete environmental sensing and obstacle detection. Sensor fusion combines adaptive weighting, multi-modal data integration, and statistical filtering to create high-confidence maps for reactive path planning and collision avoidance. The robot's payload system features machine vision for item recognition, telescopic lifts, variable-width grippers, and real-time toolhead verification to handle a variety of goods. Fleet management is achieved through dynamic task reallocation that considers robot location, battery level, and operational delays, all coordinated by a cross-platform middleware architecture that ensures standardized communication, remote monitoring, and over-the-air updates among diverse robot brands. Energy management optimizes power usage via predictive routing, autonomous return-to-charge, and auction-based scheduling. Safety is maintained through proximity detection, behavior-based intervention, and human-robot cohabitation protocols while advanced localization is enhanced by fusing ultra-wideband positioning with visual landmark alignment, inertial sensing, and machine learning to deliver high accuracy in non-line-of-sight conditions. An onboard edge AI module further refines navigation and task prioritization through neural network inference, ensuring robust, adaptive operation in dynamic, unstructured warehouse environments.

The invention relates to an insulator image segmentation method based on an infrared enhancement image of an unmanned aerial vehicle, and relates to the technical field of image processing, and the method comprises the steps: collecting an infrared image sequence through an infrared enhancement camera, obtaining a standard infrared image sequence, and carrying out the state recognition through an image perception prior engine, outputting a predicted insulator attention heat map and a predicted boundary confidence map; optimizing and adjusting the initial threshold segmentation algorithm and the initial region growing algorithm, and obtaining an adaptive threshold segmentation algorithm and an adaptive region growing algorithm; and carrying out image segmentation on the standard infrared image sequence to output an insulator image sequence, and carrying out early warning judgment on the insulator image sequence based on an adaptive early warning mechanism. The method solves the problems that a traditional insulator image segmentation method cannot effectively deal with the problems of low infrared image temperature contrast ratio, large noise interference and complex background, so that the insulator segmentation is easy to cause mistaken segmentation and missing segmentation, and the high-precision requirement of fault detection is difficult to meet.

The invention discloses a skin disease diagnosis method based on image recognition, and the method comprises the following steps: obtaining a to-be-detected skin image, and recording an equipment identifier; carrying out white balance correction, reflection suppression, hair suppression and scale normalization to generate a standardized image; positioning the focus to generate a focus mask; generating an enhanced view set in the focus mask by the enhanced view set according to the enhanced times and the enhanced view set; calculating a color invariant graph, a sign response graph and an artifact confidence graph, and synthesizing a potential energy field set according to a fusion weight; calculating a local structure tensor and generating a Riemannian metric matrix; constructing an improved Morse-Smal complex under a Riemannian metric matrix, carrying out cross-complex clustering on the critical point set to obtain a stable score, carrying out complex simplification to generate a stable Morse-Smal complex, and calculating the confidence coefficient of the cell boundary; and extracting structured feature vectors, diagnosing, judging and outputting disease categories, grading results and re-checking and marking. According to the invention, the anti-artifact and cross-device stability is improved.

An image background blurring method includes: obtaining an image to be blurred; performing identification processing on the image by an image semantic segmentation model in response to a target object confirmation instruction to obtain a confidence map indicating a target region of a target object in the image; generating a background image of the image according to the confidence map; performing blur processing on the background image in response to a blurring instruction including a blurring degree to obtain a blurred background image, wherein the blur processing includes determining a reduction factor according to the blurring degree and performing reduction processing on the background image according to the reduction factor; and merging the blurred background image and the image according to the confidence map to obtain a target image. The present application enhances blur processing efficiency by reducing a data computation amount during the blur processing.

A deep neural network(s) (DNN) may be used to perform panoptic segmentation by performing pixel-level class and instance segmentation of a scene using a single pass of the DNN. Generally, one or more images and / or other sensor data may be stitched together, stacked, and / or combined, and fed into a DNN that includes a common trunk and several heads that predict different outputs. The DNN may include a class confidence head that predicts a confidence map representing pixels that belong to particular classes, an instance regression head that predicts object instance data for detected objects, an instance clustering head that predicts a confidence map of pixels that belong to particular instances, and / or a depth head that predicts range values. These outputs may be decoded to identify bounding shapes, class labels, instance labels, and / or range values for detected objects, and used to enable safe path planning and control of an autonomous vehicle.

The embodiment of the invention provides a face forgery detection method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring a pre-training detection model and a to-be-detected face image, and inputting the to-be-detected face image into the pre-training detection model, so that a face key point detector obtains a face key point graph according to the input to-be-detected face image; a semantic segmentation graph generator generates a semantic segmentation graph according to the face key point graph; the feature extractor performs basic feature extraction according to the to-be-detected face image to obtain basic features; the global feature extraction branch performs global feature extraction according to the basic features to obtain a global feature map; the feature learning branch obtains a region-level forgery confidence map according to the basic features and the semantic segmentation map; the feature fusion device carries out fusion processing according to the global feature map and the region-level forged confidence map to obtain classification features; and the classifier obtains a face forgery detection result according to the classification features. According to the invention, the accuracy of face forgery detection is improved.

The invention belongs to the technical field of pavement detection, and discloses a highway pavement roadbed detection method and system based on laser three-dimensional scanning. By establishing an analytic optical or radar scattering physical model, a distance measurement deviation initial value conforming to a multi-path reflection physical mechanism is provided for each laser measurement point, and then refined correction is performed on the initial value through a residual error regression network fusing local point cloud context information, so that advantage complementation of physical prior and data driving is realized, and the accuracy of distance measurement is improved. And finally, in combination with adaptive filtering based on a confidence map, ranging errors caused by a multi-path effect can be effectively distinguished and corrected in the presence of complex road surface conditions such as slippery road surface, greasy dirt or thin water film, and meanwhile, geometric details such as real road surface pits, textures and the like are completely reserved. Therefore, the reconstruction error and the disease false alarm rate of the pavement digital elevation model are obviously reduced, and the precision and the reliability of highway pavement roadbed detection are improved.

The invention discloses a power line image super-resolution recovery system and method for power transmission line patrol, and relates to the technical field of unmanned aerial vehicle image processing. The system provided by the invention comprises three core modules, namely a physical perception generator, a differentiable physical discriminator and a spatial transformation network (STN), wherein the generator is combined with a U-Net and a residual structure, and the output is ensured to accord with the geometry and material characteristics of a power line through physical constraint loss; the discriminator adopts a double-branch structure, and synchronously evaluates image authenticity and physical rationality; and the STN module corrects perspective distortion by using the pose data of the unmanned aerial vehicle to realize geometric alignment of the power line. The system integrates advanced components such as Swin Transformer, ViT and differentiable RANSAC, supports end-to-end processing from original Bayer data to a high-definition image, and outputs a defect detection confidence map with enhanced physical constraints. Compared with a traditional data driving method, the method strictly guarantees the physical credibility of a reconstruction result while improving the resolution, and is suitable for a power line inspection task in a complex environment.

The invention relates to a pavement crack intelligent detection method, system and device based on image recognition and a medium. The method comprises the steps of obtaining an original pavement image, and performing preprocessing according to global brightness distribution of the original pavement image to obtain an image after illumination equalization; inputting the illumination equalization image into a pre-trained convolutional neural network to extract multi-scale texture features, and generating a multi-scale texture feature map; mapping the feature map into a crack confidence map through a segmentation network; performing threshold segmentation based on the confidence map to obtain a crack binary image; performing sub-pixel-level edge refinement on the binary image to obtain a sub-pixel-level crack edge set; and based on the edge set, crack width and length quantization data are calculated through geometric parameters, and a detection result is output. By adopting the method, local shielding interference can be effectively resisted, a scene with dramatic illumination change is adapted, sub-pixel-level precise positioning and quantification of the crack edge are realized, the recognition capability and detection continuity of the fine crack are improved, and the reliability and generalization capability in a complex road environment are enhanced.

The invention provides a multi-modal image fusion system and method for neurosurgery, and the method comprises the steps: carrying out the rigid transformation of a multi-modal magnetic resonance image and an ultrasonic image, and obtaining a preoperative magnetic resonance image and an intraoperative ultrasonic image after preliminary alignment; inputting the preoperative magnetic resonance image and the intraoperative ultrasonic image into a pre-trained lightweight deformation registration network for deformation modeling to obtain a voxel-level dense deformation field, and determining a confidence map of deformation field region registration based on the dense deformation field; according to the confidence map, multi-scale feature fusion based on confidence weighting is carried out on the deformed preoperative multi-modal labeling information and the corresponding features of the intra-operative ultrasonic image, and a three-dimensional anatomical structure chart for real-time navigation in the neurosurgery operation is obtained. Based on the scheme, brain tissue drift can be dynamically compensated in real time in a neurosurgery operation, and multi-modal image fusion is guided based on a compensation result.