59results about How to "Improve generalization performance" patented technology

The invention discloses a pulmonary nodule diagnosis method based on a local receptive field and semi-supervised depth self coding. Firstly, the local receptive field is adopted to carry out multilevel feature extraction on a pulmonary nodule image; then, semi-supervised sparse self coding is used to autonomously learn nodule features in a pulmonary image; and finally, multiple pieces of clinical information are fused to realize accurate detection on the pulmonary nodule. The recognition accuracy is high, the inputted CT image can be subjected to feature multiple classification and then self coding learning, the network diagnosis is more accurate, and an important role in auxiliary diagnosis is played for a doctor.

The invention relates to an Adaboost classifier on-line learning method and an Adaboost classifier on-line learning system. The Adaboost classifier on-line learning method comprises steps that: object detection is carried out by a strong classifier acquired through employing off-line training, and an object detection result is acquired; object detection is acquired by employing a background model to acquire a motion object; the object detection result acquired by the strong classifier is compared with the motion object acquired through detection by employing the background model to acquire an error classifier object; the error classifier object is taken as an on-line training sample to carry out on-line training to acquire a strong classifier after updating. According to the Adaboost classifier on-line learning method and the Adaboost classifier on-line learning system, the object detection result acquired by the off-line classifier is compared with the motion object acquired through detection by employing the background model to acquire the error classifier object, the error classifier object is taken as the on-line training sample to acquire the strong classifier after updating. Generalization performance of the object detection classifier is effectively improved, so the object detection classifier can adapt to a monitoring scene during operation, a detection ratio is improved, and a rate of false alarm is reduced.

The present invention discloses a smart home control system and method based on Alljoyn and machine learning and belongs to the machine learning, Internet of things and mobile Internet technical field. The method includes the following steps that: S1, sample sets are acquired, a neural network model is obtained by means of training; S2, a machine learning system is deployed, and communication between the machine learning system and a server is established; S3, an Alljoyn gateway is deployed, and communication between the Alljoyn gateway and the server is established; S4, a device and a data acquisition module are connected into the Alljoyn gateway; S5, a mobile end application controls the device; S6, the neural network model is updated; and S7, smart scenes are seamlessly switched. According to the smart home control system and method of the invention, artificial intelligence is introduced into a smart home and is combined with a mobile Internet, and therefore, a convenient, efficient, safe and user-friendly smart home control scheme is designed, and the internal / external network real-time synchronization of the state of an intelligent device, the seamless switching of the smart scenes and the self-management of the smart home system are realized.

The invention discloses an electrocardiosignal graph classification method based on deep learning, and belongs to the field of image recognition. A traditional single convolutional neural network or single recurrent neural network has the problem of poor classification in part of categories, and has the problem of poor generalization ability of an automatic classification algorithm caused by different representations under the same disease due to individual differences such as heartbeat intensity, heart rate and the like. The invention discloses an electrocardiosignal graph classification method based on deep learning. The method comprises the following steps: preprocessing acquired image data of the electrocardiosignal graph, wherein the preprocessing includes wavelet denoising of the image data of the electrocardiosignal graph and data enhancement of the image data of the electrocardiosignal graph; designing a convolutional neural network, and classifying the electrocardiosignal graph by utilizing the designed convolutional neural network. The electrocardiosignal recognition and classification effect can be remarkably improved.

The embodiment of the invention discloses a pedestrian recognition method based on dynamic occlusion samples. The method comprises: constructing an original image feature learning network framework; inputting a pedestrian image of a training set to obtain n local features, and learning and optimizing the local features, serially connecting the obtained n optimized local features as the original image features of the training pedestrian image; building the generator; generating an occluded pedestrian image; inputting the occluded pedestrian image into the generator to obtain n local features, and learning and optimizing the local features, serially connecting the obtained n local features as occluded image features; obtaining the final features of the training pedestrian image by using theoriginal image features and the occlusion image features, and performing pedestrian recognition by using the final features of the training pedestrian image. The invention makes full use of the advantages of the convolution neural network, learns the original image feature and the occlusion image feature of the pedestrian, and finally fuses the two features to represent the pedestrian image, thereby further improving the matching accuracy of the pedestrian recognition.

The invention provides a classification model training method based on the crowdsourcing technology. The method comprises a step of estimating a level of providing annotation information of a user on crowdsourcing annotation information corresponding to a few samples, a step of taking an observed annotation people level as prior knowledge to determine the annotation information used by training samples, a step of taining a classification model on the training sample and the annotation information, a step of using the classification model to select a training sample which allows a model expected error to be minimum, and predicting a category to which the sample belongs, a step of adding the selected sample and annotation information provided by a user with a highest annotation level in the category into a training set, and a step of carrying out iterative execution of the above steps on an updated training set until the precision of the classification model or the number of the training samples reaches a preset standard. The method has the advantages that disadvantageous influence of the low quality annotation information provided by a user with a low annotation level on the classification model training is avoided, and an effect of training a high generalization ability classification model in a crowdsourcing environment is guaranteed.

The invention discloses a method and a device for estimating indoor scene layout based on a conditional generation confrontation network. The method comprises the following steps: a confrontation network is generated by training conditions of a training set; an indoor image to be tested is inputted to a conditional generation confrontation network after training; and a layout edge map with the same size as an input image is predicted and generated. the vanishing points of the indoor image to be measured are estimated, rays are extracted from each vanishing point at equal angular intervals, anda plurality of fan-shaped regions are generated; a sampling sector region is determined according to the criterion of maximum average edge strength; Gaussian blur is added to that predict layout edgemap, and then the sampling sector region is sampled to generate a layout candidate item; The spatial layout which is most similar to the predicted layout edge map is selected as the final layout estimation result. The invention provides more complete original information for generating scene layout boundary map, does not need explicit hypothesis data parameter distribution, can improve layout estimation accuracy, and has important application value in indoor scene understanding and three-dimensional reconstruction task.

The embodiment of the invention provides a method for automatically identifying organs endangered by radiotherapy in a CT image based on a deep semantic network. The method comprises the following steps: S1, preprocessing a CT three-dimensional image; s2, obtaining a part to which each two-dimensional image in the CT three-dimensional image belongs; s3, respectively constructing deep semantic segmentation models for the pelvic cavity, the abdomen, the chest, the head and the neck; s4, inputting the two-dimensional images belonging to the pelvic cavity, the belly, the chest, the head and the neck into a trained deep semantic segmentation model for the corresponding pelvic cavity, the belly, the chest, the head and the neck to identify organs endangered by respective radiotherapy; and S5, combining results output by the deep semantic segmentation models of the pelvic cavity, the abdomen, the chest, the head and the neck. According to the method, artificial intelligence-assisted radiotherapy endangered organ contour sketching is implemented in the working process of radiotherapy planning, preoperative evaluation of surgical operation and operation planning, and the working efficiencyof medical workers can be effectively improved.

The present invention discloses a learning prediction-based indoor layout estimation method and system. The method comprises the following steps that: a training set is constructed, and training samples in the training set are utilized to train a deconvolution network, wherein the training samples are room layout maps and edge maps corresponding to the room layout maps, and the room layout maps and the edge maps corresponding to the room layout maps are adopted as the input and output of the deconvolution network; a to-be-tested room layout map is inputted into the trained deconvolution network, and a predicted edge map is outputted; vanishing points in a preset direction, of the to-be-tested room layout map, are calculated, so that a plurality of sectors are generated; and sectors with local maximum edge strength are selected from the plurality of generated sectors as sampling sectors on the basis of the predicted edge map; the sampling sectors are sampled, so that a series of candidate room layout estimated maps are obtained; and a room layout map which is most similar to the predicted edge map is selected from the candidate room layout estimated maps as a final room layout map according to the similarity of the obtained room layout estimated maps and the obtained edge map.

The invention relates to a natural language processing technology in artificial intelligence, in particular to a text recognition method and device, computer equipment and a storage medium, and can be applied to scenes such as electronic commerce, news information, microblog forum and vehicle-mounted recommendation. The method comprises the following steps: acquiring a grammatical relationship sequence of a to-be-recognized text; if the grammatical relationship sequence of the to-be-recognized text comprises at least two grammatical relationship sequences in preset grammatical relationship sequences, determining a feature word position tag in the to-be-recognized text according to one grammatical relationship sequence in the at least two grammatical relationship sequences; determining feature words in the to-be-recognized text according to the feature word position tags in the to-be-recognized text; and determining a text recognition result of the to-be-recognized text according to the feature words in the to-be-recognized text. By adopting the method, the determination accuracy of the feature words in the to-be-recognized text is improved, so that the text recognition accuracy is improved, and the effectiveness of a data analysis result of big data is ensured.

The invention relates to the technical field of agricultural product detection, in particular to a method for detecting quality of figs based on near infrared spectrum. A random forest regression algorithm is an ensemble learning algorithm taking a decision tree as a base learner, CART decision trees are trained by using a Bagging ensemble learning technology and forming a forest, the decision trees in the forest are not associated (FIG1), and an average value of the results output by the decision trees is used as a regression result (FIG2), so that the problem of overfitting is solved, and the overall model has relatively high precision and generalization performance; and low RMSEC and RMSEP show that the random forest algorithm is used for predicting the internal quality of the green-peel figs, a good effect is achieved through experiments, and a large amount of experimental data proves that the random forest algorithm has higher fitting capacity than a partial least squares method and is not prone to over-fitting.

The invention discloses an ELM-based optimized classification method for land use and cover of hyperspectral remote sensing images, which comprises the following steps: firstly, a plurality of ELM-based classifiers are constructed, and a training data set is constructed for each ELM-based classifier; Then, T ELM-based classifiers are trained based on the training data sets of each ELM-based classifier, and the classification and prediction results of training samples in each training data set are obtained. Then, the ELM-based classifier set is pruned based on the classification prediction results. Finally, the hyperspectral remote sensing images to be classified, The spectral features of each pixel point are extracted, and the feature data of the object to be classified are obtained and inputted into the ELM-based classifiers in the set of classifiers retained after pruning, and the hyperspectral remote sensing images to be classified are classified and judged by ensemble, and the classification results of the hyperspectral remote sensing images to be classified are outputted. The invention realizes an optimized classification method for land use and cover of hyperspectral remote sensing images, which improves classification accuracy and classification processing efficiency.

The invention discloses an intention recognition method and system, electronic equipment and a readable storage medium. The method comprises the steps: crawling that question-answer data from the Internet through a crawler technology; identifying an intention category of an answer corresponding to each question in the question-answer data through an entity identification model to obtain question-intention data; taking the obtained problem-intention data as training data to train a deep learning model, and obtaining a problem and intention recognition model; inputting the question into the question intention recognition model to obtain an intention corresponding to the question. The problem that training data is difficult to obtain is solved, the cost of manual data annotation is greatly saved, and the method has good generalization ability.

The invention discloses a blind face restoration method and system, and the method comprises the steps: collecting a blind face data set, evaluating the quality of the blind face data set through theLaplace gradient, and removing blurred and non-face images; enhancing image data for the blind face data set, and randomly distributing to obtain a training set and a test set; constructing an AFFNetnetwork; inputting the images of the training set into an AFFNet network, training the AFFNet network in combination with a reconstruction loss function, a perception loss function, a style loss function and a resistance loss function, and training and optimizing the AFFNet network by using an SGD optimization algorithm to obtain an optimal blind face restoration model; and inputting the image ofthe test set into the optimal blind face restoration model, and carrying out matching and selection to obtain an image with the highest accuracy as a final retrieval result. According to the scheme, the method has the advantages of being simple in logic, accurate, reliable and the like, and has high practical value and promotional value in the technical field of image processing.

The invention relates to a pedestrian re-identification method and device, computer equipment and a storage medium. The method comprises the steps that a target image is input into a pedestrian re-identification network, the pedestrian re-identification network comprises a depth separable convolution layer, and the depth separable convolution layer comprises a plurality of parallel branches and afeature fusion layer cascaded with the plurality of parallel branches, each parallel branch is formed by connecting different numbers of depth separable convolution modules in series; feature extraction is performed on the target image through the parallel branches, and feature maps corresponding to different scale information and output by the parallel branches are obtained; feature fusion processing is performed on the feature maps corresponding to different scale information and output by the parallel branches through the feature fusion layer to obtain a fused feature map; pedestrian re-identification is performed according to the fusion feature map. By adopting the method, the precision of a pedestrian re-identification result can be improved.

The invention belongs to the technical field of medical image processing, and particularly relates to a medical image segmentation method based on a 3D dynamic edge insensitivity loss function. The model adopts a dynamic edge insensitivity loss function, and the design principle of the loss function is as follows: in each iterative training process, the farther the prediction error pixel point isfrom the edge, the more sensitive the network is to the pixel point, and the larger the punishment weight is. In this way, the untrusted edge sensitivity can be reduced, and the influence of the edgeuncertainty of different expert annotation data on the model is reduced, and the generalization performance of the model for medical image segmentation is improved. And meanwhile, a U-net architecturebased on an attention mechanism is adopted, and the deviation of the network model is relatively small by fusing the weight distribution of the attention module self-adaptive feature map, so that theinfluence of annotation noise on model learning is reduced, and the generalization and accuracy of the medical image segmentation model are improved.

The invention discloses a series arc fault identification method of an extreme learning machine based on a dynamic online sequence. The method comprises the following steps: 1) noise reduction of electric energy waveform sampling data; 2) real-time data segmenting and intercepting; 3) waveform data calculating and processing; and 4) ELM fault arc identification: based on an ELM algorithm, arc identification is converted into a fault classification problem, and weights from an input layer to a hidden layer are given randomly; after the weights from the input layer to the hidden layer exist, weights from the hidden layer to an output layer are obtained according to a least square method, and thus fault arc identification is realized. Through the dynamic online ELM learning algorithm which is efficient in calculation and high in universality, an accurate and effective way is provided for series arc fault identification of a power grid under different load conditions.

The invention discloses an OCT retina image field adaptive segmentation method and system, and the method comprises the following steps: 1, obtaining OCT retina image data, and carrying out the image preprocessing; 2, respectively constructing an intra-domain segmentation model and a cross-domain segmentation model in combination with the feature extraction network, the feature classification network and the feature identification network; 3, according to the intra-domain segmentation model and the cross-domain segmentation model, obtaining pseudo labels of the target domain; 4, selecting an effective pseudo tag and an original image by using a self-selection active learning strategy, adding source domain rich image data, and training the cross-domain segmentation model again; 5, segmenting the target domain OCT retina image by using the trained cross-domain segmentation model. Through the technical scheme of the invention, the generalization performance of the segmentation model can be effectively improved, reliable quantitative data is provided for clinical ophthalmic disease diagnosis, and computer-aided personalized treatment is realized.

The invention discloses a random weight network generalization ability improvement method and device, and a computer readable storage medium. The random weight network generalization ability improvement method disclosed by the invention has the benefits that firstly, an initial output layer weight of a weak random weight network is analytically calculated on a pseudo-residual data set; then an objective function considering the loss and the complexity of a current integrated learning model is designed, and the optimization criterion of an optimal output layer weight is obtained by minimizing the objective function; finally, the optimal output layer weight of the weak random weight network is calculated by taking the initial output layer weight as a heuristic method and in combination withthe derived optimization criterion; the process can be regarded as the re-optimization of the initial output layer weight of the weak random weight network; optimization rules are obtained through theobjective function, and the benefit of re-optimizing the initial output layer weight of the weak random weight network is mainly reflected in that the integrated learning model with a simple structure can get better generalization performance, better over-fitting control ability and smaller prediction variance.

With the increasing development of the deep learning technology, the requirement on the generalization performance of the model in a real environment is increased increasingly, and the influence of the differences of illumination, background and the like on the generalization performance of the model has aroused wide attention. The invention discloses a scene adaptive target detection method based on a motion foreground. According to the method, the target frame of the motion foreground is effectively utilized by utilizing the priori of the distribution consistency of the motion foreground and the global target data, and meanwhile, the instance feature similarity is calculated through the decoder, so that the effect of the model in the target domain is greatly improved. Experimental results show that the target detection effect of the method provided by the invention is greatly improved in a real environment.

The invention relates to a deepfake detection method based on video frame sequence prediction, and aims to improve the attention of a time sequence model on time sequence characteristics. According to a technical scheme in the invention, the deepfake detection method based on the video frame sequence prediction is characterized by comprising the steps that a suspicious video is input into a trained time sequence model, features of the suspicious video are extracted through the time sequence model, the features are input into a true and false classifier, and the true and false classifier outputs the true and false probability of the suspicious video. The training of a time sequence model comprises the following steps: randomly disorganizing original continuous video frames of video clips, and recording a disorganizing mode; inputting the disordered video frames into the time sequence model to extract features, and sending the features into a frame sequence classifier and the true and false classifier at the same time; and calculating frame sequence prediction loss between the result of the frame sequence classifier and the disorganizing mode, and calculating true and false classification loss between the result of the true and false classifier and true and false labels of the video clips. The method is suitable for the fields of machine learning and computer vision.

The invention relates to a non-intrusive electrical load decomposition method based on a variable weight time domain convolutional network, and the method comprises the steps: model training: respectively training decomposition models used for the electrical power decomposition of each piece of electrical equipment, the decomposition models comprising a plurality of time domain convolutional networks for the electrical power estimation, and during the training, the time domain convolutional networks are used for the electrical power estimation; for the same electric device, training the corresponding time domain convolutional network by using the electric load data of different time periods; model application: inputting the total electricity utilization power sequence to be decomposed into a decomposition model of each device, and performing electricity utilization power estimation by adopting a plurality of time domain convolutional networks to obtain a plurality of groups of electricity utilization power estimation values of each time point, and carrying out point-by-point variable weight weighted summation on the plurality of groups of electric power estimation values to obtain an electric power decomposition result of the electric equipment at each time point. Compared with the prior art, the method is high in decomposition precision and good in generalization performance.

The invention relates to the technical field of electricity price prediction, in particular to a day-ahead electricity price prediction method based on a crisscross algorithm and a deep learning model, and the method comprises the following steps: 1), collecting the original electricity price data of a high-proportion new energy electricity market, and carrying out the preprocessing of the original electricity price data; 2) establishing an LSTM prediction model, and taking the day-ahead electricity price, load, wind power and photovoltaic generating capacity before a prediction day as feature input of the LSTM prediction model; 3) adopting a conventional gradient descent method to train the LSTM prediction model for the first time; and 4) taking the minimum mean square error as an objective function, carrying out fine adjustment on the weight coefficient and bias between the fully connected layers based on a crisscross algorithm, and obtaining a final optimized long and short term memory network deep learning model. The method can effectively prevent the weight coefficient and bias of the deep learning model from falling into local optimum, improves the generalization performance of the model, and improves the prediction precision of the day-ahead electricity price.

A textile formaldehyde detection system. In order to improve the generalization performance under low concentration and small sample conditions, the autoregressive AR prediction model was used to interpolate the response data in the low concentration range of gaseous formaldehyde. In order to suppress the cross-interference of electrochemical reactions, empirical mode decomposition was applied to realize the separation of gaseous formaldehyde response data and high-frequency interference. On this basis, the actual gaseous formaldehyde response is approximated by interval regression modeling, and the low-frequency interference added to the gaseous formaldehyde response data is further eliminated. Combined with the above algorithm, the invention was invented, which can effectively suppress the cross-interference of electrochemical reactions, reduce the lower limit of detection, and improve the detection sensitivity.

The invention discloses a neural network reinforcement learning method and system based on a fitness track. The neural network reinforcement learning method comprises the following steps: initializinga neural network weight, reinforcement learning parameters and a fitness track; Obtaining a current environment state and an immediate return value; Calculating a Q value function of reinforcement learning; Acquiring a fitness track and updating a neural network weight; Detecting a new environment state and an immediate return value; When the new environment state and the immediate return value meet the end condition, reinforcement learning is finished, the end condition is not met,obtaining the current environment state and the immediate return value through return re-detection. The method has the advantages that the function approximation problem of reinforcement learning for continuous state space is solved, meanwhile, a fitness track is introduced to effectively store an access path with correct experienced state actions, the generalization performance of the neural network is improved, and finally the convergence speed of the algorithm is increased.

The invention provides a supraventricular premature beat intelligent analysis method based on an attention conversion mechanism, and the method can improve the recognition accuracy of supraventricular premature beat heart beats, greatly saves the manual workload, and improves the recognition efficiency. According to the technical scheme, all RR interval stationary heart rate segments are found based on a time RR interval scatter diagram, all the RR interval stationary heart rate segments are connected based on a forward and reverse propagation algorithm, and a dominant heart rate curve corresponding to electrocardiogram data to be recognized is obtained; and constructing a rhythm classification and recognition model based on a convolutional neural network, recognizing the rhythm type of each heart beat, and judging the heart beat of each supraventricular premature beat type in the electrocardiogram data to be classified based on the dominant heart rate, the RR interval deviation degree, the rhythm type and the QRS wave form of each heart beat.

The invention relates to a wind turbine planetary gearbox fault diagnosis method and system based on an improved capsule network. The method comprises the following steps: collecting vibration signals of four states of a wind turbine planetary gearbox as four groups of original vibration signals; wherein the four states comprise a normal operation state, tooth surface abrasion, planet gear tooth breakage and rolling body bearing missing; carrying out point number segmentation on each group of original vibration signals, and constructing a sample data set; training an improved capsule neural network model by using the sample data set, and performing fault diagnosis on the planetary gearbox of the wind turbine generator under the severe working condition by using the trained improved capsule neural network. According to the method, an intelligent diagnosis network integrating feature extraction and fault diagnosis can be constructed, a large amount of manual operation is saved, and the method still has the characteristics of high diagnosis precision and strong robustness under severe working conditions. The method can be widely applied to the technical field of mechanical fault diagnosis.

The invention discloses a method and system for estimating indoor layout based on learning prediction. The method includes: constructing a training set, and using training samples in the training set to train a deconvolution network; the training samples are room layout diagrams and their The corresponding edge map, the room layout map and its corresponding edge map are used as the input and output of the deconvolution network respectively; the room layout map to be tested is input to the trained deconvolution network, and the predicted edge map is output; The vanishing point in the preset direction in the room layout diagram generates several sectors; then based on the predicted edge map, selects the sector with the local maximum edge intensity as the sampling sector from the generated sectors; the sampling sector is sampled , to obtain a series of candidate room layout estimation maps; then according to the similarity between the room layout estimation map and the obtained edge map, select the room layout estimation map closest to the predicted edge map from the candidate room layout estimation maps as the final layout of the room.

The invention discloses a rolling bearing life stage identification method based on MAMTL, and the method comprises the following steps: S1, carrying out the life stage division of the full life data of a rolling bearing, and dividing the full life data into four stages: a normal stage, an early degradation stage, a medium degradation stage, and a complete failure stage; s2, collecting the vibration acceleration of the rolling bearing of which the life stage division is completed in the whole life stage as a source domain sample set, and collecting the vibration acceleration of the rolling bearing to be identified as a target domain sample set; s3, training an MAMTL network, wherein the MAMTL is composed of an inner ring parallel network, an outer ring element learning network and a prototype network; and S4, identifying class labels of to-be-tested samples of the target domain: completing classification of the to-be-tested samples of the target domain by using the trained MAMTL, namely completing life stage identification of the rolling bearing. According to the method, a small number of non-equal life stage samples under the historical working condition of the rolling bearing can be used for carrying out high-precision life stage identification on the to-be-detected sample under the current working condition.

The present application provides a method and device for detecting abnormal traffic, a storage medium, and an electronic device, wherein the method includes: obtaining network traffic generated by a mobile terminal in a terminal set accessing a target web page in a web page set; Any one of the IP address and the webpage IP address of the target webpage, sort the access records of all network traffic according to the access time, and obtain multiple record sequences, among which, one record sequence contains the same terminal IP address according to the access time multiple access records, or multiple access records under the same web page IP address; by analyzing the time series data characteristics of multiple record sequences, the target record sequence in the multiple record sequences is determined, and the target record sequence is accessed Record the corresponding network traffic as abnormal traffic. Through the present application, the problem of poor generalization of abnormal traffic detection in the related art is solved.