14 results about "Local statistics" patented technology

The invention provides a small-sample defect detection method and system based on a visual language large model, and relates to the technical field of computer vision, and the method comprises the steps: obtaining an original image of a to-be-detected workpiece, and extracting a first feature; based on the image and / or prior information, semantic prompts are generated by utilizing a visual language large model and are converted into second features capable of being aligned with the first features, and cross-modal fusion is carried out to obtain fusion features; in a candidate position neighborhood, subspace projection is performed on the fusion features, anisotropy measurement is determined based on local statistics, end member number estimation and intra-packet or extra-packet discrimination are performed under the measurement, and an active end member number is determined in combination with a cross-scale stability criterion; a decoding channel and a threshold value are adjusted according to the number of the active end members, and a result is fed back to the fusion process; decoding the adjusted fusion features, and outputting a pixel-level defect segmentation image; therefore, the positioning precision and segmentation stability of weak feature defects are improved under the condition that defect samples are limited.

The application discloses a strip steel hot rolling process multi-performance index abnormality monitoring method and system, and belongs to the technical field of industrial process control and fault diagnosis, and the method comprises the following steps: collecting process variable data and performance index data in a historical strip steel hot rolling process and performing standardization pretreatment; dividing the strip steel hot rolling process into multiple different process subblocks; for each process subblock, performing space-time feature extraction on the corresponding process variable data to obtain a space-time feature representation; constructing a local statistic quantity of each performance index; fusing the local statistic quantities to obtain a global statistic quantity and a global control limit; and based on this, realizing multi-performance index abnormality monitoring and alarm. The application effectively solves the problems that cross-process time lag dependence is difficult to capture, space-time feature extraction is not comprehensive, and single performance index monitoring leads to abnormality missed reports in the strip steel hot rolling process, significantly improves the fault detection rate and reduces the false alarm rate.

The application provides a visual mapping method in a low-light environment, and relates to the technical field of visual mapping. In the application, a brightness channel is separated from an input image and is divided into a plurality of tiles, local statistics of each tile are calculated to adaptively generate a contrast limiting parameter and an enhancement parameter vector, the brightness channel is enhanced based on the parameter, candidate points corresponding to feature points of an existing map are extracted after reconstruction of the image, a comprehensive confidence score is calculated, whether the candidate points and the corresponding map points are below a dynamic threshold in terms of the distance of the enhancement parameter vector is judged, and the confidence score is combined to determine whether the candidate points are adopted as new map points. Finally, the system dynamically optimizes the enhancement parameter according to the matching survival rate of feature points of each tile, continuously improves the mapping quality and stability through iterative updating, and realizes robust visual mapping in a low-light environment.

The application discloses a kind of based on deep learning electric power load multi-period prediction method and system, it utilizes instance-level reversible distribution adaptation mechanism, and the original tensor is stationary mapping in input end dynamic extraction local statistics, convert time-varying distribution into unified feature space that model is easy to converge.On this basis, design time-frequency double-flow interactive network, and extract local mutation feature in time domain and global periodicity feature in frequency domain in parallel, and compensate the information missing of single perspective through cross-domain gate fusion mechanism.Finally, cooperate with future-oriented generative decoder and distribution recovery module, and the stationary prediction result is reversely mapped back to the original dimension space with physical meaning.Through this closed-loop logic of stationary reasoning first and distribution restoration later, effectively eliminate the negative influence of statistical drift on model generalization, realize high-precision multi-period prediction under the premise of retaining environment-driven physical characteristics.

The invention discloses a road ecological landscape dynamic classification method and system based on landscape visual features, and relates to the field of artificial intelligence, and the method comprises the steps: constructing a data set containing image-level and pixel-level double-layer annotations; preprocessing the image through self-adaptive illumination-color normalization in combination with local statistics and global estimation; carrying out image block division and recombination by adopting semantic guidance; constructing an end-to-end classification model, deeply fusing visual and semantic features, modeling a spatial layout and semantic similarity relationship, and realizing landscape type discrimination; and utilizing the training model to automatically classify the new image. According to the method, the robustness to complex illumination can be effectively improved, structured image blocks rich in semantics are formed, multi-scale feature interaction and discrimination are enhanced, and feature representation consistent in semantics is learned.

This invention provides a method and system for identifying underground structure boundaries based on distributed optical fiber sensing, relating to the field of underground structure detection and boundary identification technology. The invention first deploys a sensor array in the area to be measured and establishes a correspondence between channels and spatial coordinates, acquiring an unexcited baseline; through low-energy probing to estimate the medium and noise, the excitation frequency band, energy, and repetition rate are optimized, emitted by an adaptive seismic source; reflection and transmission responses are synchronously acquired under a unified time reference; wavelet denoising, temperature and dispersion compensation, and time alignment are applied to the data to extract amplitude, phase, and frequency features; gradients are calculated on the feature field and non-maximum suppression is performed; a stable boundary point is obtained by employing an adaptive double threshold based on local statistics combined with temporal continuity and spatial connectivity constraints; a two-dimensional profile is obtained by spline fitting, and a three-dimensional model is reconstructed under multi-profile consistency constraints, outputting the profile diagram and the three-dimensional model, achieving high-precision, real-time, and visualized detection of small-sized boundaries.

The application relates to an intelligent detection method and system for uneven settlement of a super-high building foundation, and belongs to the technical field of settlement detection. The method comprises the following steps: collecting settlement monitoring data and constructing a training data set; calculating local statistics according to the spatial topological relationship of monitoring points, combining adjacent point set standardized original data to obtain standardized settlement values; amplifying the contribution of the mutation area by a second-order difference algorithm curvature mutation factor and a local variance weight factor, splicing the standardized settlement characteristic vector and the curvature mutation factor to obtain an enhanced characteristic vector; constructing an uneven settlement detection model, inputting the enhanced characteristic vector into the model to obtain an uneven settlement probability and a local anomaly heat map vector; optimizing the training process of the model through a loss function to obtain a trained model; and inputting the processed real-time settlement monitoring data collected newly into the trained model to obtain an intelligent detection result. The application can improve the detection capability of the uneven settlement of the foundation.

The application discloses a kind of narrow river detection methods suitable for spaceborne wide swath interferometric radar altimeter image, comprising: reading into spaceborne wide swath interferometric radar altimeter image;It is enhanced river linear feature by Gaussian difference preprocessing to suppress background;Curvature structure perception detector is constructed, curvature response map is calculated based on Hessian matrix eigenvalue, and feature fusion enhancement is carried out in conjunction with multi-direction structure consistency score;The binary image obtained by adaptive threshold segmentation based on local statistics is used as region label by carrying out adaptive threshold segmentation to enhanced feature map, maximum value is extracted in the region corresponding to enhanced feature map, seed point set is formed, and adaptive region growth based on queue priority is executed, to generate initial river center line;Markov random field model is introduced, global structure optimization and topological connectivity correction are carried out using graph cut algorithm, to obtain optimized river center line;Direction and radiation characteristic constraint is applied again, morphological dilation is carried out, and finally complete river mask is generated.

The invention relates to a splicing seam repairing method and device, electronic equipment and a readable medium, and the method comprises the steps: obtaining an original DR image collected by DR equipment, and determining a first candidate region corresponding to the position of a grid of the DR equipment on the original DR image; performing multi-algorithm joint calculation on the original DR image based on the first candidate region to determine a splicing seam region in the first candidate region; performing local statistics on the splicing seam region and the neighborhood thereof to obtain local gray feature information, and taking the local gray feature information as a dynamic threshold constraint; and under the constraint of a dynamic threshold value, a weighted diffusion equation is constructed based on the main structure direction and the diffusion weight determined by a structure tensor, iterative repair is carried out on the splicing seam area according to the weighted diffusion equation, and the structure tensor is obtained by calculating gradient information of the splicing seam area and a neighborhood of the splicing seam area. The problem of how to ensure the structure continuity and the detail integrity while repairing the splicing seam of the DR image is solved.

This invention provides a method for suppressing and enhancing infrared image defects and thermal noise, relating to the field of image enhancement. The method aims to construct thermal target protection weights and data fidelity weights by normalizing, performing local statistics, and estimating gradient and fringe principal directions on the original infrared image. The observed image is decomposed into structural components, point-like defective pixel components, structured thermal noise components, and low-frequency thermal bias components. Through a thermal consistency-guided defective pixel weight generation mechanism, direction-aware structured thermal noise separation, background-gated low-frequency thermal bias separation, anisotropic edge-preserving enhancement for thermal target protection, and multi-scale local thermal contrast preservation, a joint optimization solution is achieved, and the enhanced image is output after local thermal contrast remapping. This method effectively suppresses defective pixels, fringe noise, and background thermal bias, reduces the risk of accidental deletion of small thermal targets, maintains the target outline and the thermal difference between the target and the background, and improves the clarity, recognizability, and engineering applicability of infrared images.

A collaborative learning framework is presented. The collaborative framework is implemented by multiple network nodes interconnected by a network. The network nodes belong to multiple client systems of the framework. A network node belonging to a first client system constructs a predictive model for the first client system by using a pattern dictionary that is a built based on a consensus among the multiple client systems. The network node calculates a set of local statistics for the first client system based on raw data of the first client system. The network node computes a consensus set of local statistics by aggregating sets of local statistics from the multiple client systems. The network node updates the pattern dictionary based on current values of the pattern dictionary and the consensus set of local statistics.

The invention discloses a narrow river detection method suitable for a satellite-borne wide-cradling interference radar altimeter image. The method comprises the following steps: reading in the satellite-borne wide-cradling interference radar altimeter image; gaussian difference preprocessing is carried out to suppress background and enhance the linear features of the river channel; constructing a curvature structure sensing detector, calculating a curvature response diagram based on a Hessian matrix eigenvalue, and performing feature fusion enhancement in combination with a multi-direction structure consistency score; performing self-adaptive threshold segmentation based on local statistics on the enhanced feature map to obtain a second map as a region mark, extracting a maximum value in a corresponding region of the enhanced feature map, forming a seed point set, executing self-adaptive region growth based on queue priority, and generating an initial river channel center line; introducing a Markov random field model, and performing global structure optimization and topological connectivity correction by using a graph cut algorithm to obtain an optimized river channel center line; and then applying direction and radiation characteristic constraints, performing morphological expansion, and generating a final complete riverway mask.

This application provides a method and system for few-sample defect detection based on a large visual language model, relating to the field of computer vision technology. The method includes: acquiring the original image of the workpiece to be inspected and extracting a first feature; generating semantic cues based on the image and / or prior information using a large visual language model, and converting them into second features that can be aligned with the first feature, and performing cross-modal fusion to obtain fused features; performing subspace projection on the fused features in the neighborhood of candidate locations, determining anisotropy measures based on local statistics, and estimating the number of endmembers, distinguishing between in-bag and out-of-bag operations under this measure, and determining the number of active endmembers in conjunction with cross-scale stability criteria; adjusting the decoding channels and thresholds according to the number of active endmembers, and feeding the results back to the fusion process; decoding the adjusted fused features and outputting a pixel-level defect segmentation map; thus improving the localization accuracy and segmentation stability of weak feature defects under the condition of limited defect samples.

The invention discloses a community event data processing method and system based on a double-layer Dccu architecture, and belongs to the technical field of smart communities and grassroots governance, and the method comprises the steps: a cell Dccu server deployed in a cell collects data, and generates a trusted event through the recognition of a built-in AI model; a community Dccu server deployed in a community receives the trusted event through a private network, carries out local statistics and non-cloud processing, generates a report form and a work order and issues the report form and the work order; and finally, receiving a disposal result to realize closed-loop management. The system comprises the cell Dccu server and the community Dccu server. According to the invention, through a double-layer cooperative processing architecture of a cell and a community, in combination with private network transmission and local AI processing, full-process closed-loop management of event data from identification, generation, transmission, disposal to feedback is realized, and the real-time performance of event processing, the data credibility and the grassroots governance efficiency are effectively improved.