69results about How to "Solve the problem of low recognition accuracy" patented technology

The invention discloses a reading method of a pointer type meter based on depth learning, belonging to the field of depth learning and computer vision. The method of the invention utilizes Mask- RCNNobject detection and instance segmentation algorithm to divide the dial and pointer images firstly, then correcting the dial by perspective transformation, Then using PCA (Principal Component Analysis) algorithm to fit the segmented instrument pointer, Then judging the direction of the pointer according to the center coordinate of the smallest oblique circumscribed rectangle of the pointer and thecenter coordinate of the dial; Finally, calculating the pointer reading according to the slope and direction of the pointer by using the angle method. As that method of the invention can accurately classify the type of the instrument, high precision pixel level segmentation of the pointer and dial, under non-uniform illumination conditions, different scales still have good robustness, to solve the traditional pointer instrument recognition field of the dial and pointer positioning difficulties, uneven light, mirror reflection, blurred pictures caused by low recognition accuracy.

The invention relates to a method for detecting driving fatigue based on electroencephalogram. The method specifically comprises the following steps: (1) selecting characteristic quantity related to a driving fatigue state; (2) performing electroencephalogram power spectrum analysis on the driving fatigue state; (3) performing electroencephalogram sample entropy analysis on the driving fatigue state; (4) performing electroencephalogram Kc complexity analysis on the driving fatigue state; (5) using support vector machine (SVM); (6) using least squares support vector machine (LS-SVM); (7) setting model training parameters (C and g) through a particle swarm optimization (PSO) algorithm. According to the method disclosed by the invention, electroencephalograms extracted in different driving states are respectively researched from a power spectrum angle by using related methods in non-linear dynamics, so that excellent effects can be achieved in accuracy and reliability.

The invention discloses a Caffe architecture based deep learning license plate character recognition method. The method includes a classifier training process and a character recognition process. The classifier training process includes processing characters, separating the characters into a Chinese character set and a non-Chinese character set and constructing a Caffe architecture for learning network structures, and training respectively for obtaining corresponding classifiers. The character recognition process includes creating an index table in advance and processing a captured license plate image, recognizing the captured license plate image with the corresponding classifiers and then obtaining a recognition result through inquiring an index table, and obtaining the final license plate recognition result after combination in sequence. According to the invention, on the basis of deep learning based on the Caffe architecture, a problem of insufficient precision of recognition on inclined, fractured and similar characters in the prior license plate character recognition method is solved and license plate character recognition precision is improved substantially.

The invention relates to a Chinese author identification method based on a double-layer classification model and a device for realizing the Chinese author identification method, belonging to the field of information security. Aiming at the problem of low identification accuracy caused by excessive authors, an author grouping layer is added in an author identification model; each author is represented into an author vector; authors are grouped by a clustering algorithm; a second layer is an author identification layer; a dependence relationship, a function word, a punctuation mark and a word class mark are extracted from the second layer to use as characteristics; and author identification is carried out in the group. According to the method or the device, the problem that the identification accuracy is lowered because of excessive authors can be effectively solved. Meanwhile, with a proposed characteristic dimensionality reduction and optimization method based on a main ingredient analysis method, the problem that the identification accuracy is affected by noise comprised by a high-dimensionality characteristic vector is solved. The Chinese author identification method can be applied to the author textual research field of a literature and also can be applied to the field of information security, such as copyright protection.

The invention discloses an identity recognition method based on multi-mode fusion of face characteristics and palm print characteristics of a single image. The identity recognition method is characterized by comprising the following steps: 1, acquiring the face and the palm print of one same person in one image, and establishing a database; 2, respectively detecting and partitioning face areas and palm print areas of the image, thereby obtaining ROI areas; 3, respectively calculating and authenticating chi-square distance of face characteristics and palm print characteristics of an authenticated image and each image in the database according to a face recognition algorithm and a palm print recognition algorithm; 4, fusing two characteristic distances according to a multi-mode characteristic fusion algorithm, thereby achieving identity recognition of persons. As the face characteristics and the palm print characteristics are fused, the accuracy rate of identity recognition is increased.

The invention relates to a sugarcane distribution identification method based on optical remote sensing data, which comprises the following steps: acquiring phenological information of sugarcane and history Sentinel-2 and calculate NDVI for each scene image; According to phenological information, phenological stages are divided and several NDVI maximum images corresponding to each phenological stage are synthesized. Cubic dimensionality reduction of a plurality of NDVI maximum images to one-dimensional images; The gradient operator is used for calculating the gradient for the one-dimensional image to obtain an edge amplitude image; A watershed segmentation algorithm is used to extract the distribution of sugarcane from the edge amplitude image. The sugarcane distribution identification method based on the optical remote sensing data solves the problem of low identification accuracy caused by land fragmentation and cloud disturbance, and has the advantages of accurately reflecting the spatial distribution of sugarcane and being beneficial to sugarcane management and planning.

The invention discloses a power transmission line multi-target detection method which mainly performs target identification on three types of insulators, two types of insulator defects, a shockproof hammer, an interphase rod and a bird nest, and belongs to the technical field of power transmission line target identification. The method comprises the following steps: firstly, increasing the order of magnitudes of sample data by utilizing a sample generation technology, enhancing the detection effect of deep learning, then dividing newly generated experimental data into a training set, a test set and a verification set, constructing a PyTorch deep learning environment, and establishing ResNet101 and 6 layers of FPN networks to extract image features by adopting four paths of GPU distributedtraining; wherein the output of the ResNet101 and six layers of FPN networks serves as the input of the RPN network to train a Cascade R-CNN deep learning network model, and finally target recognitionis achieved according to a Softmax classifier and a frame regression result. The method is high in operation speed, high in target recognition accuracy and high in multi-target recognition capability.

The invention relates to a robust recognition method for an AR code marked on a cylinder, which comprises the steps of performing binarization on an AR code image, extracting the contour of a binarization image of the AR code, and performing straightening processing on an AR code regional image in the contour; performing arc distortion detection on a straightened AR code region, and rotating the straightened AR code region so as to enable a concave arc edge to face upward, and straightening the concave arc edge; straightening a convex arc edge by using a local adaptive compression algorithm so as to acquire an AR code region with the contour being rectangular; and performing nonlinear marking recognition on the AR code region. The robust recognition method provided by the invention solves problems that some traditional methods are low in recognition accuracy for the AR code and that the AR code cannot be correctly recognized when arc distortion is generated, and improves the robustness of AR code recognition.

The invention relates to a human body behavior recognition method of a non-local double-flow convolutional neural network model. Two shunt networks are improved on the basis of a double-flow convolutional neural network model; a non-local feature extraction module is added into the spatial flow CNN and the time flow CNN for extracting a more comprehensive and clearer feature map. According to themethod, the depth of the network is deepened to a certain extent, network over-fitting is effectively relieved, non-local features of a sample can be extracted, an input feature map is subjected to de-noising processing, and the problem of low recognition accuracy caused by reasons such as complex background environment, diverse human body behaviors and high action similarity in a behavior video is solved. According to the method, an A-softmax loss function is adopted for training in a loss layer; on the basis of a softmax function, m times of limitation is added to a classification angle, andthe weight W and bias b of a full connection layer are limited, so that the inter-class distance of samples is larger, the intra-class distance of the samples is smaller, better recognition precisionis obtained, and finally a deep learning model with higher identification capability is obtained.

The invention provides an anti-collision early warning device and method for a top beam of a hydraulic support and a roller of a coal mining machine. The early warning device comprises a movable monitoring early warning device and a laser receiving device. The mobile monitoring and early warning device mainly comprises a bearing box, a rotary laser range finder, a reference seat, a first hinge lug plate, a walking supporting plate, a walking wheel, a supporting wheel, a wheel shaft, a data storage module, a calculation module, a control module and an early warning module. The laser receiving device mainly comprises a fixing plate and a laser receiving module; the movable monitoring and early warning device is provided with four walking wheels and four supporting wheels, the movable monitoring and early warning device is placed on the cable trough, when the walking wheels make contact with a side plate of the cable trough, the walking wheels can roll on the side plate of the cable trough to drive the movable monitoring and early warning device to move, and when the cable trough inclines, the walking wheels can roll on the side plate of the cable trough to drive the movable monitoring and early warning device to move. The supporting wheels are in contact with the side plates of the cable trough, so that the mobile monitoring and early warning device can move smoothly. By adopting the anti-collision early warning device and method for the hydraulic support top beam and the coal mining machine roller, interference detection and early warning of the roller and the hydraulic support top beam in the coal cutting process of the coal mining machine can be achieved, and the anti-collision early warning device and method have the advantages of being easy to operate, high in adaptability, high in precision and the like.

The invention relates to a method and a device for identifying a target object, a storage medium and a vehicle. The method comprises the following steps that: obtaining a current frame of image and aprevious frame of image in a preset range around the vehicle; according to the current frame of image and the previous frame of image, obtaining target characteristics, determining the first positioninformation of the target characteristics in the current frame of image, and determining the light stream of the target characteristic according to the first position information; according to the first position information and the light stream, carrying out clustering on the target characteristic to obtain a first target object; obtaining the second position information and the speed informationof a target to be clustered in the preset range around the vehicle, and according to the second position information and the speed information, clustering the target to be clustered to obtain a secondtarget object; and according to the first position information, the second position information, the light stream and the speed information, identifying whether the first target object and the secondtarget object are the same target object or not.

The embodiment of the invention discloses a domain name recognition method and device, a domain name recognition model generation method and device and a storage medium, relates to the technical fieldof networks, the method comprises the following steps of: obtaining a one-dimensional ordered vector corresponding to the domain name to be classified, wherein the one-dimensional ordered vector comprises a number obtained by performing character-number conversion on the domain name to be classified; according to the one-dimensional ordered vector and a preset neural network model, determining adetection value of the to-be-classified domain name, the neural network model comprising a neural network model which is trained according to a preset loss function and an optimization algorithm and satisfies a preset condition, comparing the detection value with a specified intermediate value, and determining a detection result of the to-be-classified domain name according to a comparison result.Through the embodiment of the invention, the recognition accuracy can be improved.

A fingerprint identification method of using the image field-based non-rigid registration disclosed by the present invention comprises the following steps of S1 obtaining the image fields of a floating fingerprint and a reference fingerprint, namely extracting the respective direction field and gradient field information of two frames of to-be-registered fingerprint images, wherein the two framesof to-be-registered fingerprint images are separately a floating fingerprint image and a corresponding reference fingerprint image; S2 calculating a deformation field corresponding to the floating fingerprint image according to the direction field and gradient field information of the two frames of to-be-registered fingerprint images; S3 applying the deformation field to the floating fingerprint image to obtain a registered fingerprint image; S4 using the registered fingerprint image to match the reference fingerprint image. The fingerprint identification method of the present invention can solve the problem that due to the fingerprint non-rigid deformation, the fingerprint identification precision is lower.

The invention relates to a pipeline leakage detection method, and belongs to the technical field of intelligent sensing mechatronics. The method specifically comprises the steps that S1, audio data in a pipeline are collected in an ultrasonic sensing mode, and key information is obtained; s2, MELGAN is used to expand the collected small sample practical audio data so as to enhance the generalization ability of the trained model; s3, the small sample audio data and the sample enhanced data are combined and sent to a yov3 model for training so as to obtain a pipeline leakage detection identification model with improved identification accuracy. When the pipeline leaks, the leakage position in the pipeline can be accurately judged in time, the leakage phenomenon of the pipeline can be efficiently and accurately found, the pipeline can be timely processed, and targeted guidance can be obtained; and the method can effectively solve the problem that the recognition accuracy is not high when a small sample training model is used.

The embodiment of the invention provides a lip language identifying method and device, and relates to the technical field of big data. The lip language identifying method includes the steps of obtaining multiple frames of facial images of a user, determining a plurality of lip key points in the each frame of facial image and coordinates corresponding to each of the lip key points; and generating lip language coding corresponding to the multiple frames of facial images according to the coordinates corresponding to each of the lip key points in the each frame of facial image, and inputting the lip language coding into a preset lip language identifying model so as to identify the content of a lip language. Therefore, generation of the corresponding lip language coding is achieved through thecoordinates of the lip key points in the multiple frames of facial images, and further, the content of the lip language is identified through the lip language coding, so that the influences of skin colors, textures and other factors in the facial images on lip language identification are avoided, and the generalization ability and recognition accuracy of the lip language identifying method are improved. Therefore, the lip language identifying method and device can solve the technical problem of low accuracy of lip language identification in the prior art.

The present invention relates to a speech search method which comprises a step of searching a speech signal of a current frame according to a WFST network and a previous stage search result and obtaining a search result at a current stage, a step of resetting the search state of the WFST network if the current stage search result is matched with preset template information, a step of carrying outpre-search through the WFST network with the resetting of the search state according to the preset template information matched with the current stage search result to obtain a template path network,and a step of searching a speech signal of a next frame according to the template path network and the current stage search result until the search results of speech signals of all frames are outputted. The invention also discloses a speed search system. By resetting the search state of the WFST network when the current stage search result is matched with the preset template information, thus thepre-search is carried out in the WFST network with the resetting of the search state according to the preset template information, the template path network is obtained, and the search of the speech signal of the next frame is continued according to the template path network. The accuracy of speech recognition is greatly improved.

The invention relates to the technical field of target identification, and discloses an expression recognition method based on BN parameter transfer learning. The method includes constructing a facialexpression recognition BN model structure according to the relationship between the facial expressions and the action unit tags; secondly, calculating BN initial parameters by utilizing the BN parameters calculated by the human face source domain data set and the human face target domain data set respectively, obtaining final human face expression recognition BN parameters according to a migration mechanism, performing BN reasoning by utilizing a reasoning algorithm in a BN theory, and recognizing facial expressions. According to the invention, a transfer learning mechanism is fully utilizedto apply knowledge learned in a certain field to different but related fields; the method can effectively solve the problem of insufficient data volume of facial expression modeling samples caused byillumination, shooting angles and the like in facial expression recognition, reduces the influence of insufficient samples on parameter learning precision and recognition results, and can be widely applied to noisy and uncertain environments in which a large amount of face target data is difficult to obtain.

The invention discloses a pedestrian re-identification algorithm implementation method based on HSV and SDF. The method comprises the following steps of acquiring pedestrian video data with a camera;extracting the moving object by using the discrete fourier and local frequency domain features, and generating a pedestrian picture library; selecting a pedestrian picture from the pedestrian picturelibrary, converting the RGB three-channel picture into a picture represented by an HSV color space; distinguishing the pedestrian target and the background through a Graph Cut algorithm, and blockingthe pedestrian targets; calculating the HSV histogram by adopting a spatial distribution coverage operator and a color bilateral operator, obtaining a pedestrian feature descriptor, and calculating the similarity of the pictures by using Euclidean distance; and sorting the pedestrian pictures in the pedestrian picture library with a penalty function and outputting the first six pedestrian picturesto obtain the final result set of pedestrian detection. The method can effectively solve the problem of low detection precision existing in the current pedestrian re-identification, has the advantages of clear algorithm, easy understanding and high pedestrian re-identification precision.

The invention discloses a bottom-up optical character recognition method suitable for a terminal strip. The method comprises the steps: acquiring a transformer substation terminal strip content imageand performing preprocessing; adopting a bottom-up method for the preprocessed image; employing a CAM thermodynamic diagram to assist a VGG16 to detect fine-grained characters, judging whether the characters are in the same text line or not according to the distance and angle information between the characters, then adding a long-segment memory network LSTM into a detection network, and storing the context features of the text line to finally form a coarse-grained text area; in the identification network ResNet, taking CTC as a loss function, inputting the characteristic information into a training model, performing greedy coding on a model output result, and finally outputting a terminal strip identification result. According to the invention, the problem of low identification accuracy possibly generated by the conventional optical character identification technology in the actual application scene of the transformer substation terminal strip is solved, and the label of the cable sleeve of the transformer substation terminal strip can be rapidly and accurately identified.

The invention relates to the technical field of text recognition in artificial intelligence, and provides a training method and device for a text similarity recognition model, and related equipment, and the method comprises the steps: obtaining a plurality of first sample groups comprising a first text sample and a second text sample; taking an element of which the literal similarity with the first text sample reaches a preset threshold value as a third text sample; labeling the third text sample to obtain a negative text sample, and forming a plurality of second sample groups; representing the samples in each second sample group with representation vectors; calculating a first similarity and a second similarity; and according to the first similarity and the second similarity, adjusting the parameters, and repeatedly obtaining the representation vector to the step to obtain a trained text similarity recognition model. Through the implementation of the text similarity recognition methodand device, the problem that in the prior art, a text similarity recognition method is low in recognition accuracy can be solved. Meanwhile, the invention also relates to a blockchain technology, andthe first sample group and the second sample group can be stored in the blockchain node.

The invention discloses a wood counting method based on deep learning, and the method comprises the specific steps of photographing a set number of wood pictures, marking the wood contour in the images, and forming a data set; inputting the data set into a Mask RCNN model for training; performing preprocessing operation on the to-be-detected picture, and enabling the picture to be clearer by using an image enhancement algorithm; inputting the preprocessed to-be-detected picture into the trained Mask RCNN model to obtain a mask area of the wood section and wood area frame coordinates; performing overlapping judgment on the wood area by utilizing the coordinates of the wood area frame, and deleting the area coordinate points which are judged to be overlapped; carrying out false detection judgment on the area around the wood by utilizing the coordinates of the wood area frame, and deleting the area coordinate points of the wood which is judged to be false detected; and counting the coordinates of the remaining areas to obtain the number of the woods. The invention is not interfered by the environment, the robustness of the deep network is high, and the invention is more suitable for the actual production environment.

The invention discloses an anomaly detection algorithm based on relative density and belongs to the fields of machine learning and data mining. According to the anomaly detection algorithm, a locally relative density method is adopted based on the nearest neighbor ideology, an abnormal point is judged according to the difference between density of data points and density of neighbors of the data points, and the larger the relative density difference between a given data point and a neighbor, the higher the abnormity of the data point. Compared with a traditional density-based mode, the relative density method has higher accuracy, the problem that a local anomaly cannot be detected through a distance-based method can be solved, and the defect that a density-based method is invalid to sparse data can be overcome. Points of different anomaly types can be detected for different data.

The invention provides a casting identification character recognition method based on CRNN-CTC. The method comprises the following steps: collecting pictures containing casting identification characters to construct a data set; carrying out data augmentation on a data set, and solving the problem of small number of pictures by utilizing a method of rotating, adding noise and adjusting brightness and contrast so as to enhance the robustness of the model; preprocessing a data set picture, inputting the preprocessed data set picture into the established network model, carrying out feature extraction on the picture through a CNN, then outputting features to an RNN, and carrying out transcription through CTC; and calculating the loss function of the CTC, continuously optimizing the network model through back propagation, and ending the training until the best prediction effect is achieved. According to the OCR recognition method combining CRNN and CTC in deep learning, the recognition accuracy can reach 98.8%, and the model obtained through training has good generalization ability and fault-tolerant ability.

The embodiment of the invention provides a license plate recognition method and device based on deep learning, and the method comprises the steps: obtaining a plurality of image frames containing vehicle license plate information, extracting the features of each image frame through a convolutional neural network, and obtaining the feature representation of each image frame; detecting the feature representation of each image frame through a predetermined target detection network to obtain the category and position information of the license plate detection frame of each image frame, and segmenting the license plate of each image frame to obtain a feature map of each segmented image frame; obtaining original license plate label information in each image frame, and training to obtain a license plate recognition model; and inputting a to-be-recognized picture into the license plate recognition model for license plate recognition, and recognizing to obtain license plate characters in the to-be-recognized picture. According to the method, the problem of low license plate recognition precision caused by a large inclination angle is effectively solved, the characteristics of the license plate are fully utilized, and the license plate recognition precision and recognition efficiency are greatly improved.

The invention discloses an animal target detection method based on a single-order deep neural network, and belongs to the field of visual detection. In the animal breeding process, an adopted manual counting method is large in workload, long in consumed time and prone to counting errors. The animal target detection method based on the single-order deep neural network comprises the steps of animal data sample collection, animal data sample labeling, VOC data set making, animal sample data set training, detection model construction, model performance adjustment and performance evaluation, and animal individual target detection in a picture by using the adjusted model. The improved animal target detection algorithm provided by the invention can effectively solve the problem of low recognition precision caused by environmental influence and animal shielding in a pasture while ensuring the detection speed, so that the animal individual can be accurately detected.

The invention provides a vehicle axle number determination method and device, and the method comprises the steps: obtaining the infrared video data of a monitoring region, carrying out the image enhancement of a current frame image of the infrared video data, and obtaining an image enhancement image of the current frame image; determining a foreground image of the current frame image according tothe image enhancement image; determining a single vehicle image in the current frame image according to the foreground image and the image enhancement image; and determining the axle number of each vehicle in the current frame image according to the single vehicle image. The problem of low identification precision of a vehicle axle number identification mode in related technologies can be solved,the vehicle axle number identification in a highway free flow scene is performed by adopting an infrared thermal imaging technology, and the vehicle axle number identification precision is improved.

The invention discloses an information point recognition method and device and electronic equipment, and relates to the technical field of computer vision and deep learning. According to the specificimplementation scheme, the method comprises the steps of obtaining first text information and a first image of a first information point, and obtaining second text information and a second image of asecond information point; based on the first text information and the second text information, determining text similarity of the first information point and the second information point; determiningthe image similarity of the first information point and the second information point based on the first image and the second image; and determining whether the first information point and the second information point are the same information point based on the text similarity and the image similarity. According to the technology provided by the invention, the problem of relatively low identification accuracy of an information point identification technology is solved, and the accuracy of information point identification is improved.

The invention discloses a power plant water supply pipeline leakage detection method and system. The method comprises the following steps: collecting a video image and performing background preprocessing on the video image to acquire a pipeline leakage candidate area image; processing the candidate area image according to an automatic data enhancement technology and a cost sensitive learning method, and training to obtain an OfficientDet classification recognition model; determining m candidate area images and inputting the m candidate area images into an OfficientDet classification recognition model to obtain a detection result; and determining a plurality of corresponding video images in the video images according to a detection result, and performing secondary judgment on the detection result by adopting a Hausdorff Distance image similarity measurement method to obtain a final detection result. According to the method, the detection result is obtained by preprocessing the background data and constructing the OfficientDet classification recognition model, and secondary judgment is further performed on the detection result, so that the accuracy of pipeline leakage recognition is improved.

The invention relates to a voice search optimization method. The method comprises: an input signal is obtained and a matched analysis is carried out on the input signal and a preset template; if one preset template matching the input signal exists, a search state of a WFST network is reset; according to the WFST network after search state resetting, pre searching is carried out on the preset template matching the input signal to obtain an optimized WFST network corresponding to the preset template; and on the basis of the optimized WFST network, searching of all frames of voice signals is completed to obtain a searching result. In addition, the invention also discloses a voice search system. When the input signal matches the preset template, the search state of the WFST network is reset; pre searching is carried out in the WFST network after search state resetting according to the preset template, so that the optimized WFST network is obtained; and according to the optimized WFST network, all frames of voice signals are searched and the searching result is outputted. Therefore, a problem of low identification accuracy in the traditional voice identification mode is solved; and thusthe voice identification accuracy is improved substantially.

The invention discloses a lane line extraction method and device, a vehicle and a storage medium. The method comprises the following steps: acquiring a current position point of a vehicle and high-precision map data, wherein the high-precision map data comprises lane line data and planned path data of the vehicle; based on the current position point and the planned path data, extracting target lane line data within a set lane line length in front of a lane where the vehicle is located from the lane line data; and performing curve fitting on the target lane line data to obtain a lane line. Through the technical scheme of the invention, the lane line can be extracted based on the high-precision map, and the problem that the lane line identification accuracy is not high due to the constraint of weather, ground light reflection and lane line wear degree in the traditional method for identifying the lane line on the ground through image acquisition equipment is solved. The invention can be independently used or used in combination with a traditional lane line recognition method, and the effects of reducing the constraint conditions of lane line recognition and improving the lane line recognition accuracy are achieved.