134results about How to "Enhanced Feature Representation" patented technology

The invention relates to a vehicle type recognition method based on a CNN and domain adaptive learning. The method comprises steps of: establishing a CNN-based initial model by adding a rotation-invariant layer in an Alexnet network, distinguishing a discriminant layer and designing a new objective function; separately extracting the feature maps of different-domain sample convolution layers by using the established initial model, calculating the cosine similarity between the sample feature maps, determining the shared convolution kernel or the non-shared convolution kernel of the CNN, retaining the weight and the offset of the shared convolution kernel, and updating the weight and the offset of the non-shared convolution kernel; based on a target-domain training sample, calculating the cosine similarity between respective feature map layers and the average similarity of the entire target domain, and clustering each type of similar feature maps according to the average similarity; expanding source-domain samples with similar distribution characteristics in the target domain to new samples in the target domain, adjusting the entire CNN model by using the new samples in the target domain, and then using a softmax classifier to classify the vehicle types of the test samples in the target domain.

The invention discloses a spatial spectrum attention hyperspectral image classification method based on Octave convolution, and solves the problems of large class spacing, small different class spacing and low classification accuracy in the prior art. The scheme is as follows: inputting images to be classified and preprocessing data, dividing a training set and a test set, constructing an Octave convolutional neural network, determining a loss function of the Octave convolutional neural network, training and updating the Octave convolutional neural network, testing the data of the test set, and completing hyperspectral image classification. According to the method, Octave convolution operation is used to reinforce feature representation, and a spatial attention mechanism and a spectral attention mechanism are introduced, so that the network can more accurately find an area which is more beneficial to classification and contains more comprehensive and detailed information. The method ishigh in classification precision and strong in robustness, and can be applied to analysis and management of hyperspectral image data.

The invention relates to an intelligent railway roadbed disease recognition method based on a ground penetrating radar and belongs to the technical field of railway roadbed detection. The method includes the following steps of: establishing an intelligent railway roadbed disease recognition software system including ground penetrating radar data acquisition, preprocessing, two-dimensional discrete detection, feature extraction, feature dimension reduction and recognition model creation; and using the established intelligent recognition software to recognize a railway roadbed disease. According to the method, machine vision and pattern recognition technology replace manual reading of ground penetrating radar data so ad to achieve rapid, accurate and lossless intelligent recognition of a plurality of railway roadbed diseases, thereby improving the timeliness of the detection of the ground penetrating radar and promoting the intelligence of railway roadbed detection.

The invention discloses a semantic segmentation method based on feature pyramid attention and mixed attention cascading. The method comprises the steps of building a semantic segmentation training set; constructing a deep convolutional neural network, wherein the deep convolutional neural network comprises an encoder part, two feature pyramid attention modules, a mixed attention module, a decoding branch, a feature fusion part and a deep separable convolutional layer; training the deep convolutional neural network by using the semantic segmentation training set, and correcting network parameters; and inputting the streetscape road scene image to be segmented into the trained deep convolutional neural network to obtain a segmentation result. The invention can better adapt to the requirements of unmanned vehicle equipment for precision and speed.

The invention relates to a hybrid recommendation method based on sparse edge noise reduction automatic coding, belonging to the personalized recommendation field. The invention comprises the followingsteps: firstly, merging each commodity comment into a comment document; using TF-IDF processes the review text for each product and selects the word with the highest TF-IDF value to construct the commodity content vector as the feature representation of the commodity. Secondly, the sparse edge denoising automatic coding model is trained with commodity content vector, the commodity features are extracted from the commodity content vector with the trained model, and the similarity of the commodity feature vector is calculated with cosine similarity, so as to obtain the influence of the nearestneighbor commodity. Finally a predication score is obtained through combination of neighbor commodity influence and user-evaluation matrix deposition. The invention effectively improves the algorithmefficiency and the recommendation accuracy.

The invention discloses a music audio classification method based on a convolutional recurrent neural network. The method comprises the following steps: S1, annotating music audios to obtain a music annotation data set; S2, enhancing the training data of the data set by adopting a music data enhancement method; S3, framing and windowing the audio signals of the music in the data set, and obtaininga Mel sound spectrum corresponding to the audio through short-time Fourier transform and Mel scale transform; S4, constructing a music audio classification model based on a convolutional recurrent neural network; S5, inputting the Mel sound spectrum of the training data into the music audio classification model based on a convolutional recurrent neural network for iterative training; and S6, inputting a Mel sound spectrum corresponding to the music, and predicting the label of the music. The method provided by the invention can improve the ability of the network to extract the sound spectrumfeatures, and obtain better music overall feature representation, thereby improving the accuracy of music audio classification.

The invention discloses a long-tail distribution image data identification method based on dual-channel learning. The method comprises the following steps: 1) constructing a dual-channel learning model combining unbalanced learning and small sample learning; 2) updating all parameters in the dual-channel learning model by utilizing dual-channel learning total loss and back propagation, and storingoptimal dual-channel learning model parameters; and 3) inputting the image data of the test set to the optimal dual-channel learning model, and obtaining the prediction label of the image. Accordingto the invention, unbalanced learning and small sample learning are combined to solve the problem of long-tail distribution image data identification; the unbalanced learning channel can improve the identification accuracy of the unbalanced data set; the small sample learning channel can improve the feature representation of model learning, and the dual-channel total loss enables the model to focus on the unbalanced learning channel in the early stage of training and focus on the small sample learning channel in the later stage of training, thereby improving the recognition accuracy of the long-tail distribution image data on the whole.

The invention belongs to the technical field of network security, and particularly relates to an Android malicious software detection method and device based on a capsule network, and the method comprises the steps: collecting an Android software file sample, decompressing a to-be-processed file, converting the to-be-processed file into an RGB three-channel color image, and enabling the RGB three-channel color image to serve as training sample data; constructing a capsule network, and training the capsule network by using the sample data to obtain a trained network model containing a graph structure and network parameters, the capsule network realizing transmission between feature vectors in a capsule layer through an iterative dynamic routing algorithm; and inputting the to-be-detected target file into the trained capsule network model for testing, and judging whether the to-be-detected target file is a malicious software file or not through an output result. The method and device canefficiently run on the Android operation platform, occupy few resources, are high in efficiency and accuracy, can realize high-accuracy classification detection tasks even under the condition of small-scale training samples, and achieve the purpose of protecting the Android intelligent mobile terminal.

The invention discloses a super-resolution reconstruction method and system for an image. The method is based on a channel attention mechanism and a position attention mechanism of an image. Featureson each image block are extracted at different depths and then fused, and the importance of the characteristic channel information and the position information is effectively identified; and since thefeatures of different depths make different contributions to the image super-resolution, the method and the system make full use of and fuses the features of different depths, the obtained image features retain a larger receptive field and better detail features at the same time, and the super-resolution effect of the obtained image is better after reconstruction based on the features. Besides, the relation between the features is fully mined so that the performance bottleneck problem caused by the too deep network can be alleviated, the feature representation capability of the image is great, and the robustness can be realized.

The invention discloses a deformable context coding network model and a liver and liver tumor segmentation method, which can accurately determine the contour positions of a liver and a liver tumor, realize more accurate liver and liver tumor segmentation, have a wide application prospect, and are suitable for popularization and application. According to the network model of the invention, deformable convolution is utilized to enhance the feature representation capability of a traditional encoder, help the traditional encoder to learn a convolution kernel with adaptive spatial structure information, and eliminate interference of different sizes of liver tumor positions; according to the method, the global feature information in the image is coded by using the Ladder spatial pyramid poolingmodule for extracting the multi-scale context information, so that the contour positions of the liver and the liver tumor are determined more accurately, more accurate liver and liver tumor segmentation is realized, and the method has a wide application prospect.

The invention discloses a video expression recognition method based on a deep residual attention network. The method comprises the following steps: S1, performing video data preprocessing on a video sample; S2, performing facial expression feature extraction on the face image by adopting a deep residual attention network; and S3, carrying out certain processing on the features extracted in the step S2, then carrying out training and testing, and outputting a final classification result of the facial expressions. The method is realized by adopting a spatial attention mechanism, weights in spatial distribution are generated for an input feature map, and thenweighted summation are performed with the feature map, so network learning is supervised to allocate different attention (weights) to different areas closely related to expressions in the face image, feature learning of a target area closely related to expressions in the face image can be focused, the feature representation capabilityof the deep residual network is improved, and the performance of video expression recognition is further improved.

The invention discloses a remote sensing image target detection method based on multi-scale feature fusion. The method comprises the following steps: constructing a target detection basic network composed of a truncation type VGG and a multi-scale feature map; constructing a multi-scale feature map fusion module composed of a deconvolution module and a prediction module, the multi-scale feature map fusion module being used for extracting context information of a multi-scale feature map in the target detection basic network and generating a multi-scale feature map; predefining a multi-length-to-width ratio and a multi-scale default frame according to the size of the receptive field of the multi-scale feature map, and completing positioning and classification of the target by using the default frame; training a target detection basic network; accessing a multi-scale feature map fusion module to the target detection basic network, fixing parameters of the target detection basic network during training, and only adjusting the parameters of the multi-scale feature map fusion module; and at the last stage, finely adjusting parameters of the target detection basic network and the multi-scale feature map fusion module at the same time to achieve a coupling effect.

The invention provides a knowledge graph completion-oriented link prediction method, which comprises the following steps of: sequentially regarding other entities irrelevant to a head entity as tail entities by utilizing a given head entity and a given relationship, and sequentially regarding other entities irrelevant to the tail entities as head entities by utilizing a given tail entity and a given relationship, calculating scores of the triple through a link prediction model, taking the tail entity with the highest score as a predicted tail entity, and taking the head entity with the highest score as a predicted head entity; according to the link prediction model, a global context coding module based on an attention mechanism is introduced into a convolutional neural network, global context information is learned by aggregating local features, and feature representation used for knowledge graph completion is enhanced. The method can solve the technical problem that when a convolutional neural network model carries out knowledge graph embedding learning, due to the fact that context information is not utilized, triads with many complex relations have wrong link prediction.

The invention discloses a traffic signal control method and system based on a random strategy gradient, and electronic equipment. The method comprises steps of obtaining the static road network data of at least one control signal intersection; visually drawing a traffic simulation road network according to the static road network data; acquiring real-time traffic operation state data of at least one control signalized intersection; performing parameter check on simulation parameters in the traffic simulation road network according to the traffic operation state data to obtain an optimized traffic simulation road network; inputting a traffic state obtained by observing the optimized traffic simulation road network into a value network to obtain an evaluation value of each signal control scheme in the traffic state; inputting the traffic state into a strategy network to obtain a probability value of each signal control scheme; and updating parameters of the strategy network through a random strategy gradient based on the evaluation value of each signal control scheme in the traffic state and one signal control scheme. According to the method provided by the invention, a problem of dimension explosion of signal control can be solved.

The invention discloses a medical image super-resolution reconstruction method based on a dense mixed attention network. According to the method, a hybrid attention mechanism is introduced on the basis of the dense neural network, so that the neural network pays more attention to channels and regions containing rich high-frequency information, network convergence is accelerated, and the super-resolution precision is further improved. The method mainly comprises the following steps: designing and constructing a network based on a dense neural network and a hybrid attention mechanism; preprocessing the data set, enhancing the data, and constructing a training sample; training the network model by using L2 loss until the network model is converged; and in a super-resolution reconstruction stage, inputting a low-resolution medical image, and performing super-resolution reconstruction by using the trained network model to obtain a final high-resolution image. Compared with a mainstream super-resolution method, the method provided by the invention is higher in precision, and is an effective medical image super-resolution reconstruction method.

The invention discloses a video pedestrian re-identification method based on a multi-attention heterogeneous network, and belongs to the field of image processing. The method comprises the following steps: constructing and training a multi-attention heterogeneous network; and performing feature extraction on the video of the known pedestrian ID and the video of the undetermined pedestrian ID by using the trained network, and the pedestrian ID is judged according to the cosine distance between the two features. According to the method, Soft attention and non-local attention are introduced intothe OSNet network; according to the method, the Soft attention is utilized to pay attention to pedestrian area features in an image, the learning ability of non-local attention to space-time featuresin a video sequence is utilized to improve feature representation of the video sequence, features which are more robust and more discriminative are extracted, and the recognition accuracy is improved.And meanwhile, the features of the specific frame are selected as local feature learning network branches, so that the learning of pedestrian local features is enhanced while the pedestrian global features in the video sequence are learned, and the performance of the network in video pedestrian re-identification is improved.

The invention relates to the technical field of natural language processing and deep learning, in particular to a method and system for processing repeated entities in remote supervision relation extraction. According to the main technical scheme, the method comprises the steps that corpora form a sentence packet according to entity pairs, and the position of a target entity in a statement containing repeated entities is determined; a word vector containing semantic and position information is construncted; sentence vectors are constructed through a multi-angle convolutional neural network; and through a dynamic routing mechanism, a sentence packet level vector is constructed and sentence packets are classified. According to the relation extraction method and system, the problem of attention deviation caused by repeated entities which are not mentioned in an existing remote supervision relation extraction method and system can be positioned and effectively solved.

The invention provides an MRI tumor optimal segmentation method and system based on multi-modal image fusion. The MRI tumor optimal segmentation method comprises the steps that step 1, creating an MRItumor multi-modal image fusion network step 2, constructing a multi-modal 3D network for enhancing tumor image segmentation; step 3, constructing a significance loss function based on GAN image fusion; step 4, constructing a Mask attention mechanism contrast loss function; step 5, constructing an SSIM loss function; and step 6, performing MRI tumor optimal segmentation according to the MRI tumormulti-modal image fusion network, the multi-modal 3D network and the three loss functions. When a deep architecture is trained, one remaining unit helps. Feature accumulation is carried out by using arecursive residual convolution layer, so that better feature representation is provided for a segmentation task; a U-NET system structure with the same network parameters and good performance is designed for medical image segmentation.

The invention discloses a group behavior recognition method based on multi-modal information fusion and decision optimization, and the method comprises the steps: firstly obtaining a group member candidate box sequence for a to-be-recognized video, extracting the corresponding optical flow features, and extracting the human body posture segmentation features as a third visual clue; then acquiringa double-flow model of the spatial and temporal features of the human body target and performing multi-modal information fusion (MMF) on the double-flow model; and finally, connecting the two branchesobtained after MMF fusion with a GRU, and performing decision optimization by adopting a multi-classifier fusion method based on adaptive category weight, thereby obtaining a group behavior label. According to the scheme of the invention, during feature fusion, an MMF feature fusion algorithm is designed, so that space-time features supplement each other, information supplements each other, and finally better feature representation is obtained; in the aspect of decision optimization, a multi-classification fusion method based on self-adaptive class weights is designed, classifier acceptance and rejection and each class weight are calculated more accurately, and therefore high recognition precision is obtained.

The invention relates to a hyperspectral image classification method, and concretely relates to a hyperspectral image classification method based on spectral space attention fusion and a deformable convolutional residual network. The objective of the invention is to solve the problem of low classification accuracy of hyperspectral images caused by insufficient spectrum and spatial feature extraction and overfitting under small samples due to the fact that the hyperspectral images contain abundant information in the existing hyperspectral image classification. The method comprises the following steps: 1, acquiring a hyperspectral image data set and a corresponding label vector data set; 2, establishing an SSAF-DCR network based on spectrum space attention fusion and a deformable convolution residual error; 3, inputting the x1, the x2, the Y1 and the Y2 into a network SSAF-DCR, and performing iterative optimization by adopting an Adam algorithm to obtain an optimal network; and 4, inputting x3 into the optimal network to carry out classification result prediction. The method is applied to the field of hyperspectral image classification.

The invention discloses a method for text classification using diverse text features. The method comprises the steps of generating a plurality of sets of different text feature representations by using a multi-dimensional text representation algorithm, and generating a multi-dimensional text feature representation longitudinally; using a plurality of different text representation algorithms to generate multiple sets of different text feature representations, and generating a multi-dimensional text feature representation transversely; combining different feature representation vectors of each sample as a new feature vector of the sample, thereby obtaining a new feature representation of a data set. The method improves existing text representation algorithms, proposes to use more text representations with lower dimensions and larger differences to mine different internal structures of texts, enhances a text representation ability, improves the effect of tasks such as text categorizationand the like, and greatly reduces the dimension of the text features.

The invention discloses a cloth defect detection method and system based on a deep neural network, and belongs to the technical field of pattern recognition. The method comprises the steps that a defect cloth image training set is used for training a deep neural network model, labels are defect types and real frame position information, the deep neural network model is composed of a backbone network and a detection network, and the backbone network is used for extracting three feature maps with different scales from defect cloth images; the detection network includesthree detection sub-networks and the detection result fusion module, wherein the three detection sub-networks are the same in structure. Each detection sub-network is used for detecting a defect type and prediction frame position information from the feature map, and consists of three dense connecting blocks, and the feature channel connection between the dense blocks is used for enhancing feature transfer, and the detection result fusion module is used for performing non-maximum suppression on the prediction result to obtain a final prediction frame and a defect type, and inputting to-be-detected cloth into the traineddeep neural network model to obtain a detection result, so that the type and the position of the defect in the cloth can be detected more quickly and accurately.

The invention relates to the field of face age estimation, in particular to a multi-scale face age estimation method and system embedded with high-order information, and the multi-scale face age estimation method comprises the steps: inputting a face image, and carrying out the preprocessing of the face image; inputting the face image into a residual network to perform global feature extraction soas to construct a global branch; inserting blocks for extracting high-order age information at different positions of the global branch; taking the output feature map of the first convolution layer of ResNets as the input of a long-term and short-term memory network, obtaining the position information of the age-sensitive region, and obtaining a local feature map through cutting to construct a local branch; minimizing a loss function through back propagation, performing joint optimization on the two branches, and performing iterative training on the neural network; and inputting the test setinto a trained neural network model, and calculating and outputting a final predicted age according to the age characteristics. The multi-scale face age estimation method and system have the advantages that the network model provided by the invention is relatively low in calculation cost and high in precision, and the applicability of related products is relatively high.

The embodiment of the invention provides an image segmentation method, device and apparatus, and a storage medium. The method comprises: extracting an initial feature map of a to-be-processed image through a base network of an image segmentation model; pooling the initial feature map through an average pooling sub-model of the image segmentation model to obtain a first feature map carrying short-distance dependency relationship information; processing the initial feature map through at least one branch sub-model to obtain at least one target feature map, wherein the target feature map comprises a second feature map carrying global dependency relationship information and/or a third feature map carrying long-distance dependency relationship information; and cascading the first feature map with the target feature map, and then performing convolution to obtain an image segmentation result and outputting the image segmentation result. According to the method, the branch sub-models are arranged in parallel with the average pooling sub-model to obtain the feature map carrying global dependency relationship information and/or long-distance dependency relationship information, and the feature map carrying global dependency relationship information and/or long-distance dependency relationship information is cascaded with the feature map carrying short-distance dependency relationship information obtained by the average pooling sub-model, so that the feature representation capability is enhanced, and the image segmentation accuracy is improved.

The invention discloses a space-time channel constraint correlation filtering tracking method based on suppressible abnormity. The method comprises the steps of S1, extracting HOG features, first depth features and second depth features of a t-th frame; S2, performing fusion processing on the HOG feature, the first depth feature and the second depth feature to obtain a first fusion feature X, anddetermining the position and the scale of a target in the t-th frame of image based on the first fusion feature X and a filter; S3, according to the feature map of the t-th frame, updating the filterbased on a space-time channel constraint related filtering model capable of suppressing abnormality; and S4, repeating the step S2S4 until all frames are tracked, and finally obtaining a tracking result. According to the method, the feature representation capability of the target template is remarkably improved by combining manual features with depth features, self-adaptive channel feature selection is realized through l2, 1 norm, and the problems of boundary effect and background clutter are effectively solved.