14results about How to "Avoid error propagation" patented technology

This invention provides a depth prior-guided multi-view stereo 3D reconstruction method. It acquires N original images of the target to be reconstructed from different perspectives; it iterates through the N original images, using each image as a reference image and the remaining N-1 images as source images, inputting them into a trained depth map generation network to obtain N depth maps; and then fuses these N depth maps to obtain a 3D point cloud of the target to be reconstructed. Based on its network architecture, the subsequent depth map generation module uses the depth maps output by the preceding depth map generation module to generate a depth prior. This depth prior guides the Transformer in the subsequent depth map generation module to complete cost volume aggregation, generating the final depth map, thereby improving the final quality of the depth map and point cloud.

This invention discloses a multidimensional indexed modulation chaotic communication method and apparatus based on neural network-assisted detection, belonging to the field of wireless communication and chaotic spread spectrum communication technology. The method includes: firstly, segmenting the input bitstream to obtain modulation bits, domain index bits, codeword index bits, and reference index bits; then, selecting the corresponding transform domain codebook based on the domain index and the corresponding codeword basis vector based on the codeword index, while simultaneously generating an orthogonal chaotic sequence pool and selecting the corresponding reference chaotic sequence; next, performing MPSK modulation on the modulation bits, and combining the codeword basis vector and the reference chaotic sequence to generate a multidimensional indexed modulation chaotic signal, which is then transmitted through a wireless channel after multi-carrier modulation; the receiving end demodulates, filters, and samples the received signal, and inputs it into a multi-task convolutional neural network detector to achieve joint detection of modulation symbols, domain index, codeword index, and reference index, thereby recovering the transmitted bitstream.

This invention discloses an orthogonal frequency division multiplexing (OFDM) transmission method based on dirty paper coding, belonging to the field of wireless communication technology. At the transmitting end, at least two layers of frequency domain symbols to be transmitted are generated. A first-layer OFDM signal is generated from the first-layer frequency domain symbols and non-negatively clipped in the time domain to generate first-layer clipped noise. The cumulative clipped noise from the first layer to the (h-1)th (h≥2)th layer is used as interference signal to perform dirty paper coding pre-coding on the h-th layer of the frequency domain symbols to be transmitted, generating pre-coded frequency domain symbols, and then generating the h-th layer OFDM signal. The OFDM signals of each layer are superimposed and transmitted. At the receiving end, after performing a fast Fourier transform on the received signal, the received frequency domain symbols of each layer are extracted from the received frequency domain signal and decoded independently. This invention eliminates deterministic interference between layered signals through dirty paper coding pre-coding at the transmitting end, eliminating the need for serial interference cancellation at the receiving end, avoiding error propagation problems, and significantly reducing the computational complexity at the receiving end.

The invention discloses a radar interference suppression and target identification integrated method and system, and the method comprises the steps: building an end-to-end complex valued neural network which is used for carrying out the radar target identification under radio frequency interference; training the network by using a non-interference and interference-containing complex-valued radar training data set, taking non-interference complex-valued radar image data and a target category as supervision tags, and learning anti-interference and target identification at the same time; the network structure comprises an anti-interference and target feature enhancement module and a target identification module. The anti-interference and target feature enhancement module is used for separating a target and radio frequency interference from the input complex-valued radar image, suppressing the radio frequency interference and highlighting the features of the target; and the target identification module is used for learning and extracting the features of the target from the output of the anti-interference and target feature enhancement module. The method can solve the problem that the recognition performance is reduced due to error accumulation in a cascade processing flow in the prior art, and is suitable for radar image target recognition tasks under radio frequency suppression interference.

The present application relates to the technical field of water turbine production equipment, and particularly relates to a stator shell machining method capable of quickly installing an iron core, comprising the following steps: one, a prefabrication machining step; two, a pre-positioning step; three, a tool installation and precision adjustment step; four, a measurement and adjustment step; and five, a deformation-resistant welding step, the present application is provided with a pre-positioning structure of a clamping groove-clamping block, and the initial placement of the dove tail rib can achieve high coarse positioning accuracy, the tedious repeated alignment and auxiliary tool building time in the traditional method are saved, the three-dimensional adjustable tool makes the radial and circumferential precision adjustment one-step, different tools do not need to be repeatedly loosened and tightened, preliminary estimation shows that the installation time of a single dove tail rib can be shortened by more than 50%, and the overall stator shell machining cycle is shortened by 30%.

This application provides a method and system for intelligent chromosome karyotype segmentation that integrates deep learning. The method involves: enhancing chromosome microscopic images; performing region segmentation based on the enhanced chromosome images to obtain multiple chromosome region sub-images; for any chromosome region sub-image, extracting the chromosome pixel area and band count corresponding to that sub-image, and determining the structural coverage index of the sub-image based on these values; comparing the structural coverage index with a preset chromosome banding distribution model to perform anomaly screening and identify candidate chromosome adhesion regions; and segmenting the candidate adhesion regions using a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image. This application can screen and identify candidate adhesion regions based on the chromosome skeleton banding density, improving the accuracy of chromosome image segmentation.

The invention discloses an amplitude-phase error self-calibration direction of arrival greedy estimation method. The method comprises the following steps: S1, establishing an array receiving signal model containing an amplitude-phase error; s2, calculating an effective grid size and an iteration stop threshold value of the search grid; s3, performing coarse search on the search grid, and detecting an initial direction of arrival and an initial signal corresponding to the maximum projection energy; if the maximum projection energy is lower than an iteration stop threshold, skipping to S6; s4, performing Newton iteration refinement on the initial direction of arrival to obtain more accurate continuous angle estimation and signal estimation; s5, alternately updating the signal waveform of the signal source and the diagonal calibration matrix, updating the residual signal according to the signal waveform and the diagonal calibration matrix, rejecting the current signal source, and returning to S3; and S6, removing a weak source through pruning, performing optimization, and outputting a final direction of arrival and diagonal calibration matrix estimation. According to the method, error transmission is avoided, and the direction-of-arrival estimation precision under the condition of unknown array errors is remarkably improved.

The application relates to the field of marine environment monitoring and optical sensing technology, and discloses a seawater surface layer temperature and salinity synchronous measurement system and a measurement method. The system comprises a light source module; an optical sensing module arranged at the seawater surface layer and directly contacted with seawater medium, which forms an output light field with statistical characteristics under the interaction of light and seawater; a light field sampling module which samples the intensity distribution of the output light field and obtains light field statistical data; and a data processing module which constructs a probability model of the light field intensity distribution based on the light field statistical data, extracts a statistical entropy characteristic quantity representing the overall statistical state of the output light field, and inversely obtains the seawater surface layer temperature and salinity by using the response relationship of the statistical entropy characteristic quantity with the seawater temperature and salinity. The application takes the overall statistical state of the light field as the sensing core, has the advantages of strong environmental disturbance resistance, universal structure, easy platform integration and the like, and is suitable for long-term in-situ monitoring of marine buoys, floating balls and the like platforms.

The invention discloses a sign language vocabulary segmentation and recognition method based on continuous integral triggering, which comprises the following steps: (1) acquiring a sign language video frame sequence, and extracting frame-level visual features to obtain a frame-level feature sequence; (2) inputting the frame-level feature sequence into a CIF module, generating a feature vector sequence corresponding to the vocabulary sequence by the CIF module through a continuous integral triggering method, and outputting a time division point corresponding to each vocabulary; (3) inputting the feature vector sequence into a vocabulary decoder, and outputting a vocabulary recognition result; and (4) aligning a vocabulary recognition result with the segmentation time point to form a vocabulary sequence with a timestamp, and outputting the vocabulary sequence. By utilizing the method, the problems of high labeling cost, two-stage error accumulation and lack of explicit segmentation capability of the existing end-to-end method in the traditional method can be effectively solved.

The invention discloses a legal map construction method based on multi-subject behavior recognition and role extraction, and belongs to the technical field of natural language processing and judicial intelligent analysis. The technical problems that multi-subject behavior semantics in criminal case texts are difficult to completely recognize, master-slave roles are inaccurate to judge, and behavior cross relations are difficult to extract are solved. According to the technical scheme, the method comprises the following steps of 1, data acquisition and preprocessing; step 2, semantic role labeling SRL; step 3, multi-level semantic feature extraction; 4, performing combined extraction on entities and relationships; step 5, behavior cross relation identification; and step 6, knowledge storage and application. According to the method, the behavior logic and role relationship of multiple subjects can be accurately described in a single case context, and efficient data support is provided for judicial semantic understanding, evidence chain construction and sentencing assistance.

The invention provides a lane number detection method and a lane number detection model training method and device, and relates to the technical field of artificial intelligence, in particular to the technical field of image processing. In the training process, a lane number detection model comprising an encoder, a decoder, a lane number detection branch and a detection quality confidence detection branch is constructed and trained, wherein the lane number detection branch and the detection quality confidence detection branch are arranged in parallel. In the reasoning stage, the lane number detection model obtained through training is called to process a to-be-processed road image, and a lane number detection result and a detection quality confidence score are obtained. The lane number detection model realizes the end-to-end generation of the lane number detection result, does not need intermediate steps, effectively avoids the error propagation of the intermediate steps, and improves the accuracy and precision of the detection of the lane number. The detection quality confidence score can jointly optimize model parameters in the training stage, the accuracy degree of the lane number detection model is improved, and a credible basis is provided for processing a lane number detection result in the reasoning stage.

The application discloses a kind of amplitude-phase error self-calibration's wave direction greedy estimation method, comprising: S1: the array received signal model including amplitude-phase error is established;S2: the effective grid size of search grid, iteration stop threshold is calculated;S3: coarse search is carried out on search grid, and the initial wave direction and initial signal corresponding to the maximum projection energy of detection are detected;If the maximum projection energy is lower than iteration stop threshold, then jump to S6;S4: Newton iteration refinement is carried out to initial wave direction, and more accurate continuous angle estimation and signal estimation are obtained;S5: source signal waveform and diagonal calibration matrix are alternately updated, and residual signal is updated accordingly, and current source is eliminated, and return S3;S6: weak source is eliminated and optimized by pruning, and the final wave direction and diagonal calibration matrix estimation are output.The application avoids error transmission, and significantly improves the wave direction estimation precision under unknown array error condition.

This application discloses an inertial integral positioning and downhole constraint correction method and system based on a miniature downhole sensor. The method includes: acquiring triaxial angular velocity and triaxial acceleration data from the miniature downhole sensor; recursively obtaining predicted position, predicted velocity, and predicted attitude based on quaternion attitude calculation and inertial integral algorithm; constructing a wellbore spatial constraint model based on known wellbore trajectory and wellbore diameter parameters, and constructing a fluid property constraint model based on downhole fluid property parameters; generating the two types of constraint correction quantities independently and without coupling; employing an error state Kalman filter framework to filter and update the two types of correction quantities as observation information, and injecting the error state into the nominal state to output a three-dimensional motion trajectory estimate. This method continuously corrects inertial positioning drift by introducing wellbore and fluid constraint information, achieving usable positioning results for the sensor even without external positioning signals.

This application provides a target-oriented crop plot-level classification method and device for fragmented farmland. The method includes: using a U-TAE encoder-decoder backbone network to extract multi-scale features from the temporal data of multi-temporal remote sensing image sequences in the input information; using a parallel spatiotemporal self-attention module to enhance the temporal and spatial features of the features output from the encoding stage of the backbone network; using a parameterless plot aggregation operator to aggregate the fused features into a plot-level representation; performing plot-level classification on the plot-level representation; and back-broadcasting the classification results to the corresponding pixel location of the plot based on the plot ID map. This application captures long-range dependencies through U-TAE and spatiotemporal attention, achieves end-to-end plot-level classification through parameterless plot aggregation and bi-branch prediction, eliminates salt-and-pepper noise while maintaining accurate boundaries, suppresses SAR speckle noise, and achieves efficient and accurate identification of fragmented farmland.