520 results about "Phase image" patented technology

The invention discloses an intelligent change detection method for non-aligned two-time-phase images of an inspection unmanned aerial vehicle, and relates to the technical field of unmanned aerial vehicle image processing and change detection, and the method comprises the steps: obtaining two-phase images collected by a low-altitude unmanned aerial vehicle under a fixed route and same sensor parameters; a lightweight registration model is constructed and trained, feature point matching is utilized to predict matching point pairs, a homography matrix is calculated, and accurate registration of non-aligned images is achieved; and constructing and training a change detection model based on image pair interaction feature fusion, analyzing the aligned image after registration, and outputting a change information binary image. The method can effectively solve the problem of non-alignment caused by position and angle differences during two-time-phase image acquisition of an unmanned aerial vehicle, and the technical problems of low precision, poor robustness and insufficient calculation efficiency of a traditional method in change detection, effectively improves the automation level of low-altitude safety monitoring of ground highways and railways, reduces the maintenance cost, and improves the safety of the unmanned aerial vehicle. The important application value is realized.

The invention discloses a method and a device for correcting laser focusing aberration in a transparent material, and belongs to the technical field of laser processing. The method comprises the steps that a machining light path containing an object plane, an aspheric reflector and a 4F device is built in Zemax, the surface type of the aspheric reflector is optimized with the minimum focus aberration as the target, and a rise table is derived; an optical path difference and a laser phase are calculated through Matlab, and a phase diagram is generated; and an actual machining light path is built, and a phase diagram is loaded to achieve low-aberration machining. Simulation optimization and phase modulation are combined, aberration caused by refractive index difference is effectively counteracted, wavefront errors are reduced by 58%, the thickness of a machining damage layer is reduced by 70%, the method is suitable for various transparent materials, the machining precision and the material utilization rate are improved, and the method is suitable for high-precision laser machining scenes.

The invention discloses a cultured seaweed identification method and system based on time sequence remote sensing and morphological constraint, and the method comprises the steps: obtaining and preprocessing an original multispectral remote sensing image, forming a multispectral remote sensing image set, and generating a water body range mask; under the constraint of a water body range mask, synthesizing each time phase image, enhancing the texture through local maximum filtering, calculating a seaweed light absorption index, and generating a feature enhanced image; performing threshold segmentation on the feature enhanced image, extracting and vectorizing a potential seaweed pattern spot raster image, and calculating morphological indexes such as area and firmness of each vector pattern spot; and screening the vector pattern spots based on a preset area and a firmness threshold, and outputting a cultured seaweed distribution result. According to the method, the self-defined index and the morphological constraint are combined, the recognition precision of regular form targets such as culture rafts is effectively improved, area time sequence change analysis can be generated, and an automatic and high-precision technical scheme is provided for dynamic monitoring of culture seaweed resources.

The invention provides a water chlorophyll a concentration inversion method and system based on multi-modal data and a lightweight model, and relates to the technical field of water environment remote sensing evaluation. The method comprises the following steps: firstly, acquiring a Gaofeng No.5 satellite remote sensing image, a sentinel No.3 satellite image and ground actual measurement data, and completing image preprocessing and water body pixel extraction; constructing a hyperspectral index and an aquatic vegetation index, and fusing the hyperspectral index and the aquatic vegetation index with the water body temperature, the pH environmental factors and the spectral reflectivity to form a multi-dimensional feature sample set; a core feature subset is obtained through random forest and XGBoost coupling feature selection, and a lightweight student model is trained based on knowledge distillation; and constructing a to-be-predicted feature sample for the to-be-predicted time phase image and the environment factor, inputting the to-be-predicted feature sample into the lightweight student model to obtain a chlorophyll a concentration predicted value, and generating a spatial distribution map and a quality control map layer. According to the invention, high-precision, low-redundancy and efficient deployment chlorophyll a concentration inversion is realized.

The invention relates to the technical field of minimally invasive ablation, in particular to a liver tumor microwave ablation postoperative contraction deformation prediction method, which comprises the following steps: performing unified standardization processing on different periods of liver MRI images; accurately marking a preoperative liver tumor area, a postoperative ablation area and a postoperative liver tumor ghost area of the liver MRI image to obtain masks of the corresponding areas; designing a distance perception function as an attention parameter of different areas of the liver MRI image; constructing a multi-sequence distance guide complementary network model; selecting a loss function; selecting an elastic registration method to perform pre-operation and post-operation registration; and calculating an ablation safety boundary through the ablation area mask and the adjusted tumor mask. The liver tumor microwave ablation postoperative curative effect evaluation accuracy can be effectively improved, and a foundation is laid for formulating a subsequent treatment plan of a patient.

The invention relates to the technical field of optical precision interference measurement, in particular to a correction method for projection distortion detection of a primary mirror of a dynamic interferometer and a related product, and the method comprises the steps: obtaining a first coordinate and a second coordinate; determining a zero distortion point; obtaining an ideal imaging point set and an actual mapping point set after system projection through ray tracing; fitting to obtain a distortion function representing the distortion of the optical system; performing point-by-point correction on a phase diagram obtained by interference measurement, mapping pixel coordinates of the phase diagram into workpiece coordinates, obtaining interference measurement data corresponding to the actual surface of the primary mirror, and completing distortion correction; according to the method for correcting the surface shape projection distortion during interference measurement of the high-precision primary mirror of the large-aperture dynamic interferometer, the problem that in the process of assembling, repairing and developing the large-aperture Twyman type dynamic interferometer, the projection distortion effect of a detection system causes the loss of an accurate matching relation between a workpiece coordinate system and a detection coordinate system is solved.

The invention provides a random forest shallow sea sediment classification method based on multi-temporal remote sensing image fusion, and relates to the technical field of sediment information extraction. Comprising the following steps: 1, collecting and preprocessing multi-temporal image data to obtain remote sensing reflectivity; 2, the water depth of each single-time-phase image is inverted, and the optimal water depth is obtained; 3, calculating bottom reflectivity characteristics of blue and green wave bands based on the optimal remote sensing image; 4, respectively calculating topographic features and spectral features based on the optimal water depth and the optimal remote sensing image; and 5, in combination with the bottom reflectivity features, the topographic features and the spectral features, carrying out random forest feature optimization and classification model training, and generating a substrate classification result. On the basis, the method solves the problems that an existing remote sensing image substrate classification method is insufficient in feature consideration, noise in a single-time-phase image can cause low classification precision, and therefore negative effects can be generated on accurate acquisition of substrate information.

The invention discloses a logic chip mask defect detection method, and relates to the technical field of integrated circuit manufacturing, and the method comprises the following steps: S1, constructing a scattering response model of sub-wavelength inorganic particles on the surface of a mask, and constructing a scattering response model of sub-wavelength inorganic particles on the surface of the mask under a set multi-angle polarized light incidence condition; non-linear disturbance influence of inorganic particles with different particle sizes on the reflected light phase in each incident polarization state is simulated, and a scattering response template containing standard phase disturbance characteristics is generated; and S2, based on a scattering response template, carrying out pixel-by-pixel matching analysis on the acquired multi-angle polarization phase diagram on the surface of the mask, identifying a local area matched with the template, and marking the local area as a suspected scattering interference area. According to the method, through construction of the scattering response template and multi-angle polarization consistency analysis, accurate distinguishing of artifacts and real defects is achieved, the recognition precision and stability under complex interference are improved in combination with structure retention type phase reconstruction and dynamic threshold recognition, and mask quality control and photoetching yield improvement are facilitated.

The invention provides a structured light projector calibration method and system and a medium, and the method comprises the steps: projecting a multi-frequency stripe pattern to a target region based on an MEMS micromirror array module, obtaining a distorted image sequence, and carrying out the phase decoding, and obtaining a phase diagram corresponding to each frequency; resolving the phase diagram based on a phase shift method and an optical path triangulation principle to obtain a preliminary three-dimensional point coordinate; constructing a neural network model and a calibration sample set, obtaining a real three-dimensional coordinate by using a standard calibration plate or a known curved surface, and calculating the deviation between the initial three-dimensional point coordinate and the real three-dimensional coordinate to obtain a training sample set; and training the neural network model according to a training sample set based on an Euclidean distance loss function, outputting a three-dimensional offset compensation amount according to the neural network model, correcting the initial three-dimensional point coordinates, and realizing an end-to-end error self-compensation mechanism through the neural network model, so that deviation compensation is effectively performed on the three-dimensional point coordinates, and the accuracy of the three-dimensional point coordinates is improved. And the three-dimensional reconstruction precision is improved.

The invention relates to the field of image processors and biological medicines, and discloses an NK cell activity rapid detection method based on image processing. The method comprises the following steps: acquiring an unmarked time sequence phase image sequence of an NK cell and target cell co-culture system; performing cell instance segmentation to track individual cells; extracting a morphological dynamic characteristic parameter set of the target cell, wherein the morphological dynamic characteristic parameter set comprises a volume change rate, a phase gradient entropy, a cytoplasm phase fluctuation frequency and a nuclear region phase mean value; inputting the parameters into a pre-trained death state discrimination model, and outputting a death probability; and calculating a killing efficiency index based on the death probability evolution curve, and judging the activity level of the NK cells. The system comprises a phase image acquisition unit, a cell segmentation unit, a feature extraction unit, a death judgment unit and an activity judgment unit. Through unmarked imaging and deep learning fusion analysis, high-precision, real-time, quantitative and ultra-early NK cell activity evaluation is realized, and the method is suitable for clinical instant inspection and immunotherapy monitoring.

The invention relates to the technical field of flaw detection, and discloses a thermal insulation decorative plate appearance flaw detection method and system based on image recognition, and the method comprises the steps: executing optical excitation operation on the surface of a thermal insulation decorative plate through a periodic light pulse; acquiring an infrared image time sequence corresponding to the surface of the thermal insulation decorative plate according to the transient optical response sequence, and performing time sequence analysis on each pixel point in the infrared image time sequence; converting the gray time sequence change curve into a frequency domain change curve, and calculating the optical phase lag amount of each pixel point according to the frequency domain change curve; performing pixel-level comparison according to the optical phase lag and a reference phase diagram corresponding to the target thermal insulation decoration sample, generating a phase difference distribution diagram, and identifying a connected region with abnormal phase in the phase difference distribution diagram; recognizing surface flaws of the thermal insulation decorative plate according to the communication area, and determining a flaw detection result of the thermal insulation decorative plate according to the surface flaws. According to the invention, the accuracy of appearance flaw detection of the thermal insulation decorative plate can be improved.

The invention discloses a remote sensing semantic change detection method, and the specific process is as follows: in a dual-temporal VSS-Mama encoder, executing the four-stage feature extraction of a dual-temporal image pair, and outputting a dual-temporal feature map; cascading the double-time-phase characteristic patterns according to a channel of a change decoder, and respectively placing the double-time-phase characteristic patterns into an STSS module at four stages of the change decoder to carry out learning cross-time-sequence interaction and output a binary change pattern; a parallel bilingual decoder is adopted to execute step-by-step feature up-sampling on the dual-time-phase feature maps and output bilingual change maps; repeating the training until convergence and obtaining an optimal weight file; generating a semantic change graph by using the binary change graph, the bilingual change graph and the optimal weight file; according to the remote sensing semantic change detection method, a twinning network architecture of encoder-change decoder-bilingual semantic decoder is constructed, so that global space-time modeling and local detail enhancement are complementary, and the problems that the remote sensing semantic change detection method is low in long-range dependence modeling efficiency, broken in edge continuity, easy to ignore subtle change, high in reasoning overhead and the like are solved.

A method for generating a training dataset for determining grain boundaries and minerals in a thin section of a rock sample, includes receiving the thin section of the rock sample, generating optical images of the thin section with an optical tool, generating mineral phase images of the thin section with an electron microscopy tool, computing first and second pseudo-images based on different features extracted from the optical images, generating the training dataset based on (1) the optical images, (2) the mineral phase images, and (3) the pseudo-images, and training a single deep neural network, DNN, based on the training dataset to simultaneously determine a mineral type and grain boundaries in the thin section of the rock sample.

The invention discloses a building change detection method fusing building explicit prior and multi-stage feature aggregation, and belongs to the technical field of image processing. The method comprises the steps of obtaining a double-time-phase image at a to-be-detected position, inputting the obtained double-time-phase image into a trained building change detection network to obtain a changed building prediction map of the double-time-phase image, and determining a building change result according to the changed building prediction map. According to the method, in the feature extraction stage, the explicit prior of the edge of the building is introduced, and the edge information extraction capacity of the network is enhanced. Meanwhile, in the feature fusion stage, interaction of local-global information is promoted in a multi-level feature fusion mode, and the reusability of the multi-scale feature map is enhanced. According to the method, the edge and corner information is utilized again at the output end to enhance the optimization capability of the network on the changed building edge, and the building change detection precision is jointly improved.

The invention provides a Flash wafer detection method and device based on deep learning, and relates to the technical field of semiconductor detection. The method comprises the following steps: firstly, receiving an initial electric signal response sequence which is acquired by a scanning probe and comprises reflection intensity and phase deviation; performing time-frequency domain joint transformation processing on the sequence to generate a frequency domain energy distribution map and a time domain attenuation characteristic curve; further constructing a three-dimensional space mapping model containing a frequency domain energy amplitude and a time domain attenuation time constant; a pre-trained condition generation network is called to reconstruct the model, and a reconstructed phase image set with the resolution consistent with that of a standard template is generated; and finally, performing pixel-by-pixel comparison on the reconstructed image and a standard template to generate a detection report containing defect space positioning coordinates and a contour boundary sequence. According to the method, deep mining and visual reconstruction of Flash wafer microscopic electric signal features are realized, and the accuracy and the automation degree of defect detection are remarkably improved.

The invention relates to the field of image recognition, discloses an image recognition method and an image recognition system, and systematically solves the core pain point of traditional document recognition through optical-topology fusion processing and a dynamic resource allocation mechanism. The image recognition method is composed of an acquisition module, a grid module, a phase module, a setting module and a distribution module, pixel brightness is calculated and coded based on document RGB data, and a two-dimensional coding matrix is generated; constructing a geometric correction grid, forming a nonlinear constraint field, and enhancing the anti-deformation capability of the image; generating spiral phase light waves in the constraint field by using a spatial light modulator, and generating a time-varying phase map; determining a local topology index by detecting the number of phase jump times; and extracting the closed boundary region as a character block, and generating an analysis result file. The system breaks through traditional limitation, improves image geometric correction precision, feature extraction sensitivity and boundary judgment accuracy, efficiently completes document analysis, and is suitable for scenes such as document digitization and information retrieval.

The invention provides a dual-temporal image change detection method, system and device based on multi-scale feature fusion and an attention mechanism, and a medium, and relates to the field of computer vision and remote sensing image processing, and the method comprises the steps: taking a dual-temporal remote sensing image semantic change detection data set as original remote sensing image data; preprocessing the original remote sensing image data; extracting multi-scale features of the processed remote sensing image through a double-branch feature extraction network, wherein the double-branch feature extraction network adopts a ResNet architecture as a backbone network; a cross-time-phase feature fusion module is adopted to fuse multi-scale features of different time phases, a multi-stage feature decoding module is adopted to perform up-sampling and scale fusion on the fused features, and a final remote sensing image change detection probability graph is output after decoding operation. According to the invention, by introducing technical means such as deep learning, multi-scale feature extraction and a self-attention mechanism, efficient detection of a multi-time remote sensing image change area is realized, and the detection precision is improved.

The invention discloses a zero sample anomaly detection method and device based on multi-mode prompt learning, and the method comprises the steps: carrying out the modeling of an image through a frequency domain dynamic prompt module, obtaining a frequency spectrum feature, and carrying out the fusion of the frequency spectrum feature and a multi-mode prompt learning template, and obtaining a text prompt; inputting the text prompt into an encoder to obtain an initial text feature; the cross attention guidance anomaly graph generation module updates the initial text features to obtain target text features, and calculates the target text features and the stage image features to obtain an anomaly graph; the double-branch pooling attention module pools the staged image features to obtain local image features; and adding the local image feature and the category representation feature to obtain a global image feature, and calculating the global image feature and the initial text feature to obtain an abnormal score. The method can reduce the dependence of the model on fixed prompts, enhances the discrimination capability of different abnormal modes, improves the precision of abnormal positioning, and can be widely applied to the technical field of computers.

The invention relates to the technical field of image processing, and discloses a laser holographic aberration compensation system and method based on deep learning, and the system comprises a first module which determines a phase-to-gray lookup table; the second module is used for calculating to obtain stacking strength; the third module is used for obtaining a Zernike polynomial coefficient through a convolutional neural network regression device; the fourth module is used for calculating to obtain predicted intensity and an aberration estimator; the fifth module is used for outputting two grey-scale maps; and the sixth module outputs two phase diagrams. According to the method, self-supervised training data is constructed through three-plane intensity collection, and an angular spectrum method physical model and Zernike polynomial coefficient low-dimensional representation are combined, so that a convolutional neural network regression device can stably learn aberration mapping, and compensation precision and generalization ability are both considered; double-phase two-frame time division coding is adopted, the output characteristics of a phase type spatial light modulator can be adapted, and cooperative control of aberration compensation and target complex field reproduction is achieved.

The invention provides a switch cabinet fault identification method and system based on a reconstruction modal phase image, and belongs to the technical field of intelligent fault diagnosis of a power system, and the method comprises the steps: dividing the operation data of a switch cabinet into a plurality of modal signals based on the operation data of the switch cabinet collected in real time; based on a preset reconstruction threshold, reconstructing the plurality of modal signals into high-frequency signals and low-frequency signals; constructing a two-dimensional trend confrontation image based on the high-frequency signal and the low-frequency signal; and classifying fault features in the two-dimensional trend confrontation image based on a preset neural network algorithm to identify the switch cabinet fault. The method solves the problem of poor fault identification reliability caused by the fact that fault type judgment cannot be completed independently in the prior art. According to the method, the real-time monitored data is reconstructed, and the two-dimensional graph highlighting the feature redundant information is generated, so that the fault recognition accuracy is improved.

The invention discloses a variable view field Linnik interference microscopic system and a measurement method thereof, and belongs to the field of laser measurement instruments, and a variable diaphragm is introduced into a reference arm of an interference system to realize matching of objective lenses of different test arms, so that light field spatial correlation under speckle illumination is maintained; meanwhile, a light source system with a delay line structure is adopted to accurately compensate the optical path difference introduced by the asymmetric objective lens configuration, so that the acquisition of a high-contrast interference pattern is realized. The method can effectively improve the quality of an interference image, has a depth selection function, and is suitable for dynamic phase imaging in different view field ranges. The technical scheme of the invention shows good imaging effect and adaptability in the aspects of dynamic imaging and depth layering.

The invention discloses a high-resolution optical remote sensing image building change detection method, system and equipment based on texture frequency domain perception and a medium, and the method comprises the steps: obtaining a public building change detection data set LEVIR-CD which comprises double-time-phase images T1 and T2, cutting the data set into non-overlapping image pairs, and dividing the image pairs into a training set, a verification set and a test set according to a proportion; constructing a texture frequency domain sensing network, wherein the texture frequency domain sensing network comprises a twin MIT-B0 encoder, a texture sensing frequency domain attention module and a multi-layer perceptron decoder; training the texture frequency domain sensing network; performing result prediction on the test set by using the trained texture frequency domain sensing network to obtain a pixel-level prediction result; performing evaluation index calculation on each category and overall quality of the pixel-level prediction result, and evaluating network change detection performance; systems, devices, and media for implementing the method; according to the method, the precision and reliability of building change detection are effectively improved, and more reliable technical support is provided for application in related fields.

The invention provides an image phase unwrapping method and device. The method comprises the steps of obtaining a to-be-unwrapped phase image and performing normalization processing; performing phase average value calculation on the normalized phase image to be unwrapped to obtain a column average value array of phase average values of each column of pixels; phase change trend judgment is carried out according to the column mean value array, and a phase change trend judgment result of the to-be-unwrapped phase image is obtained; zero point detection and verification are carried out according to the column mean value array, and a true zero point in the phase image to be unwrapped is determined; determining a phase jump boundary in the normalized phase image to be unwrapped according to the phase change trend judgment result and the true zero point; performing region division on the normalized phase image to be unwrapped according to the phase jump boundary and the true zero point, and performing phase compensation on each region to obtain an unwrapped image; the method can reduce the calculation cost.

The invention relates to the technical field of metal surface coating preparation, and discloses a cladding layer surface and deep defect detection method based on pulse laser, which comprises the following steps: clamping and fixing a workpiece; pulse laser excitation and bimodal signal synchronous excitation are carried out; synchronously acquiring and recording dual-channel signals; thermal wave signal processing and surface / near-surface flaw analysis; performing ultrasonic signal processing and deep defect analysis; and carrying out information fusion and three-dimensional flaw reconstruction. According to the invention, the pulse laser beam is utilized to excite the thermal wave and the ultrasonic wave at the same time, various flaws from the surface to the deep layer can be covered through one-time detection, the limitation of a single detection mode is solved, and meanwhile, the depth information of the flaws can be obtained by analyzing the phase diagram of the thermal wave signal; three-dimensional distribution of flaws can be reconstructed by combining ultrasonic time-of-flight (ToF) analysis and infrared sequence images, and accurate positioning and quantitative evaluation are realized.

The invention discloses a light field display method and system with holography as a representation medium, and the method comprises the following steps: S1, collecting a multi-view image of a target scene, carrying out the coding of the multi-view image, and generating a light field tensor containing space and angle information; s2, mapping the light field tensor in the step S1 into a two-dimensional space phase diagram; s3, converting the two-dimensional space phase diagram in the step S2 into a modulation diagram used for driving a spatial light modulator; and S4, driving the spatial light modulator to perform light wavefront modulation according to the modulation graph in the step S3, and reproducing the image of the target scene through optical diffraction. According to the invention, integration from dynamic scene information acquisition to high-quality three-dimensional image reproduction is realized, a low-redundancy and high-fidelity processing link is realized, and the efficiency and quality of holographic display are greatly improved.

The invention discloses a phase unwrapping method for an SAR (Synthetic Aperture Radar) interferometric phase diagram. The phase unwrapping method comprises the following steps: acquiring prior information of the interferometric phase diagram; based on prior information, unwrapping the interferometric phase diagram to obtain an absolute phase; elevation information is extracted according to the absolute phase; according to the elevation information, the prior information is reversely optimized. The acquisition of the prior information is based on image segmentation of the interferometric phase diagram. According to the phase unwrapping method for the SAR interferometric phase diagram and the related product, all segmented areas are executed one by one, all reconstruction results are further geocoded and integrated, fine three-dimensional reconstruction of a building can be achieved, target extraction is conducted by fusing structural features in the two-dimensional imaging direction, and the construction efficiency is improved. And the accuracy of building identification is obviously improved. The structure auxiliary information of image segmentation effectively reduces the risk of phase discontinuity in the phase unwrapping process, significantly reduces the unwrapping error, and achieves the high-precision and low-cost three-dimensional reconstruction of a building under the condition of a double-track long baseline.

The invention discloses a laser shearing speckle interference phase unwrapping method based on deep learning. According to the method, eight speckle patterns before and after deformation are used as input, multi-scale feature extraction is carried out through layer-by-layer convolution and down-sampling of an encoder, features are sent to an attention fusion module, response weights of channels and spatial positions are adaptively adjusted through channel attention and position attention, then the response weights are input into a decoder, and jump connection with an encoding end is combined, so that multi-scale feature fusion is realized. And step-by-step reconstruction of multi-scale features is realized, and a phase diagram is output. Performing comprehensive constraint on prediction and reference phases in the aspects of numerical deviation, structural consistency and gradient smoothness and updating network parameters by adopting composite loss formed by mean square error, mean absolute error, structural similarity, gradient loss and out-of-plane displacement calculation items; therefore, an unwrapping result which is globally continuous and has clear and stable local details is obtained, and a feasible way is provided for intelligent processing and automatic analysis of the laser shearing speckle interference image.

The invention discloses a building remote sensing image change detection method based on a twin network and an attention mechanism, and the method comprises the steps: extracting a feature sequence of a building remote sensing image at the same place at different times through employing the twin network, and adding the attention mechanism of a spatial dimension to obtain a multi-scale fusion feature. The method comprises the following steps: firstly, preprocessing a dual-time-phase remote sensing image and constructing a data set; then, an improved twin neural network model is constructed, the model embeds attention mechanisms in an encoder and a decoder, and multi-scale features of the dual-time-phase image are fused; training the model by adopting a mixed loss function; and finally, carrying out change detection by utilizing the trained model and outputting a result. According to the method, feature extraction consistency is guaranteed through the twin network, attention to a building change area is enhanced by using an attention mechanism, large-scale and detail features are highlighted by using feature fusion, the precision and robustness of change detection are effectively improved, and the method is suitable for building change monitoring in a complex environment.

The invention discloses an object level change detection method based on semantic flow feature alignment and global attention fusion, and belongs to the field of change detection. The method comprises the following steps: inputting a double-time-phase image and carrying out preprocessing; constructing a twin feature encoder to extract multi-scale features; a self-global cross attention module is introduced in a coding stage to realize global semantic fusion and coarse registration of double-temporal image features; a semantic flow alignment module is further designed, and pixel-level alignment of a feature level is realized by generating a semantic flow field, so that view angle difference and structure offset are compensated; in the decoding stage, a feature optimization and jump connection fusion mechanism is adopted, shallow spatial information and deep semantic features are integrated, and boundary expression of a change region is enhanced; and finally, outputting an object level change detection result through a detection head. Experiments show that the method has object-level change detection capability in scenes with visual angle differences, weak textures and the like.

The invention discloses a visual defect detection method, device and equipment based on knowledge prompt and a storage medium, and relates to the technical field of appearance defect detection.The method comprises the steps that multi-dimensional image data corresponding to a to-be-detected object is collected, and the multi-dimensional image data comprises a multi-band image, a multi-polarization image and an active structure light phase image; generating knowledge prompt information of the to-be-detected object according to a preset knowledge graph; and performing fusion defect detection based on the multi-dimensional image data and the knowledge prompt information to obtain a defect detection result. According to the method, the comprehensive information of the to-be-detected object is obtained from different dimensions by collecting the multi-dimensional image data, semantic guidance is provided by combining the knowledge prompt information generated by the knowledge graph, cross-modal fusion defect detection of the to-be-detected object is achieved, and therefore the visual defect detection precision is improved.