41results about How to "Strong feature expression ability" patented technology

The invention provides a vehicle attribute identification method based on a multi-task convolutional neural network. The method comprises a training process and an identification process. Particularly the method comprises the steps of acquiring a picture of a to-be-identified vehicle, designing a multi-task convolutional neural network structure and training a network model vehicle attribute identification, identifying a vehicle model and returning vehicle window position coordinate of the vehicle, designing a vehicle image mask and generating a new vehicle image, extracting a multi-task convolutional neural network characteristic of the new vehicle image, training an SVM classification model, and identifying vehicle color. The vehicle attribute identification method is advantageous in that manual characteristic definition and re-classification by a user are not required; the multi-task convolutional neural network structure can simultaneously receive and process a plurality of tasks; and furthermore based on the multi-task convolutional neural network, structure information of the vehicle in the vehicle image is acquired for realizing an effective vehicle color identification method and improving identification accuracy, thereby supplying accurate basis for intelligent traffic.

The invention discloses a facial feature recognition method and system based on the multi-region characteristic and metric learning. The method comprises the steps that convolution neural network parameters of the corresponding location and scale are obtained through the multi-scale facial area training, and corresponding facial area features are extracted according to the neural network parameters; the above features are filtered to obtain the high dimensional facial features; metric learning is conducted according to the high dimensional facial features, the after defined loss function of the feature expression is obtained through the dimension reduction processing of the features, a network model of the metric learning is obtained through the training of the loss function; the images to be recognized are inputted into the network model, the facial features are dimension reduced using the Euclidean distance in order to be recognized. In the method, multiscale is used to select multiple areas, the convolutional neural networks are trained, and the expression skills of the characteristics are improved. Meanwhile, through the selection of the acquired multi-scale features, the efficiency of expression of characteristics is improved, and the accuracy of face recognition is effectively improved.

The invention relates to the technical field of image super-resolution reconstruction, in particular to an image reconstruction system and method based on a cross residual channel-spatial attention network. The system comprises a shallow feature extraction module, a depth feature extraction module, an up-sampling module and a reconstruction layer. The input of the shallow feature extraction moduleis a low-resolution image, and the shallow feature extraction module is used for extracting shallow features; the depth feature extraction module comprises a frequency division module and a cross residual group, the input of the depth feature extraction module is the output of the shallow feature module, and the depth feature extraction module is used for extracting deep features; the input of the up-sampling module is a deep feature and is used for up-sampling; and the reconstruction layer is used for reconstructing features to obtain a high-resolution image. The reconstruction network provided by the invention has stronger feature expression capability and differentiated learning capability, and can reconstruct a high-quality high-resolution image.

The invention relates to a pest image identification method based on a multi-space convolution neural network, and solves the shortcomings of low image identification rate and poor robustness compared with the prior art. The method comprises the following steps of collecting and pre-processing training images, constructing a multi-scale MS-CNN network model and a multi-core classification model, collecting and pre-processing image to be tested, taking a test sample as input and training the multi-core model and the MS-CNN model, and conducting automatic identification of pest images after the training. According to the pest image identification method based on the multi-space convolution neural network, the accuracy rate of pest identification is improved, the robustness of a pest identification algorithm is enhanced and the practical application level is achieved.

The invention discloses a deep learning high under-sampling hyperpolarized gas lung MRI reconstruction method. The method comprises the following steps: constructing and constructing a hyperpolarizedgas lung MRI image training set; according to the method, the cascade CNN model is used, lung contour information is added into a loss function, an accurate reconstructed image can be obtained under the high under-sampling multiple, and the imaging speed is remarkably increased.

The invention discloses a hexagonal bolt looseness detection method based on deep learning and Hough transform, and belongs to the technical field of computer vision. According to the method, a deep learning technology is used, and detection of bolt looseness is realized through an area-based full convolutional neural network target detection algorithm. The depth characteristics of an input picture is calculated to obtain a bolt position, a bolt image is extracted from an original picture according to the calculated bolt position, edge information is obtained through a series of image processing algorithms, and a bolt angle is calculated through Hough transform. By adopting the designed detection scheme, whether the bolt is loosened or not can be accurately detected.

The invention discloses a casting surface defect identification method based on a deep convolutional neural network. The method comprises the following steps: 1, collecting a casting surface defect image, marking the image, and establishing a data set of common casting surface defects; 2, constructing a deep convolutional neural network defect recognition model; 3, constructing a network loss function; 4, dividing the data set into a training set and a test set, and training the defect recognition network by using the training set; 5, inputting the test image into the trained network to identify the position, the type and the size of the defect. According to the invention, the recognition precision and recognition performance of the casting surface defects are improved, and the online, intelligent and automatic development of casting quality detection is promoted.

The invention discloses a remote sensing image unsupervised change detection method based on a Siamese network structure. The remote sensing image unsupervised change detection method comprises the following steps: (1) initializing parameters; (2) obtaining a difference image; (3) performing pixel-level fusion on the two information complementary difference images in the step (2) by using an adaptive local energy weighting algorithm to obtain a new difference image; (4) adopting a clustering algorithm to realize pre-classification; (5) taking a pre-classification result as a label, and realizing the precise detection of an SAR image change region through a DFF-Siamese network; according to the method, unsupervised change detection of the SAR image is realized, priori knowledge is introduced into the deep convolutional neural network, feature mining is deeper by adding a layer-by-layer difference measurement module in the Siamese network, the learning ability of the network is effectively improved, and a more ideal change detection result can be obtained.

The invention relates to a multi-language scene character recognition method and recognition system. The method comprises the steps that the language type of characters in a scene character image is determined; a deep convolutional neural network model is determined according to the language type of the characters; the convolutional layer characteristics of the scene character image are extracted by using the deep convolutional neural network model; a space pyramid model is established based on the convolutional layer characteristics; high-order coding is performed on each space area on the space pyramid model by using a Gaussian model; the results of high-order coding are spliced to act as scene character descriptors; and the scene character descriptors are classified by using a classifier so as to realize multi-language scene character recognition. The multi-language scene character recognition method and recognition system have great recognition effect on the multi-language scene character image so that the method is a general character recognition method and has great adaptability for multi-language scene character recognition.

The invention relates to a multimedia resource searching method and device, computer equipment and a storage medium, and belongs to the technical field of computers. The multimedia resource searchingmethod comprises the steps of receiving a search request, respectively fusing text features of keywords carried by the search request with multimedia features of a plurality of multimedia resources; obtaining a plurality of fusion features with stronger feature expression capability; inputting the plurality of fusion features into a click rate prediction model; carrying out convolution processingon the plurality of fusion features through the click rate prediction model; and outputting the estimated click rates of the multiple multimedia resources, and generating a search result on the basisof the estimated click rates of the multiple multimedia resources, wherein the estimated click rates of the multimedia resources in the search result meet the target condition, and the accuracy of theserver during multimedia resource search is improved, and the search experience of a user is improved.

The invention discloses a road driving area efficient segmentation method based on a depth feature compression convolutional network. The method aims to solve the problem that most current road segmentation methods based on deep learning are difficult to meet accuracy and real-time requirements at the same time. The method comprises: establishing a deep feature compression convolutional neural network; firstly, designing a standard convolutional layer and a pooling layer to perform preliminary compression on extracted road characteristics; by means of advantage of the expanded convolution layer that a receptive field can be increased, and optimizing the advantage, to make up road spatial position information loss caused by feature initial compression, then fusing and decomposing a convolutional layer to realize deep feature compression, finally proposing a layer-by-layer hierarchical up-sampling strategy with learnable parameters to decouple the deeply compressed features, then training the network, and inputting the road image to obtain a segmentation result. The depth feature compression convolutional neural network designed by the invention obtains a good balance between accuracy and real-time performance, and realizes efficient segmentation of a road driving area.

The invention provides a method for detecting a specific kind objective in a movement scene in real time. The method includes the steps that single-frame significance detection is conducted on an obtained video frame sequence, and a significance area which most probably comprises a suspected objective is obtained; an offline training deep learning specific objective classifier is used for conducting target classification judgment on various significance areas, and the property of each significance area is determined; after a concerned specific kind objective is found, a current frame significance detection result serves as the start, and tracking and recording of the subsequent movement track of the objective are achieved. According to the method for detecting the specific kind objective in the movement scene in real time, on the condition that a camera bearing platform moves, the significance areas with few suspected objectives are rapidly determined based on a single-frame image, the calculated amount of full figure searching is reduced, and the algorithm meets the condition of real-time calculation. An adopted deep reliability network has multiple implied layers, has the more excellent feature expression capability than a superficial network and still has the superior classification performance on target images with greatly-changed illumination and appearances.

The invention is suitable for high spectral remote sensing image classification, and provides a high spectral remote sensing image classification method based on three-dimensional Gabor feature selection. The method comprises the following steps: A: according to a set frequency and a direction parameter value, generating a three-dimensional Gabor filter; B: carrying out convolution operation on the high spectral remote sensing image and the three-dimensional Gabor filter to obtain three-dimensional Gabor features; C: selecting a plurality of three-dimensional Gabor features which meet the requirements of each class of classification contribution degrees from the three-dimensional Gabor features; and D: using the selected three-dimensional Gabor features to classify the high spectral remote sensing images through a multi-task sparse classification method. The method is based on the three-dimensional Gabor features, wherein the adopted three-dimensional Gabor features comprise local change information with rich signals, and therefore, feature expression capability is high; the three-dimensional Gabor features are selected through a Fisher discriminant criterion, hidden high-level semantics among features can be fully utilized, redundant information is removed, and classification time complexity is lowered; and further, sparse coding is used to combine the three-dimensional Gabor features with multiple tasks to greatly improve classification precision.

The invention discloses a method of multi-attribute inference based on user node embedding. The method builds a user-commodity dual directed graph G with side weights and performs biased random walk thereon to obtain user-commodity sequences. The user-commodity sequences are placed in in a CBOW model for training to get the real value vector representation of all users in low-dimensional space. Amulti-attribute inference neural network model is constructed, and a multi-attribute inference model is obtained by training the low-dimensional vector representation of the user and corresponding multi-attribute representation as a training set. The real value vector representation of the user who needs to infer the user attribute in the low dimensional space is inputted into the trained multi-attribute inference model, and the multi-attribute values of the user are obtained. The invention can be applied to the fields closely related to user attributes such as defining different customer types in market analysis and mining user attribute information in depth to optimize personalized recommendation algorithm.

The invention discloses a surveillance camera occlusion detection method based on video classification. The surveillance camera occlusion detection method comprises three parts of data preparation, data preprocessing and occlusion detection. In the data preparation stage, monitoring video data is manually labeled, and occlusion videos and normal videos are subjected to dichotomy. In a data preprocessing stage, the video data set is converted into an image data set, and the video frame image is enhanced by using adaptive histogram equalization with limited contrast. In the occlusion detection stage, a video classification method based on deep learning is utilized, video apparent characteristics and time sequence characteristics are comprehensively considered, whether occlusion abnormity occurs in the camera or not is intelligently judged, and high accuracy is achieved. The method is reliable and high in detection accuracy, reduces the requirements of hardware equipment, and can be applied to various bad weather factors.

The invention provides a rapid multi-scale estimation target tracking algorithm related to depth characteristics and re-detection. The characteristics of the target are represented through a deep learning method, and the characteristic expression capability of the target is improved. In the tracking stage, when characteristics of image blocks with different scales are extracted, through PCA dimension reduction, the calculated amount can be reduced, and the overall calculation speed is increased. On the basis of two discrimination indexes, namely a peak sidelobe ratio (PSR) and a confidence coefficient smooth constraint (SCCM), a new detection index is provided, so that the tracking reliability of the current frame can be more accurately measured. If the reliability of the current frame isrelatively low, a series of target candidate boxes are generated through an Edgeboxes method so as to carry out re-detection.

The invention discloses a copper plate surface defect detection and automatic classification method based on machine vision and deep learning. The method comprises the steps: 1, conveying a copper plate to a sensor fixing position through a conveying device; 2, controlling the conveying device to stop moving by a sensor, and triggering an industrial camera to collect images; 3, preprocessing the collected images; 4, inputting the preprocessed defect images into a pre-trained defect detection model to carry out intelligent identification on the surface of a copper part; 5, judging whether the surface of the copper plate has defects or not by the defect detection model; and 6, driving a mechanical arm to grab the defective copper plate into a corresponding defective product groove by a PC. The system comprises the industrial camera, a light source, the sensor, the conveying device, the defective product groove, the PC and the mechanical arm. The problems of low manual detection efficiency, low accuracy, high omission ratio and the like can be solved, meanwhile, the mechanical arm is controlled to autonomously complete the sorting task, and the method has the characteristics of high robustness and high automation level.

The invention discloses a face recognition method based on weighted collaborative representation, and the method comprises the steps: to-be-recognized images are linearly represented as a linear combination of all training images, and distance information of a to-be-recognized image and each type of sample serves as prior information to be introduced into a feature representation function; The reconstruction weights of a certain type of samples closer to the to-be-recognized images are enhanced, then the least square method is utilized to solve the representation coefficient, and finally the type of each to-be-recognized image is judged according to the reconstruction residual error between each to-be-recognized image and each type of training image. The optimization problem is solved based on an L2 norm, so that calculation speed is relatively high, in addition, the category information of training samples and the priori distance information between each to-be-recognized sample and each type of training samples are used as weights for constructing a feature representation equation, so that the feature expression capability of the proposed model can be enhanced to a certain extent;therefore, the influence of changes of image illumination, face postures, expressions and the like on the recognition effect can be effectively avoided.

The invention discloses a medical image classification method and device based on multi-view learning and a deep supervision auto-encoder, and the method comprises the following steps: 1, carrying outthe wavelet decomposition of a region of interest of a medical image, and obtaining a multi-frequency sub-band; step 2, defining each sub-band as a view, and quantitatively extracting an image omicsfeature from each view so as to obtain a multi-view feature; step 3, constructing a classification network of a deep supervision auto-encoder based on multi-view feature learning, and training the classification network based on morphological multi-view feature vectors of the image samples and classification tags thereof to obtain a trained classification model; and step 4, classifying the imageswith unknown classification labels based on the trained classification model. According to the invention, the classification accuracy of medical images can be improved.

The invention relates to an SAR image ship target detection method and system based on training from scratch According to the SAR image ship target detection method based on training from scratch, a ship target detection model comprising a backbone network SAR-SDB and a front-end network SAR-SDF is designed for overcoming the defects of a deep learning detection algorithm used for SAR images so as to be used for achieving accurate detection of a target. The backbone network SAR-SDB has strong feature expression ability, reduces the number of channels, reduces the model size and calculation amount, and avoids the problem of overfitting. By adopting the front-end network SAR-SDF, the accuracy of target classification and positioning can be improved, so that the size of a target detection model is reduced and the target detection time is shortened while the detection accuracy is improved.

The invention discloses a knife switch opening and closing state identification method and device based on multistage image information, and the method comprises the steps: deploying an image collection device in a rectangular region below a knife switch arm, and collecting a knife switch image by aligning to a joint point of a knife switch contact, so as to enable the joint point to be located at the central position of the image; cutting the knife switch image to obtain an image Im and an image Im +, wherein the image Im is a rectangular frame area tightly covering a knife switch arm area, and the image Im + is a rectangular frame area tightly covering a knife switch contact area; inputting the images Im and Im + into a pre-constructed and pre-trained knife switch opening and closing state identification network model, obtaining the probability of each category of the knife switch opening and closing state, and selecting the state corresponding to the maximum value of the category probability as the identification result of the knife switch state. According to the method, the multi-level image information and the deep neural network are adopted, so that the feature expression capability is enhanced, and the robustness and recognition performance of the method are improved.

The invention belongs to the related technical field of milling finish machining and deep learning, and discloses a tool feature point recognition method and equipment combining transverse geometric features of adjacent tool paths, and the method comprises the following steps: (1) analyzing a G01 program segment of a target part to obtain three-dimensional coordinates of a tool location point in a machining tool path, sorting according to the advancing direction of the cutter to obtain a cutter location point cloud; (2) determining and calculating geometric parameters of the cutter location points, and constructing geometric feature vectors of the cutter location points; (3) generating a geometric feature matrix of the cutter location points by combining the neighborhood cutter location points in the advancing direction of the cutter; (4) topology is carried out on the cutter location point cloud to form a graph data structure; (5) establishing a communication relation between the cutter location points through the adjacent cutter location point index of each cutter location point, and calculating a cutter location point cloud adjacent matrix; and (6) inputting the cutter location point cloud data of the predicted feature points and the cutter location point cloud adjacency matrix into the trained graph neural network model to complete the recognition of the cutter feature points. The method has higher identification precision and recall ratio.

The invention discloses a text stance detection method based on text and user representation learning, which acquires a text data set from a social media platform, generates a user social relationship graph and obtains a corresponding Laplacian matrix, and determines the stance of each text Label vector, use the pre-trained BERT model to obtain the text vector of each text, build and train the position detection model, when it is necessary to detect the position of the user text, generate the user social relationship graph and obtain the corresponding Laplacian matrix, the The text vector is input into the position detection model to obtain the position detection result. The present invention respectively acquires two modal features of text and user and performs cross-modal fusion, so as to realize accurate text position detection.

The invention discloses a self-adaptive asymmetric quantization deep neural network model compression method, which comprises the following steps of: adaptively quantizing the weight of each layer offloating point of a network into an asymmetric ternary or quaternary value in the training process of each batch before forward propagation starts to calculate during deep neural network training; ina back propagation parameter updating stage, carrying out parameter updating by using the original floating point type network weight; and finally, performing compression storage on the trained quantized deep neural network. According to the method, the parameter redundancy degree of the deep neural network is reduced, the residual parameters are adaptively quantified, the network model is compressed to the greatest extent, and the identification accuracy of the quantization method on the deep network and a big data set is improved.

The present invention proposes a fast multi-scale estimation target tracking algorithm on deep features and re-detection. The characteristics of the target are represented by the method of deep learning, which improves the feature expression ability of the target. In the tracking stage, when extracting features of image blocks of different scales, PCA dimensionality reduction can reduce the amount of calculation and improve the overall calculation speed. Based on two discriminant indicators, peak side lobe ratio (PSR) and confidence smoothing constraint (SCCM), a new detection index is proposed, so that the tracking reliability of the current frame can be measured more accurately. If the reliability of the current frame is low, a series of target candidate boxes are generated by the method of Edgeboxes for re-detection.

The invention discloses a high-resolution remote sensing image road extraction method based on combination of deep learning and a multi-dimensional attention mechanism. The method comprises the following steps: extracting remote sensing image road information by adopting a full convolutional neural network UNet; a multi-dimensional attention module is combined with a coding part of the UNet network, so that a road feature map transmitted to a decoding part has higher feature expression capability; a multi-level feature fusion mode is adopted, feature information of different levels is obtained in each layer in the decoding stage, and a transmitted feature map has texture information and semantic information so as to optimize the expression ability of the feature map; a user can observe an extraction result of a high-resolution remote sensing image returned by a satellite in real time by accessing a Web front end of node.js based on a server. According to the scheme, high-accuracy remote sensing image road information is extracted, the image subjected to convolution training has higher expression ability due to introduction of the multi-dimensional attention module and the multi-level feature fusion method, and compared with a general deep learning method, the remote sensing image road extraction accuracy is improved. Meanwhile, the self-feedback mechanism of the deep learning network enables the extraction process to be more intelligent and automatic, and adaptive adjustment can be performed on images of different road scales in different regions to obtain optimal road image information, so that the method has very high practical value and popularization value.

The invention discloses a face feature recognition method and system based on multi-region feature and metric learning. The method includes: obtaining convolutional neural network parameters of corresponding positions and scales through multi-scale face region training, and according to the convolutional neural network The parameters extract the features of the corresponding area of ​​the face; filter the above features to obtain high-dimensional face features; perform metric learning according to the high-dimensional face features, perform dimensionality reduction processing on the features to obtain feature expressions, and then define a loss function. The loss function training is used to obtain a network model of metric learning; after the image to be recognized is input into the network model, the face features are reduced in dimension and then recognized by Euclidean distance. In the present invention, multi-scale selection of multi-regions is used to train the convolutional neural network, which improves the expressive ability of features. At the same time, by selecting the acquired multi-scale features, the expression efficiency of the features is improved, and the accuracy of face recognition is effectively improved.

The invention relates to artificial intelligence, and discloses a bus lane detection method based on image recognition, the method comprises the following steps: acquiring an original input image of a lane; constructing a feature extraction network, and extracting image features of the original input image, wherein the image features output by the feature extraction network execute operations and convolution operations of a CBL module for multiple times to obtain a feature map of one scale, wherein different intermediate layers of the feature extraction network respectively execute multiple times of CBL module operation, convolution, up-sampling and feature fusion operation to obtain at least three feature maps with different scales; monitoring and identifying the bus lane on the feature maps of the at least four scales by adopting an anchor frame method; and mapping the corresponding bus lane coordinates on the feature map into the coordinates on the original input image, thereby realizing bus lane detection of the original input image. The invention further provides a device, electronic equipment and a computer readable storage medium. According to the invention, the accuracy of bus lane recognition and the recall rate in a difficult scene are improved.