215results about How to "Solve the degradation problem" patented technology

The invention discloses a method for detecting and identifying continuous segmented texts in an image based on SegLink and Attention-based CRNN fusion processing, and belongs to the technical field ofoptical character recognition, aiming to solve the problems of low text detection accuracy, particularly low inclined text detection accuracy, difficulty in positioning, difficulty in font segmentation and low recognition accuracy in OCR information document digitization. The method includes the steps: establishing a SegLink + CRNN model based on a Tensorflow deep learning framework, detecting text lines in an image through a SegLink network; segmenting the segmented text according to lines; extracting single-line text features through a densely connected convolutional neural network; processing the sequence information of the context in the text by the bidirectional recurrent neural network, and adopting the CTC decoding algorithm to avoid the problem of single word segmentation, and eliminate the influence of the single word segmentation link on the recognition accuracy; and further fusing an Attention mechanism during CTC transcription to improve the recognition accuracy for the text sequence characteristics. The method is applicable to printed form and handwritten form recognition, and can be applied to recognition of multilingual texts such as English and Chinese.

The invention relates to a simultaneous localization and mapping method based on distributed edge unscented particle filter. First, a coordinate system is built and an environmental map is initialized; then subfilters are built for each landmark point with successful matching respectively; next, based on a robot motion model, a particle swarm is generated in each subfilter respectively, and the state vector and the variance of each particle are obtained; noise is introduced, particle state vectors after extension are calculated by utilization of unscented transformation, the particles after extension are updated and the particle swarms are optimized; then particle weights are calculated and normalization is carried out, and aggregated data of each subfilter are subjected to statistics and the data are sent to a master filter; next, global estimation and variance are calculated; then the effective sampling draw scale and sampling threshold of each subfiter are determined, the subfilters with severe particle degeneracy are subjected to resampling; then the state vectors and the variances of the robot are output, and stored in a map. Finally, landmark point states are updated by utilization of kalman filtering algorithm until the robot is no longer running.

The invention discloses a hyperspectral remote sensing image classification method based on a dense residual three-dimensional convolutional neural network. According to the method, original hyperspectral data are used as network input, three-dimensional spatial-spectral features of a hyperspectral remote sensing image are extracted through three-dimensional convolution, the hyperspectral image can be directly processed through three-dimensional convolution, preprocessing operations such as dimension reduction are not needed, and the spatial-spectral features of the hyperspectral image are extracted more sufficiently. The dense residual network is used to deepen the number of network layers and learn deeper spectral and spatial features, the residual network can effectively reduce the problem of gradient disappearance along with the increase of the network depth, and the structure can more effectively utilize the features and enhance the feature transfer between convolutional layers. The training time is shortened through an early stop method, classification prediction is carried out through a Soft-max classifier, and an initial classification result is obtained; and proposing a multi-label conditional random field optimization algorithm, and optimizing a classification result. The method improves the operation efficiency, and improves the classification accuracy of the remotesensing images.

Disclosed is an electrolyte comprising a compound having both a sulfonate group and a cyclic carbonate group. The electrolyte forms a more stable and dense SEI layer on the surface of an anode, and thus improves the capacity maintenance characteristics and lifespan characteristics of a battery. Also, disclosed is a compound represented by the following Formula 1, and a method for preparing the same by reacting 4-(hydroxyalkyl)-1,3-dioxolan-2-one with a sulfonyl halide compound:wherein each of R1 and R2 independently represents a C1˜C6 alkylene group optionally containing a C1˜C6 alkyl group or C2˜C6 alkenyl group introduced thereto; R3 is selected from the group consisting of a hydrogen atom, C1˜C20 alkyl group, C3˜C8 cyclic alkyl group, C2˜C6 alkenyl group, halo-substituted alkyl group, phenyl group and benzyl group.

The invention discloses a method for preventing water erosion desertification by using biological crust, which is to artificially inoculate a biological crust resource onto bare soil surface of a water erosion desertified area and cultivate the biological crust to allow the biological crust to cover the bear soil surface. The method comprises the following steps:1) selecting well developed natural biological crust, and shoving 10 to 30-millimeter-thick crust layer on surface for later use; 2) drying the collected biological crust in shade, removing other materials, crushing, fully stirring and uniformly mixing to obtain the biological crust resource; 3) mixing the prepared biological crust resource with fine soil according to a ratio of 1:(1-2), uniformly spreading the mixture on the soilsurface to be inoculated, covering a small amount of fine soil, and fully watering after the fine soil is covered; and 4) controlling a proper condition to promote the growth of the biological crust,namely watering, topdressing, killing pests and weeding. With the method of artificial inoculation and cultivation, a high-coverage biological crust can be formed in a short period, and the formed biological crust has obvious water and soil holding function; and the method is an effective method for treating water erosion desertification.

The invention discloses a method for tracking a hypersonic velocity reentry vehicle based on Gaussian mixture approximate. The method herein includes the following steps: establishing a movement tracking model of a hypersonic velocity reentry vehicle in a North East Up inertial coordinate system, and separately establishing a radar observation model and a target priori model set of a target tracking system; and based on the priori model set and the radar observation model, in combination with radar echo data, acquiring a target state smooth estimation value and a covariance matrix smooth estimation value. According to the invention, the method introduces a Gaussian mixture approximate theory, designs a Gaussian mixture smooth filter under multiple models, uses Gaussian component adaptive consolidation strategy, and selects smooth window length in a scientific manner, so that the method herein can avoid large amount of computation and particle degeneracy that often exist in particle filtering. The method has the characteristics of simple computing and easy implementation, can accurately estimate sudden maneuvering, and increases rapid and precise tracking capability of the hypersonic velocity reentry vehicle.

An automobile body-coating method which comprises optionally subjecting an aluminum material in the automobile body to a surface treatment, followed by subjecting to an oxidizing film treatment, essentially or optionally, respectively and successively forming a curable electrodeposition coating film, a water based chipping primer coating film, an intercoat coating film, and a curable topcoat coating film, or comprises applying a film racing material onto an aluminum material or coating film-coated aluminum material, and comprises forming respective multi-layer coating films onto the steel material in the automobile body respectively.

The invention discloses an improved M-Net-based RGB color remote sensing image cloud detection method and system, and belongs to the field of artificial intelligence and image recognition, RM-Net deepsemantic segmentation network is designed combining advantages of a residual error network and M-Net. The method comprises the following steps: firstly, enhancing an original data set, and labeling acorresponding pixel-level tag; multi-scale features of the image are extracted on the premise that information is not lost through pooling of the hollow space pyramid, and the network is not prone todegeneration by combining with a residual unit; and finally, extracting global context information of the image by using an encoder module and a left path, recovering the spatial resolution of the image by using a decoder module and a right path, judging the category probability of each pixel according to the fused characteristics, and inputting the category probability into a classifier for pixel-level cloud and non-cloud segmentation. According to the method, the color image is trained and tested, experimental results show that the cloud edge details can be well detected under different conditions, high-precision cloud shadow detection is obtained, and it is proved that the method has good generalization and robustness.

The invention discloses a centroid tracking framework based particle filter and mean shift cell tracking method, which mainly solves the problem of low accuracy rate of the traditional cell tracking method. The cell tracking method comprises the following steps of: performing binary segmentation to a video image, and extracting the central position of each cell; tracking the centroid of the cell, and recording the tracking trace of the cell; respectively recording the starting coordinates and the terminating coordinates of the trace into a starting coordinate set and a terminating coordinate set, and selecting a cell to be tracked; further predicting the trace of the cell to be tracked by using particle filter to obtain a predicted coordinate point in the next frame of image; selecting the subsequent tracking trace of the cell to be tracked by using the mean shift method in good time according to the predicted coordinate point; and circulating the steps of prediction and selection till the last frame of image, and completing the tracking of all cells. Compared with other traditional tracking methods, the cell tracking method has improvement in the aspects of tracking effect and accuracy rate and can be used for analyzing motor cells in a medical microscope video image.

The invention discloses a complex background SAR vehicle target detection method based on CNN, comprising the following steps: S1, collecting pattern data and processing to obtain sample data set; S2,ResNet and Faster-RCNN framework is fused to form a fusion framework, and the fusion framework is retrained on the basis of pre-training weights; S3, adopting the fusion frame after retraining to carry out target detection and recognition on the pattern data; the invention combines ResNet and Faster-RCNN framework, using Faster-RCNN framework realizes the end-to-end target detection process to realize the full automation of target detection, which is convenient for engineering application. At the same time, the residual network model is used to solve the problem of network model degradation in depth convolution network model, and the phenomenon of gradient disappearance in depth convolution network model is avoided.

The invention discloses an integrated attitude determination method based on an ant colony unscented particle filter algorithm, which relates to an inertial / astronomical integrated attitude determination method. The method comprises the following steps of: compensating gyro output data by using inertial measurement information and acquiring carrier attitude information through attitude calculation; acquiring required astronomical attitude information by using astronomical measurement information through a deterministic algorithm; fusing the astronomical attitude information with the carrier attitude information by using an ant colony unscented particle filter algorithm, solving nonlinear and noise non-Gaussian problems of a system, solving high-accuracy carrier attitude information, estimating gyro drift, and feeding back and correcting carrier attitude and compensating the gyro drift; and finally realizing on-line correction of eliminating random errors of a gyro of an inertial / astronomical integrated navigation system in real time based on the astronomical measurement information, and finishing long-term and high-accuracy combined attitude determination of a spacecraft.

The invention discloses an underwater image enhancement method based on a multi-branch generation countermeasure network. The invention takes the underwater degraded original image, the underwater clear image after fusion processing under the same scene, and the underwater clear generated image under the same scene as a training sample set, and inputs the image to an attribute branch network and adiscrimination branch network to obtain an attribute map and a discrimination map. The GAN network weights are updated by the gradient descent of the cost function of the attribute graph and the discriminant graph respectively. Until the end of this network training, the model of underwater image enhancement is obtained. The key of the invention is to emulate the enhancement strategy of the underwater image which is degraded by different factors by using the characteristics of the generation against the network data driving and the strong imitation ability. A single model can be used to solvea variety of underwater image degradation problems caused by different reasons, and the model is more versatile. Attribute branching and discriminant branching are used to enhance the comprehensiveness and robustness of learning.

The invention relates to a phase-transition self-temperature-regulating heat-preserving facing brick of an external wall, which is an integrated structure compacted by a phase-transition heat-preserving layer and an inorganic facing layer, wherein the phase-transition heat-preserving layer is mixed by proportion and compacted by hull-type expanded perlite heat-preserving aggregates, phase-transition temperature-regulating aggregates and cements; the inorganic facing layer is mixed by quartz sand, calcium carbonate, calcium oxide, cement and silicon oxide powders; and the phase-transition temperature-regulating aggregate is made of phase-transition cores and encapsulated hulls. The manufacturing method not only effectively solves the encapsulation problem of the phase-transition material, leads the prepared phase-transition self-temperature-regulating heat-preserving facing brick of the external wall to have excellent heat-preserving performance, waterproof performance and permanent anti-ageing performance, but also radically solves the obsolescence of thermo-physical property and the leakage problem of phase-transition material in the circulation process. The facing brick has the advantages of light weight, strong adhesion strength, large compressive strength, heat insulation, heat preservation, waterproof and moisture-proof performances, obvious energy-saving effect, and the like.

The invention discloses a face attribute recognition method, a device, a terminal device and a storage medium. The method comprises the following steps: obtaining a current data frame in a video stream, capturing a face from the current data frame, and obtaining a face area image. The feature information of human face is extracted from the image of human face region and input into the convolutional neural network model. Wherein the convolution neural network model is based on a ResNet architecture. According to the output result of the convolution neural network model, the age and the sex corresponding to the face are obtained. The invention can reduce the error of age estimation and gender detection when the environmental impact is large.

The invention relates to a wireless sensor network-oriented mutual inference compound chaos stream cipher implementation method, which comprises the following steps of: (1) converting typical Logistic mappings and Tent mapping into a discrete function on an integer field and enabling a chaos sequence integer model to be implemented on a sensing node of a wireless sensor network which supports integer computation; (2) on the basis of an integer chaos function, establishing three Logistic chaos mappings with different initial values as a main chaos system, and then establishing a Tent chaos mapping as a chaos controller; and (3) enabling the chaos controller to mutually interfere with the three Logistic mappings, and simultaneously, enabling the Tent chaos controller to be compounded with the three Logistic chaos mappings to form a safer random cipher stream with longer period to further form a complete chaos stream cipher. The wireless sensor network-oriented mutual interference compound chaos stream cipher implementation method has favorable chaos characteristic and enhanced safety and practicability.

The invention discloses an insulator picture defect detection method based on combination of Faster R-CNN + ResNet101 + FPN. The method comprises: collecting an insulator defect sample picture and marking defects; carrying out denoising and anti-shake preprocessing and data expansion, and dividing into a training set and a test set; building an insulator defect detection network model by combiningthe ResNet50 network framework model and the Faster R-CNN detection network model with the FPN feature pyramid network model, and performing target detection training by using the training set to obtain a primary insulator defect detection network model; testing and adjusting parameter optimization by using the test set to obtain a final model; and processing the insulator defect picture to be detected by using the final model. The method can achieve the automatic recognition of the defects of the insulator, is higher in accuracy, is good in stability, is high in anti-interference capability,is high in universality, is good in robustness, and can be used for an intelligent inspection system of a transformer substation.

The present invention discloses a method and system for removing stripe noise on the basis of the Wavelet transform and the Fourier transform. An original image is subjected to the Wavelet transform to break up the two-dimensional original image into highpass / lowpass filtering characteristics in different directions, and detailed images of multiple directions of each layer are obtained; the detailed images of each layer are divided into noise direction detailed images and non-noise direction detailed images according to the actual direction of the stripe noise in the original image, and the noise direction detailed images in the detailed images of each layer are extracted; the noise direction detailed images are subjected to localization processing through the adoption of the Wavelet transform, the noise direction detailed images of each layer are converted to superposition sum of each frequency component mode, and magnitude spectra of the noise direction detailed images of each layer are obtained; and high-frequency signals with a certain width in each magnitude spectrum are filtered, and the stripe noise with a target width is removed. According to the invention, the stripe noise with any angle is effectively removed in the condition without the addition of the time complexity and the computation complexity and on the basis of completely unaffected image quality.

The invention provides a network abnormal flow detection method, model and system based on a residual error gating circulation unit (Re-GRU) and integrated dynamic extreme learning (ELM) optimization. The method comprises the following steps: firstly, establishing a feature optimization method of Fisher Score and a maximum information coefficient; secondly, changing an original GRU candidate hidden state activation function into an unsaturated activation function, and introduing a residual structure into the GRU candidate hidden state, so that the gradient disappearance problem is avoided, the network is more sensitive to gradient change, and the purpose of relieving network degradation is achieved. Then, the model is optimized and designed into a bidirectional residual GRU structure, so that the network flow characteristic extraction performance of the model is more excellent; and finally, a two-step game integrated dynamic ELM network flow detection method is provided, and an overfitting problem is solved by using a full connection layer and a Dropout layer so as to improve the detection precision, and outputting a detection result. According to the method, the experimental simulation model is established, the validity is verified according to the comparison result of different parameters, and compared with a traditional detection method, the method has better detection effect and accuracy when detecting the abnormal traffic of the network.

The invention relates to an image tracking method based on sequential particle swarm optimization, which comprises the following steps: in a present frame image, randomly spreading an individual optimal state group in the last frame image by utilizing state transition distribution; performing the particle swarm optimization iteration on the particles generated after randomly spreading; evaluating an adaptive value of each particle by utilizing an apparent model of a spatially constrained gaussian mixture; updating the individual optimal state and the group optimal state of the particles according to the evaluating results for the adaptive values; and performing the convergence judgment: if meeting a convergence condition, outputting an observed value corresponding to the particle of a group optimal state as a tracking result of the present frame image, and if not, proceeding with the particle swarm optimization iteration. By using the method, the effective target tracking is realized and the application prospect is excellent.

The invention discloses a broad bean and corn intercropping method and belongs to the field of crop cultivation methods. The method mainly relates to intercropping of the broad bean and the corn, wherein, emergence of seedlings of the broad bean is 10-20 days earlier than that of the corn, the broad bean and the corn are intercropped at a row ratio of 2:2 and at a row spacing from 28cm to 32cm. The method can help realize the purpose of simultaneously improving yield of the broad bean and biological nitrogen fixation amount, avoid the disadvantages of low yield of a leguminous crop which is cropped separately; secondly, the method can greatly reduce manually applied nitrogen fertilizer to save the plantation cost, and relieve the pressure on the environment due to nitrogen fertilizer abuse; furthermore, the method can adequately utilize indigenous rhizobium to avoid a degeneration problem of artificially inoculated strains, reduce operation steps such as inoculation and the like and save the inoculation cost.

A tissue culturing method for high-quality papaya sprout features that the adult lateral bud of the disease-resistant female papaya tree is chosen as explant, and the conditions for tissue culture and rooting culture are disclosed for higher effect.

The invention discloses a high-nutrient microbial bacterium composite fertilizer and a production method thereof. The high-nutrient microbial bacterium composite fertilizer comprises the following components in parts by weight: 16-190 parts of double-wall layer composite microbial bacterium microcapsules (containing 3-5 parts of composite microbial bacteria), 150-450 parts of monoammonium phosphate, 0-400 parts of urea powder, 100-200 parts of potassium chloride, 200-350 parts of ammonium chloride, 20-100 parts of ammonium bicarbonate, 0-150 parts of ammonium sulfate and 0-50 parts of attapulgite. The production method comprises the following steps: adsorbing the microbial bacteria by adopting germ powder thereof, performing first wall layer embedding by adopting a chemical cross-linking method, performing second wall layer embedding by adopting a spray drying process, and finally coating on the surface of chemical fertilizer granules, and split charging. The composite fertilizer is sufficient in nutrient and high in viable count, so that the microbial bacteria can be slowly released, yield increase of crops is guaranteed, the incidence rate of insect pests is reduced, and soil hardening caused by application of chemical fertilizers is improved.

The invention discloses a many-to-many voice conversion method and system based on speaker style feature modeling. Firstly, a multi-layer sensor and a style encoder are added to a StarGAN neural network, effective extraction and constraint of speaker style features are realized, the defect of limited speaker information carried by one-hot vectors in a traditional model is overcome, then, an adaptive instance normalization method is adopted to realize full fusion of semantic features and speaker personality features, so that the network can learn more semantic information and speaker personality information, furthermore, a lightweight network module SKNet is introduced into the generator residual error network, so that the network can adaptively adjust the size of a receptive field according to multiple scales of input information, the weight of each feature channel is adjusted through an attention mechanism, the learning ability of frequency spectrum features is enhanced, and the details of the frequency spectrum features are refined.

The present invention relates to an audio quality recovery system based on GAN. The system comprises a model sharing block module, a generation network model module, a distinguishing network module and a sequence recombination module; the model sharing block module is mainly used to perform feature extraction for time-domain signals without frequency domain processing due to avoidance of information loss and extract the features to a high-level unit; the generation network model module employs a high-level extraction unit to perform analysis and reconfiguration; the distinguishing network module is configured to continuously perform adversarial training with the generation network model to continuously improve a generation effect; and the sequence recombination module analyzes the sequenceweighing recombination of the final generation output by the network. The audio quality recovery system based on the GAN can generate more vivid audio signals.

The invention discloses a residual single image rain removal method based on an attention mechanism, and mainly solves the problems that the existing single image rain removal technology has limitation and is not ideal in processing effect. According to the scheme, the method comprises the following steps: 1) preprocessing an input image to obtain a preprocessed image; 2) constructing an attention residual neural network model comprising a residual network module and a codec network module; 3) inputting the preprocessed image into an attention residual neural network model for training, constraining the attention residual neural network model by using a loss function, and then performing back propagation for parameter updating to obtain a trained rain removal neural network model; and 4) inputting a to-be-processed rain image into the rain removal neural network model for image processing to obtain a rain-free clear image. According to the invention, rain stripes in a single rain-containing image can be effectively removed, and a clear image is obtained; and meanwhile, background information in the original image is fully reserved.

The invention discloses a moving object detecting and tracking method based on compressive sensing. The moving object detecting and tracking method includes the steps that supporting set fusing and residual compensation sequence rebuilding are carried out on a video sequence, target detection is carried out through background subtraction based on compressive sensing, and a moving object is tracked according to the improved particle filtering algorithm. When background extraction is carried out, the calculated amount of compressive sensing domains will be much less than the calculated amount of spatial domains, so that a large number of storage and technology expenses are saved; meanwhile, the correlation between frames of images in an image sequence serves as prior information and is used in the compressive sensing restructuring procedure, the searching space can be reduced, and the restructuring accuracy of a restructuring algorithm can be improved. The tracking performance of the improved tracking algorithm which fuses the compressive sensing principle and sheltering judgment and is based on dynamic weight section is more accurate under the illumination shadow and sheltering conditions, and the degeneration problem is obviously solved.

The invention provides a monocular line feature map construction method based on epipolar constraint, which comprises the following steps of: 1, tracking feature points of two adjacent frames of images Ki and Kj by using an optical flow tracking method, and searching an essential matrix of a reference angular point and a tracking angular point; 2, extracting LSD line features of each frame of image, and calculating an LBD descriptor; 3, calculating the midpoint epipolar line of each straight line li of the reference frame, and calculating the included angle theta between the midpoint epipolarline and the corresponding matched straight line; 4, selecting points on the straight line from the point set P<i,k> of li, calculating intersection points of the epipolar lines corresponding to the points and the matched straight line, and the intersection point set being I<i,k>; 5, determining a point set P<j,k>; 6, normalizing and triangularizing the point set, determining a space point set Pk,fitting a straight line Lk.S7, re-projecting the straight line to the ith frame and the jth frame, constructing a re-projection error, and updating a camera pose and a space straight line; 8, determining a starting point and an ending point of an endpoint space straight line; 9, updating straight lines.

The invention discloses an enhanced Transform-based image blind denoising method and system, and the method comprises the steps: combining a dynamic convolution layer with an enhanced Transform module, and carrying out the feature fusion of a plurality of modules in a weighting manner in a self-adaptive manner; a dynamic convolutional layer is introduced, parameters are adaptively adjusted under the condition that extra network depth and width are not increased, and the expression ability of the model is greatly improved; residual learning operation is added into the Transform module, so that the problem that the Transform module is difficult to train is solved, global features and semantic information are extracted more effectively, and the denoising effect is improved; a residual learning operation is adopted, hierarchical features obtained by a convolutional layer, a dynamic convolutional layer and an enhanced Transform module are fused respectively, and the memory ability of each layer of the network is transmitted; the features of the enhanced convolutional layer, the dynamic convolutional layer and the enhanced Transform module are fused through connection operation, then the weight is obtained through Softmax, secondary extraction of the features is achieved in an attention mode, and the saliency features are further obtained. According to the method, a good effect is obtained on an image blind denoising task.