83results about How to "Improve algorithm performance" patented technology

The invention discloses a movement capturing method based on multiple binocular stereo vision. A movement video collecting device is constructed, and human movement video sequences from different orientations are collected by the movement video collecting device. Multiocular movement video sequences shot by a plurality of cameras are calibrated. Marked points matching and tracking of each binocular tracker is finished. Data fusion of three-dimensional tracking result of multiple binocular trackers is completed. The three-dimensional movement information of the marked points acquired by a multiocular fusion device is fed back to the binocular tracker to consummate binocular tracking. On the basis of binocular three-dimensional tracking realized by binocular vision, the invention fuses multiple groups of binocular three-dimensional movement data, resolves parameter acquiring problem of three-dimensional position, tracking, track fusion and the like for a plurality of marked points, increases number of traceable market points and enables the tracking effect to be comparable with three-dimensional movement acquiring device employing multi-infrared cameras for collecting.

A method for accurate and rapid automated detection of atrial fibrillation (AF), sinus rhythm (SF), and atrial flutter (AFL) is disclosed, which allows distinguishing of these cardiac signals with lowered risk of errors by implanted pacemakers and like devices. The method includes training episodes of intra-cardiac signals (called the closed data set CDS) to evaluate five feature parameters with a discriminator classifying the signal into AF, AFL or sinus rhythm (SR). Comparison with the independent decisions of experienced physicians for each episode reveals specificity, accuracy and sensitivity of greater than 97%. Each episode is a window of intracardiac signal of interval 1-2 seconds with the discriminator providing results in less than 0.25 s. In another aspect, the method is resistant to the presence of noise in the data. In yet another aspect, more feature parameters may be used in alternative implementations including for detecting signals other than AF, AFL & SR.

The invention relates to an augmented reality based mobile platform three-dimensional biomolecule display system and an augmented reality based mobile platform three-dimensional biomolecule display method. The identification of the traditional paper print biomolecule two-dimensional picture and the three-dimensional real-time display of the corresponding molecule are realized on a mobile platform. The image feature point detection and matching method is improved, so that the display is collaboratively completed by a mobile client and a cloud server, the flexibility is improved, and the client resources are reduced. Three-dimensional molecular modeling data is derived from a PDB (Protein Data Bank), so that the accuracy and the authority are ensured. The augmented reality technology is expanded, a typical application is provided in biochemistry and molecular biology education learning aspects, and the technical means and the framework can be applied to any education and teaching fields; because of the development and popularization of a mobile terminal, and the portable characteristic of the mobile terminal, the system can be used by users at any time and any place, so that the teaching efficiency and the interest in learning are greatly improved.

The invention relates to the field of human body motion recognition, and provides a human body motion recognition method based on multi-sensor data fusion. The method comprises the following steps: human body motion data are collected by using N inertial sensor nodes respectively fixed at different parts of the human body; sliding window segmentation technology is used to segment the human motiondata collected from each sensor node to obtain multiple motion data segments of each sensor node; feature extraction is carried out on the motion data segment of each sensor node to obtain the corresponding feature vector; an RLDA algorithm is used to reduce the dimension of eigenvectors of each sensor node; the reduced dimension eigenvector of each sensor node is used as training data for parameter training and modeling, and the corresponding hierarchical fusion model is obtained; using the hierarchical fusion model, human motion recognition is carried out. The invention can effectively overcome the drawbacks of the single classifier in the identification process, and can effectively improve the human motion identification accuracy.

The present invention relates to an image blur processing method and device, a storage medium and an electronic device, which are used for solving the technical problem that the image blur processingeffect is not ideal in the prior art. The method comprises the following steps of: selecting a reference pixel point of a target pixel point on an image to be processed; calculating a pixel differencevalue between each reference pixel point and a target pixel point; determining a weight value of each reference pixel point at least according to a preset correspondence between a pixel difference value and a weight value; calculating a blur value of a target pixel point according to the weight value of each reference pixel point; blurring the target pixel according to the blurred value.

The invention discloses a child emotion recognition method based on a multi-attention mechanism long-short-time memory network. the method comprises: detecting an end point of a test set speech and divides the speech into frames to extract time sequence related features, the processing algorithm of extracting temporal correlation features with different lengths based on long-short time memory network; the strategy of combining the attention mechanism with the depth of time sequence is introduced into the forgetting gate, input gate and final output of the long and short time memory network. Finally, the samples to be tested are input into the improved long-short time memory network in the training process, the emotional information can be distinguished distinctly, By introducing the depthof attention mechanism combined with timing into the forgetting gate, the output gate and the final output of the long-short time memory network, the invention not only greatly reduces the parameter quantity, but also improves the algorithm performance, increases the flexibility of the method design, and has high identification efficiency and good application prospect.

According to the method in the field of medical image processing, an efficient medical image marking and learning system is realized. According to the system, an unsupervised comparative learning pre-training module introduces an unsupervised comparative learning method in a pre-training process based on an MoCo algorithm; the training module based on active learning and comparative learning obtains a pre-trained model and an actual target task data set; and further adjusts the model to adapt to the actual demand by discovering the most valuable sample in the unlabeled samples to obtain the label and providing the difficult case mining of higher weight for the sample classified by mistake and introducing the supervised contrast learning method of labeled data; and a manual annotation and model training interaction control module provides an annotation interaction interface and a model training control module. Through the system of the architecture, a universal training framework is realized.

The invention discloses an optimized extreme learning machine binary classification method based on improved active set algorithms. The method comprises the following steps of: firstly, providing a training sample set, and determining an optimized solution problem; secondly, calculating a decreasing direction dk according to a BAS active set algorithm, and finding a maximum searching step size without violating constraint conditions along the decrease direction dk; thirdly, setting a temporary iteration step size according to an EAS active set algorithm, and continuously finding an optimal step size; and fourthly, utilizing a prediction assignment method to reduce the calculation cost of iterating a working set variable to the optimal value. An optimized extreme learning machine is adopted to carry out classification on unknown samples.

A system and method for detecting spectrum holes in a region is disclosed. The system comprises a requesting module, a detecting module and an aggregation module. The requesting module retrieves channel availability data for a region. The channel availability data describes potential availability of a channel. The detecting module obtains a set of values for the channel by spectrum sensing at a set of locations in the region based at least in part on the potential availability of the channel. The aggregation module aggregates the set of values for the channel over the set of locations in the region to obtain a first value for the channel and determines the availability of the channel based at least in part on the first value for the channel.

The invention discloses a method for intelligently forecasting the wind speed in a wind power station. The method comprises the following steps of acquiring and inputting data, layering data sequences, establishing models, forecasting and comprehensively calculating, and outputting forecasting results, wherein in the step of layering data sequences, the original unstable wind speed is decomposed into two stable wind speed data outputs by adopting a wavelet packet decomposition method, and the number of the wind speed data outputs is defined as the number of wind speed sequence layers; in the step of establishing mathematical models, each layer of data in the wind speed sequence layers are independently processed, a BP (back propagation) neural network model is established for the high-frequency sequence layer, high-frequency data are calculated and then enter a data stack; a time sequence model is established for a low-frequency layer, the low-frequency data are calculated and then enter the data stack; after entering the data stack, all the data in the data stack just enter the forecasting and comprehensive calculating step for weighting calculation, and finally the forecasted results are output. The method provided by the invention belongs to an intelligent method, and can be used for realizing multi-step advance forecasting.

The invention discloses an automatic balance control method based on APSO-BP for double counterweight plates. Based on a BP neural network, balance data of each time are fully utilized and the balancing parameters are fitted and optimized to realize automatic balancing control. The input parameters are the initial position A1 and B1 of the two balancing blocks, rotating speed N, unbalance force (shown in the description), the optimization objective is system vibration response (shown in the description), and the optimal output is the target positions A2, B2 of the two balancing blocks. Basedon the initial position, rotational speed and system vibration response of the counterweight, the BP neural network is used to fit the system vibration value, and the adaptive particle swarm optimization is used to obtain the optimal target position of the two counterweight blocks. The invention can effectively optimize the particle swarm optimization based on the neural system fitting input-output relationship by using the improved adaptive weight adjustment method, which is conducive to promoting global optimization, accelerating the convergence rate and increasing the accuracy of the automatic balance system control and the effectiveness of the data.

The application is applicable to the field of computer application technology, and provides a doctor recommendation method and device. The doctor recommendation method includes: obtaining patient information and doctor information; extracting patient features from the patient information and doctor features from the doctor information; constructing a fitness function according to the patient features and doctor features, and determining the feature weight corresponding to each doctor feature according to the fitness function; and determining the doctor information that most closely matches thepatient information according to the feature weight corresponding to the doctor features. An optimal weight combination is searched with the objective of optimizing the accuracy of the final recommendation result, then a method of weighting is used to combine the feature similarity calculation results in the collaborative filtering algorithm, and the doctor recommendation ranking result that bestmatches the patient is output. The multi-objective feature selection model improves the accuracy and stability of individual identification and classification, also improves the matching degree between doctors and patients, and achieves personalized and targeted doctor recommendation.

The invention discloses a universal descent search method for signal detection of a large-scale MIMO system. The method comprises the following steps: S1, constructing a linear inequality quadratic optimization model for detecting transmitting signals of the large-scale MIMO system according to parameter situations under a non-ideal channel, and constructing a detection matrix A; S2, adding a mapping operation to a descent search algorithm according to the model constructed in the step S1; S3, calculating a lower triangular matrix L according to a SSOR preprocessing algorithm; S4, according tothe matrix L, detecting a receiving signal matrix output by a receiving end matched filter through the descent search algorithm containing the mapping operation, and obtaining a transmitting signal estimated value. Compared with the existing iterative detection method, the method provided by the invention is suitable for a more variable channel model and a constantly expanding system scale, stillcan keep low error rate under low signal-to-noise ratio in bad conditions, and can satisfy requirements on detection technologies of the next generation mobile communication better.

The invention relates to a line feature simplification method based on the clonal selection algorithm. The clonal selection algorithm is used for solving the line feature simplification problem by fully using the advantage of optimization-problem solving of the clonal selection algorithm; according to the characteristic of line feature simplification problem, the affinity function and the constraint condition of the clonal selection algorithm are designed, and a clonal selection algorithm model applied to the line feature simplification is constructed; the local search algorithm and the clonal selection algorithm are combined and the global search capacity of the clonal selection algorithm and the local optimization capacity of the local search algorithm are fully utilized so that the line feature simplification precision and algorithm are improved. The invention provides a new thinking direction for promoting the line feature simplification technique system to develop automatically and intelligently and enriches related research content and methods. The research results provide important theoretical guidance and technical support for the cartographic generalization of actual production.

The invention provides a hybrid expert reinforcement learning method and system, which utilize a multi-head agent with shared network parameters as a plurality of experts and train the agent through adeep deterministic strategy gradient algorithm so as to learn a series of similar tasks at the same time. Each expert is introduced into an uncertainty estimate of performing an action in a state toenhance the anti-overfit Q-value assessment capability and the overall performance of the model. These enable agents to extract, migrate, and share learned knowledge (feature expressions) among different tasks, thereby improving learning efficiency of a single task and effectiveness of scheduling experts among a plurality of tasks. Different from the traditional data-driven design of a hybrid expert system, the hybrid expert system adopts a self-supervised gating network to determine an expert with the highest potential to process each interaction of an unknown task, and can completely calibrate the scheduling accuracy through uncertainty estimation fed back by the expert system under the condition of no man-made external supervision information.

The invention discloses a force-oriented graph layout method based on community discovery and clustering optimization, and relates to the technical field of visual layout of graph data. Comprising the following steps: converting original data into corresponding graph data; dividing nodes of the graph data into leaf nodes and non-leaf nodes, regarding each non-leaf node as a community, and compressing the leaf nodes to obtain compressed graph data; performing a community discovery process of maximizing modularity in a first stage of a traditional Louvain algorithm on the compressed graph data; replacing the iterative community merging process in the second stage of the traditional Louvain algorithm with selective community merging for the updated graph data in the previous step; and for the community structure obtained in the step 4 and the corresponding graph data, realizing the force-oriented graph layout based on clustering optimization by using a CombboForce layout algorithm. According to the method, the layout efficiency of the force-oriented graph layout during visual layout of the graph data is improved, and the layout effect of the force-oriented graph layout during visual layout of the graph data is optimized.

A system for detecting sleep apnea uses a PPG sensor for generating a PPG signal. An estimate of respiratory effort is derived from the PPG signal, and from this a respiratory effort signal is derived. Characteristic features are extracted from the respiratory effort signal, and sleep disordered breathing events are detected from the extracted characteristic features. This analysis can distinguish between sleep disordered breathing events and normal breathing during sleep.

The invention discloses a three-dimensional reconstruction method, a face three-dimensional reconstruction method and related devices thereof. The method comprises the following steps: acquiring an image comprising a target object; obtaining feature parameters of the target object based on the image; performing convolution calculation on the feature parameters by using a neural network to obtain three-dimensional reconstruction parameters of the target object, wherein the neural network is obtained by solidifying the association relationship between the feature parameters and the three-dimensional reconstruction parameters; and performing three-dimensional reconstruction on the target object based on the three-dimensional reconstruction parameters. Through the method, power consumption and memory consumption caused by an algorithm in a three-dimensional reconstruction technology are reduced.

The present invention relates to a fertilization precision control method for water and fertilizer integrated equipment and a control system thereof. The method comprises the following steps: step S1, establishing a fertilization precision control model of the water and fertilizer integrated equipment; and step S2, solving an optimal solution of the fertilization precision control model by adopting estimation of distribution algorithms.