630results about How to "Improve feature extraction" patented technology

In one embodiment, the invention is a method for generating geometric patterns from an image having a plurality of ridges and mesh points. The method comprising the steps of: establishing a mathematical model according to regional conditions in the image; converting the mathematical model into numerical equations; solving the numerical equations; and transferring the solutions of the numerical equations to respective regions of the image.

The invention provides a method for detecting an X-ray mammary gland lesion image based on feature pyramid network under transfer learning. The method comprises the steps: 1, establishing a source domain data set and a target domain data set; 2, establishing a deformable convolution residual network layer by a deformable convolution and extended residual network module; 3, establishing a multi-scale feature extraction sub-network based on a feature pyramid structure through a feature map up-sampling and feature fusion method in combination with a deformable convolution residual network layer;4, establishing a deformable pooling sub-network sensitive to the focus position; 5, establishing a post-processing network layer to optimize a prediction result and a loss function; and 6, migratingthe training model to a small sample molybdenum target X-ray mammary gland focus detection task so as to improve the detection precision of the network model on the focus in the small sample image. According to the method, a transfer learning strategy is combined to realize focus image processing in a small sample medical image.

The invention discloses a multi-feature fusion vehicle re-identification method based on deep learning. The method comprises five parts of model training, license plate identification, vehicle identification, similarity measurement and visualization. The method comprises the steps that a large-scale vehicle data set is used for model training, and a multi-loss function-phased joint training policyis used for training; license plate identification is carried out on each vehicle image, and a license plate identification feature vector is generated according to the license plate recognition condition; a trained model is used to extract the vehicle descriptive features and vehicle attribute features of an image to be analyzed and an image in a query library, and the vehicle descriptive features and the license plate identification vector are combined with the unique re-identification feature vector of each vehicle image; in the stage of similarity measurement, similarity measurement is carried out on the image to be analyzed and the re-identification feature vector of the image in the query library; and a search result which meets requirements is locked and visualized.

Disclosed is a stereo vision measuring device based on a binocular omni-directional vision sensor. Each ODVS composing the binocular omni-directional vision sensor adopts the design of mean angle resolution. The parameters of two image collection cameras are in complete accord and in possession of a pretty good symmetry, and can quickly realize the point-to-point matching. The device adopts a unified spherical coordinate in the process of data collection, processing, description and representation of space objects in terms of centering on human in visual space, and adopts the elements of distance sense, direction sense and color sense to express features of each characteristic point, thereby simplifying the complication of calculus, omitting the calibration of the cameras, facilitating the feature extraction and realizing the stereo image matching easily, finally realizing the purpose of high-effective, real-time, and accurate stereo vision measurement. The device can be applied in a plurality of fields of industrial detection, object identification, robot automatic guidance, astronautics, aeronautics, military affairs, etc.

The invention provides a vehicle license plate recognition method, and relates to a method for image recognition. The method comprises the following steps of preprocessing an image; partitioning a vehicle region according to color and texture characteristics; extracting a remarkable factor graph of a vehicle region diagram; extracting candidate vehicle license plates by an Adaboost classifier based on expanded Harr-like characteristic; determining the position of a real vehicle license plate from the candidate vehicle license plates; partitioning the marked vehicle license plate from the corresponding vehicle region original diagram; carrying out character segmentation according to structural characteristic; and carrying out character recognition based on the improved template matching method. With the method provided by the invention, the defects that the application scene of the traditional vehicle license plate recognition method is relatively single, some traditional vehicle license plate recognition methods are only suitably used for single vehicle license plate recognition of the single scene and difficultly used for multiple-vehicle license plate recognition of multiple scenes, and the recognition rate is easily affected by strong light, haze and weak light environments are overcome.

The embodiments of the present invention belongs to the computer vision technological field and provide a semantic segmentation network training method, an image semantic segmentation method and corresponding devices. The semantic segmentation network training method includes the following steps that: a to-be-trained image is acquired; the to-be-trained image is inputted into a pre-established semantic segmentation network, the front network layer of the semantic segmentation network is adopted to extract the features of the to-be-trained image, and a feature map containing the block, global and edge features of the to-be-trained image can be obtained; the feature image containing the block, global and edge features of the to-be-trained image is inputted to the rear network layer of the semantic segmentation network so as to be subjected to image pixel classification, so that a semantic segmentation image containing segmentation pixel types can be obtained; and the parameters of the semantic segmentation network are update according to the semantic segmentation image. Compared with the prior art, the method of the invention separately extracts and restores the edge features of theto-be-trained image, thereby improving the training recognition effect of the edge of a segmentation region.

The invention belongs to the field of image and video processing, and relates to a small target detection method based on multi-scale images and weighted fusion loss, and the method comprises the steps: extracting a plurality of groups of feature vectors from a plurality of different-scale images based on an improved Mask RCNN model, carrying out the fusion of the plurality of groups of feature vectors, and constructing a feature pyramid; generating a candidate detection box based on the feature pyramid and screening to obtain a suggested detection box; correspondingly returning the suggesteddetection boxes to the feature pyramid to generate feature maps of the suggested detection boxes, and performing aligned interception on the feature maps; inputting the aligned suggested detection boxes into a classifier layer to obtain category confidence coefficients and position offsets of the suggested detection boxes; in the test stage, screening a certain suggested detection box according tothe category confidence score of the suggested detection box, and performing non-maximum suppression; in the training stage, weighting the loss function calculated by detecting the small target feature layer and fusing with the loss function of detecting the large target layer and the middle target layer, thus the sensitivity of the model to the small target object is enhanced.

The invention relates to a design method of an establishing method of a rolling bearing intelligent diagnosis model based on a convolutional neural network. Firstly a one-dimensional vibration signalis mapped into two-dimensional image information, and the two-dimensional image data are used for training a network model; and then the structural parameters of the convolutional neural network in the application process are analyzed and the better network parameters are selected so that the convolutional neural network structure having high mechanical fault classification capacity is obtained. Accurate identification and classification of the mechanical fault under the complex condition of different loads and different rotating speed can be realized; and the convolutional neural network model can greatly enhance the feature extraction capacity of the neural network by establishing the multilayer network, mastering of manual technology for mass signal processing and dependence on the diagnosis experience can be overcome, the fault features can be directly acquired from the original time domain signal through the learning mode to perform diagnosis, and adaptive extraction of the faultfeatures and intelligent diagnosis of the state of health under the condition of TB level data volume per hour can be realized.

The invention discloses a feature pyramid-based remote-sensing image time-sensitive target recognition system and method. The system comprises a target feature extraction sub-network, a feature layersub-network, a candidate area generation sub-network and a classification and regression sub-network, wherein the target feature extraction sub-network is used for carrying out multiple layers of convolution processing on a to-be-processed image and outputting the convolution processing result of each layer as a feature layer; the feature layer sub-network is used for overlapped the last feature layer and the current feature layer to obtain the current fused feature layer, wherein the topmost fused feature layer is a topmost feature layer; the candidate area generation sub-network is used forextracting candidate areas from different layers of fused feature layers; and the classification and regression sub-network is used for mapping the candidate areas to different layers of fused featurelayers so as to obtain a plurality of mapped fused feature layers, and carrying out target judgement on the plurality of mapped fused feature layers so as to output a result. According to the systemand method, the hierarchical structures of feature pyramids are utilized to ensure that all the scales of features have rich semantic information.

Provided is a speaker separating model training method. The method includes steps: obtaining and preprocessing multiple pieces of audio data; extracting audio characteristics of the preprocessed audiodata; inputting the audio characteristics to a preset neural network for training to obtain vector characteristics; calculating a first similarity between a first vector characteristic and a second vector characteristic of a first speaker; calculating a second similarity between the first vector characteristic of the first speaker and a third vector characteristic of a second speaker; and calculating a loss function value of the first similarity and the second similarity, and when the loss function value is less than or equal to a preset loss function threshold, finishing a training process of a speaker separating model, and updating parameters in the model. The invention also provides a two-speaker separation method, a terminal and a storage medium. The characteristic extraction capability of the input voice data by the model can be substantially enhanced, the accuracy of two-speaker separation is improved, and better separating effect can be achieved especially in separation tasks with long-time dialogues.

A twin neural network training method for face verification comprises: a training sample set is prepared; the images in the training sample set are normalized in size and then input to the artificialneural network for training. Artificial neural network consists of two identical sub-neural networks. The training samples are divided into datasets data_p and data_p which are equal in number. The datasets data_p and data_p are sent to two sub-neural networks to extract the eigenvectors of the samples. By comparing the loss function to realize the iterative optimization of the neural network, until the iterative number reaches the set value, then jump out of the iteration, at this time the trained artificial neural network is the twin neural network for face verification; The contrast loss function represents similarity between two sets of eigenvectors.

The invention relates to the technical field of computer vision, and provides a pulmonary nodule image classification method when uncertain data is contained in a data set. The method comprises the following steps: firstly, collecting a pulmonary nodule CT image set, determining the category of the image through a majority voting principle by utilizing an expert voting method, and preprocessing toobtain a pulmonary nodule CT image data set; then, based on a knowledge distillation method, constructing a pulmonary nodule image classification model comprising a teacher model and a student model;next, obtaining a determined tag data set, training a teacher model on the determined tag data set, and calculating a soft tag on the pulmonary nodule CT image data set; then, training a student model on the data set combining the hard label and the soft label; and finally, inputting the preprocessed CT image to be classified into the trained lung nodule image classification model to obtain the category of the lung nodule image classification model. According to the method, the uncertain label data in the data set can be effectively utilized, the accuracy and efficiency of pulmonary nodule diagnosis are improved, and the usability and robustness are high.

The invention discloses a fault diagnosis method for a rolling bearing based on LCD-MF (Local Characteristic Scale Decomposition)-(Multifractal). The fault diagnosis method comprises the following steps: firstly, obtaining a plurality of intrinsic scale components (ISC) by using LCD, and selecting useful ISCs to respectively calculate Teager energy operators (TEO) to obtain instantaneous amplitude and instantaneous frequency corresponding to each ISC; secondly, analyzing the instantaneous amplitude of each ISC by using multifractal detrended fluctuation analysis (MFDFA), and extracting a generalized Hurst index of each ISC as the multifractal characteristic of each ISC; thirdly, performing dimension reduction by using principal component analysis (PCA) and taking a result obtained by the PCA as a fault feature vector; acquiring a vibration signal of the rolling bearing during work under variable working conditions in real time, processing the acquired vibration signal according to the steps to obtain a corresponding fault feature vector M, and identifying the fault feature vector to realize fault detection and fault positioning of the rolling bearing.

The invention discloses a real-time bridge monitoring system which comprises a kernel control processing module (7), an environment detection sensor unit (2), a flow detection sensor unit (3), a structure detection sensor unit (4), a snapshot camera (6), a GPS time service unit (1) and a wireless data transmission unit (5). The environment detection sensor unit (2), the flow detection sensor unit (3), the structure detection sensor unit (4), the snapshot camera (6), the GPS time service unit (1) and the wireless data transmission unit (5) are connected with the kernel control processing module (7). According to the real-time bridge monitoring system, time signals of a GPS are introduced in a measuring result, time marking codes are additionally arranged for output data of sensors and output data of the camera, the accurate synchronization of the data is achieved, and a synchronization standard is provided for subsequent data processing and analyzing. Meanwhile, according to the analysis of the data of structure detection sensors and the judgment of threshold values, whether a target object has dynamic deformation is determined; the snapshot camera is triggered to work according to the threshold values of signals, the dynamic deformation is monitored completely, and a bridge is monitored comprehensively and reasonably; the real-time bridge monitoring system has the advantage of saving data resources and reducing the pressure of wireless data transmission and the pressure of data storage and backup.

The invention provides a method for constructing and training a human face identification feature extraction network. The method comprises the steps of constructing a feature extraction network and a metric learning dimension reduction network, wherein an output of the feature extraction network is an input of the metric learning dimension reduction network; training the feature extraction network based on all sample sets to output feature sets; screening the feature sets by utilizing semantic sampling to obtain a pure sample set; and training the metric learning dimension reduction network based on the pure sample set. Through a natural human face identification network constructed by the method, the feature representation capability can be improved, so that feature information in data can be fully mined and an original human face picture can be accurately identified.

The invention discloses an optimization method of a convolution neural network. In a convolution neural network training process, a filter group with poorer effects and a filter group with better effects are selected out, and parts of coefficients in the filter group with poorer effects are replaced by parts of coefficients in the filter group with better effects, parameter update between convolution layer filters is achieved. After a modified network has been trained for a certain time, error magnitude generated by the network after and before the modification is compared, whether the previously achieved parameter update of the filters is valid or not, selection is performed in the two networks, and the network with good performance is used for following trainings. By continuously repeating the above processes, a network with better performance can be obtained through the training and the feature extraction ability of the network is improved compared with the traditional network.

The invention relates to an X-ray weld defect detection method based on a convolutional neural network, and the method comprises the following steps that a weld image data set containing different types of weld defects is built, and all weld images in the data set are marked with weld labels; an AF-RCNN model is established, wherein the AF-RCNN model comprises a backbone network module, a region generation module and a target classification and position regression module; the backbone network module adopts a residual network (ResNet) and a feature pyramid network (FPN) structure, an efficient convolution attention module is introduced between the residual network (ResNet) and the feature pyramid network (FPN) to enhance the learning ability of the network for unobvious defects and small target features, and a CIOU loss function is introduced to enhance the positioning ability of aiming frames; and an AF-RCNN model is trained by using the established data set, wherein the AF-RCNN model is used for classifying and positioning weld defects. The accuracy of all defects reaches 94% or above, and the detection speed is 11.65 FPS.

The invention discloses a single classifier anomaly detection method based on a multilayer random neural network. The the method comprises: only inputting a training data set of a normal class; through multilayer ELM-AE autoencoder and decoding processing, input sample data obtaining a reconstructed characteristic value; inputting the reconstructed characteristic value into the last layer of ELM to obtain actual output; sorting the obtained distance error vectors of the actual output and output tags from large to small, and determining a threshold for separating a normal class from an abnormalclass according to a set threshold parameter; and finally, inputting the test data into the multi-layer random neural network single classification abnormity detection model, and testing the recognition effect of the model. According to the method, main information is extracted more quickly and efficiently, dimensionality reduction is carried out, and then recognition and classification are carried out. And the speed is higher, the accuracy is higher, and the generalization performance is better. The method is not only suitable for small data sets, but also suitable for high-dimensional largedata sets, and has universality. And the method has important significance for practical application in future.

The invention relates to a surface defect detection method based on a cascaded convolutional neural network, and the method comprises the steps: building a defect detection network based on Cascade R-CNN, and carrying out the detection of an image of a to-be-detected product collected by an industrial camera in real time through the trained and optimized defect detection network. According to theinvention, the Cascade R-CNN is improved into a defect detection network, defect regions and defect features do not need to be manually extracted, and classification is carried out while defects are positioned; Cascade R-CNN is used as a basic detection network architecture, and the excellent detection performance of the Cascade R-CNN makes the positioning and classification precision of the defects more advantageous, and the feature extraction capability is enhanced through ResNeXt; the FPN is used for detecting small defects, deformable convolution is adopted, anchor frames are added to adapt to various shapes and sizes of the defects, the defects have large enough receptive fields, the detection effect of various extreme defects is improved, the threshold value of non-maximum suppression is adjusted, and the detection accuracy is further improved.

The invention relates to an attention mechanism-based image aesthetics quality evaluation method, and belongs to the technical field of computer vision. The method comprises the steps of firstly processing training data, then designing a network structure model, adopting a lightweight deep network as a backbone network, and integrating the backbone network into an attention mechanism module; designing a loss function for training the network based on a data equalization thought; finally, training a network structure model by using the processed training data to obtain a network model capable of automatically evaluating the aesthetic quality of an image; and performing aesthetic scoring on the input picture based on the model, and applying the model to shooting to assist a user to shoot a more beautiful picture in real time. Compared with the prior art, the network structure model adopted by the invention can more effectively extract the characteristics of images, the adopted loss function greatly enhances the data learning ability of the model. Compared with other methods, the accuracy is improved, and the parameter quantity of the model is reduced.

The invention provides a fine-grained image classification method based on an attention transfer mechanism. The fine-grained image classification method is used for improving the classification precision of fine-grained images. The method comprises the following implementation steps: obtaining a training sample set and a test sample set containing fine-grained images, constructing a global perception network model and an attention transfer network model, training the global perception network and the attention transfer network by using the training sample set, and classifying the test sample set by using the trained global perception network and attention transfer network. According to the invention, the global perception network and the attention transfer network are designed to carry outjoint feature extraction on the image; and, on the basis of extracting global features and discriminant region features, the semantic correlation between different discriminant regions is further extracted by the network, so that the feature extraction capability of the network is enhanced, and the classification accuracy of fine-grained images is effectively improved.

The invention relates to a device for assisting judicial case judgment based on machine learning, which utilizes a large amount of document data and trains a model to learn the relationship between case fact description and the fine range and relevant legal provisions, and realizes the prediction of the fine range and the law label of any given case fact description text. The invention relates toa device for assisting judicial case judgment based on machine learning. Including: defining the proper nouns in the description of the facts of a given case and dealing with them; Extracting multiplesemantic features from the text to achieve a deeper level of semantic representation; Machine learning method based on multi-label classification is used to classify the law items and obtain the lawlabels related to the description text of the case facts. Single-label classification training model based on machine learning predicts the range of possible fines in related cases. The invention applies machine learning to the judicial field for the first time, realizes deeper semantic representation by multiple feature extraction modes, improves the accuracy and generalization ability of the training model well, has higher reference significance for the final judgment of a case, and is conducive to the realization of the same case and the same judgment.

The invention relates to a novel binocular stereoscopic vision measuring device, which comprises two omnibearing vision sensors provided with the same imaging parameters, a connecting unit and a microprocessor which is used for performing three-dimensional stereoscopic vision reconstruction on images of the two omnibearing vision sensors, wherein the omnibearing vision sensors comprise hyperbolic mirror surfaces, upper covers, stand bars, transparent glass surfaces, supplementary lens frames and image units; the upper covers are arranged on the hyperbolic mirror surfaces; the stand bars are round platforms with thick upper parts and thin lower parts; upper ends of the stand bars are arranged inside small holes in bottom centers of the hyperbolic mirror surfaces, and lower ends of the stand bars are arranged inside mounting holes in the centers of the transparent glass surfaces which are embedded into the supplementary lens frames; the stand bars are perpendicular to the transparent glass surfaces; and the image units are positioned inside the supplementary lens frames. The novel binocular stereoscopic vision measuring device simplifies the complexity of calculation of stereo matching and so on, saves the operation of camera calibration, is convenient to perform feature extraction, is easy to realize stereo image matching, and finally achieves the aim of high-efficiency, real-time and accurate stereoscopic vision measurement.