183results about How to "Enhance feature expression" patented technology

The invention relates to a medical image segmentation method and system based on a generative adversarial network, and electronic equipment. The method comprises the following steps: firstly, researching how to extract pixel-level features of different types of high-quality images by a generator, and carrying out structured feature representation by utilizing a capsule model so as to generate a pixel-level labeled sample; secondly, constructing a proper discriminator for discriminating the authenticity of a generated pixel-level labeled sample, and designing a proper error optimization function; feeding back a discriminating result to models of a generator and a discriminator respectively, improving the sample generating capacity and the discriminating capacity of the generator and the discriminator respectively through continuous adversarial training, finally, adopting a trained generator to generate a pixel-level labeled sample, and achieving pixel-level segmentation of an image-level labeled medical image. According to the invention, the dependence of the segmentation model on pixel-level annotation data is effectively reduced, the adversarial training efficiency of the generated sample and the real sample can be improved, and high-precision pixel-level image segmentation can be effectively realized.

The invention discloses a film comment-oriented naming entity recognition method, and belongs to the technical field of entity recognition. The method comprises the following steps of: defining a naming entity type of a name in a film comment; constructing a training data set TrainSet; segmenting a non-marked data set by taking characters as units so as to construct a character matrix, and constructing a mapping dictionary of characters and word vectors by utilizing a skip-gram model in word2vec; generating a naming feature dictionary corresponding to each character; combining word vectors andnaming features of the characters to generate a feature vector matrix for training a name recognition model; constructing a mixed model by utilizing Bi-LSTM and CRF so as to train the name recognition model; and carrying out name recognition on the feature vector matrix through the trained name recognition model, so as to realized film comment name recognition task-oriented name definition, carryout name labeling on character levels during name recognition, obtain name features and combine word vectors obtained via word2vec training with the naming features.

The invention discloses an unsupervised cross-domain pedestrian re-identification method and an unsupervised cross-domain pedestrian re-identification system. The method comprises the following steps:constructing a source domain training set and a target domain training set; converting the training images in the source domain training set into a target domain, and generating image data related tothe target domain; training an initial pedestrian re-identification model by utilizing the generated image data; extracting local features of each training image in the target domain training set based on the trained pedestrian re-identification model; performing clustering analysis on training image data in a target domain training set by using the extracted local features; determining an optimal training sample in the target domain training set based on a clustering analysis result; utilizing the generated image data and the determined optimal training sample to retrain the pedestrian re-identification model, and repeating in sequence until an iteration stop condition is reached to obtain a final pedestrian re-identification model; and obtaining to-be-identified image data in the targetdomain, and identifying the to-be-identified image data by using the finally obtained pedestrian re-identification model.

The invention discloses an image semantic segmentation model generation method and device, equipment and a storage medium. The method comprises the steps of obtaining an image sample set; using the image sample set to train an image semantic segmentation model; wherein the image semantic segmentation model comprises a feature map extraction part and a feature map analysis part, the feature map extraction part comprises a plurality of cascaded hole convolution processing residual error modules, and the feature map analysis part is constructed based on an attention mechanism, a pixel correlationmechanism and multi-scale information. According to the technical scheme provided by the embodiment of the invention, the attention mechanism is effectively utilized to learn the dependency relationship in spatial position and channel dimension, the feature expression capability is enhanced, the pixel correlation mechanism is utilized to make the segmentation result more accurate, and meanwhile,the multi-scale feature information is also utilized to learn a global scene to improve the accuracy of pixel point classification.

The invention discloses a remote sensing image super-resolution reconstruction method based on a convolutional neural network of channel attention. According to the method, a channel attention mechanism is introduced into the proposed convolutional neural network, and the characteristic expression capability of the network can be improved by utilizing the mutual dependency relationship among the characteristic channels. The method mainly comprises the following steps: designing and building a convolutional neural network model based on channel attention; Constructing a training sample by usinga smaller remote sensing database, and pre-training the deep convolutional neural network model constructed in the step 1; 2, when model training in the step 2 reaches convergence, a larger remote sensing database is added for fine adjustment, and the performance of the network is further improved; In the remote sensing image reconstruction stage, a low-resolution image is used as input, and themodel trained in the step 3 is used for reconstructing a final high-resolution image. According to the method, a better reconstruction effect can be obtained, and the method is an effective remote sensing image super-resolution reconstruction method.

The invention discloses a remote sensing image classification method based on a deep fusion convolutional neural network, and the method comprises the steps: constructing an original remote sensing image into a data set, carrying out the preprocessing of the original remote sensing image, dividing the preprocessed image into a training set, a test set and a verification set, and carrying out the data augmentation of the training set; constructing a deep fusion convolutional neural network; training to obtain an optimal network model; and classifying the actually measured remote sensing imagesby using the optimal network model. The invention provides a new classification method. A new deep fusion convolutional neural network is constructed; an improved encoder-decoder model is combined with a VGG16 model to obtain a VGG16 model; the model fuses the deep features and the middle-layer features of the remote sensing image, so that the defect of low classification precision caused by single or redundant feature extraction of the remote sensing image in the prior art is effectively overcome, the advanced feature expression capability of the target is obtained by establishing the novel network model, and the classification accuracy of the remote sensing image is improved.

The invention discloses a rotating machine fault diagnosis method based on an optimized structure convolution neural network. Firstly, working signals are collected in the normal state and the fault state of a rotating machine, then are converted into gray scale images, and the gray scale images and the corresponding fault labels are used as training samples to train the constructed convolution neural network; and in the working process of the rotating machine, a working signal is collected and is converted into a gray scale image, and the trained convolution neural network is input for faultdiagnosis. The invention adopts the method of converting the collected working signals of the rotating machine into gray-scale images, and completes the task of multi-classification of faults of the rotating machine through a convolution neural network.

The invention relates to a long-time sheltering robust tracking method based on convolutional features and global search detection. Through adopting a convolutional feature and multi-scale relative filtering method in a tracking module, the feature representation capacity of a tracking object appearance model is enhanced, so that a tracking result is very robust to factors such as illumination variation, object scale variation and object rotation. In addition, through the introduced global search detection mechanism, a detection module can detect the object again when the tracking fails because the object is sheltered for a long time, thus the tracking module can recover from error, and the object can be continuously tracked for a long time even in the case of object appearance changes.

The invention discloses a graph-based direct-push type semi-supervised pedestrian re-identification method, and belongs to the technical field of computer vision pedestrian re-identification. The method comprises the steps of firstly, using labeled pedestrian data for training a double-channel model; after obtaining the base model, carrying out feature extraction on the label-free pedestrian data,establishing a graph model for the extracted label-free pedestrian data features, giving a pseudo label to the label-free pedestrian data according to the graph model, and constructing a positive andnegative sample pair by using the labeled pedestrian data and the label-free pedestrian data with the pseudo label; assigning confidence coefficients to positive and negative sample pairs by using the graph model, and jointly finely adjusting the base model; gradually increasing the difficulty and confidence of positive and negative sample pairs, training the base model to complete convergence byusing a course learning method, performing feature extraction and feature matching on verification set data after a final model is obtained, and completing pedestrian re-identification according to amatching result. According to the method, the negative influence caused by wrong pseudo tags is reduced, the robustness of the model is improved, and the pedestrian re-identification precision is further improved.

The invention provides a new crown diagnosis system based on a deep convolutional neural network and multi-instance learning, and the system comprises a feature extraction module which packages a CT sequence of a patient, carries out time domain convolution, screens infected slice instances through weak supervised learning, and obtains an infected segment; The system comprises a multi-branch network system configured to input the feature sequence extracted from the CT sequence of the patient into a plurality of parallel branch networks, wherein class activation sequences output by different parallel branch networks are different, so that different infection fragments are positioned, and the integrity of specific case features of the patient is modeled, an attention divergence induced by the weak supervised learning is confronted to enhance the accuracy and robustness of the infected fragment; a multi-instance learning module configured to perform feature fusion of time domain convolution on multi-instance bagging so as to enhance specific case feature expression of the patient; and a gating attention mechanism module configured to perform adaptive instance feature weighted fusion to avoid gradient disappearance in multi-instance learning.

The invention discloses an advertisement click rate prediction framework and algorithm based on user behaviours. ID characteristics and other characteristics are co-converted in different levels intomeaningful numerical characteristics; due to the characteristics, the characteristic sparsity and redundancy can be reduced; the characteristic expressiveness can be improved; simultaneously, to further improve the characteristic expressiveness, characteristic selection and characteristic combination are carried out by utilization of a GBDT model in the invention; high-dimensional characteristicsare processed by utilization of an LR model; finally, to solve a class imbalance problem, a down-sampling algorithm based on a K_Means model is provided in the invention; in an experimental process, characteristic extraction on original characteristics is carried out at first; then, characteristic classification is carried out by adoption of heuristic thinking; characteristic combination is carried out by inputting perceptual characteristics into the GBDT model; finally, rational characteristics and combination characteristics are input into the LR model with a certain weight, so that advertisement click rate prediction is carried out; and an experimental result shows that the algorithm in the invention is improved both on RMSE and R2 indexes.

The invention discloses a military target identification method based on deep learning, and belongs to the field of automatic target identification based on images. The military target identificationmethod based on deep learning disclosed by the invention can be used in the fields, such as military target identification, large-scale military target retrieving, weapon equipment intellectualizationand situation estimation. According to the method disclosed by the invention, for the problems that each layer is in one-way connection and the feature expression capability is insufficient in the traditional target identification network based on deep learning, an algorithm model of a dense connected convolutional layer is re-designed; a dense connection mode is used; the algorithm model multiplexes features of each layer; therefore, the average target identification accuracy rate of the algorithm model is increased; the algorithm model obtained by training in this way is relatively small; and simultaneously, the algorithm model solves the problems of gradient diffusion and gradient expansion.

The invention provides a depth pedestrian re-identification method based on positive sample balance constraint. A residual error network employed in the method is simple in structure and can be widely used, and the network structure which is deep sufficiently improves the feature representation capability. Moreover, there is no need to specially design the network structure. A residual error classifier is used for feature extraction of an image, so the accuracy of pedestrian re-identification can be greater than the accuracy of most of well-designed methods. Compared with two-tuple loss and three-tuple loss methods, the method does not need to intentionally generate an effective sample for improving the structural loss and can achieve the similar effect. Moreover, the method enables the learned gradient direction to be more robust and effective through the overall distribution information. On the basis of improving the structural loss, the method improves the positive sample balance constraint, can control the distance of a positive sample pair, also can balance the gradients of the distance of the positive sample pair and the distance of a positive sample pair, enables the algorithm to be easier to train, and improves the performances of an algorithm.

The embodiment of the invention provides an accompaniment and human voice extraction method and device, a word-by-word lyric generation method and device, a storage medium and electronic equipment, and relates to the technical field of audio processing. The accompaniment and human voice extraction method comprises the following steps: acquiring mixed audio data in a time domain; encoding the mixedaudio data through an encoding module in an accompaniment voice extraction model to obtain encoded audio features; performing first convolution processing on the encoded audio features through a first decoding module in the accompaniment human voice extraction model to obtain target human voice audio data; enabling the encoding module and the first decoding module to form a first U-net network; performing second convolution processing on the encoded audio features through a second decoding module in the accompaniment voice extraction model to obtain target accompaniment audio data; and enabling the encoding module and the second decoding module to form a second U-net network. According to the invention, the accompaniment and human voice extraction effect can be improved.

The invention discloses a face super-resolution method and system based on a fusion attention mechanism, and belongs to the field of face image super-resolution. The method comprises the steps of down-sampling a high-resolution face image into a target low-resolution face image, carrying out partitioning operation, separating out image blocks which are overlapped with each other, and extracting superficial features by using a superficial feature extractor; fusing features of pixel, channel and space triple attention modules, and enhancing reconstructed face structure details; constructing a fusion attention network as a deep feature extractor, inputting the superficial facial features into the fusion attention network to obtain deep features, wherein the fusion attention network includes a plurality of fusion attention groups, each fusion attention group includes a plurality of fusion attention blocks; and performing up-sampling on the deep feature map, and reconstructing the face feature map subjected to up-sampling into a high-resolution face image of the target. The method is superior to other latest face image super-resolution algorithms, and face high-resolution images with higher quality can be generated.

The embodiment of the invention discloses a model training method, a business processing method and device, a terminal and a medium, the model training method comprises the following steps: obtaininga sample set, the sample set comprising a plurality of sample images and an image relationship, and the image relationship comprising an association relationship between at least two sample images; constructing a data structure based on the image relationship and the plurality of sample images; traversing each element of the data structure and clustering the sample images stored by each element toobtain a plurality of clustering sets and category information of the sample images in each clustering set; carrying out model training on an image processing model by adopting the plurality of sample images and the category information of each sample image; the embodiment of the invention can improve the model training efficiency and reduce the labor cost.

An image privacy perception method based on deep learning, including the following steps: S1, building a privacy classification data set with labeled categories, and training a privacy perception network with a transfer learning method; S2, recognizing a privacy image using a deep convolutional neural network oriented to privacy perception; and S3, extracting an attention profile according to deep convolutional features of the neural network, and locating an attention focusing region to complete the perception of an image privacy region. The method has the following beneficial effects: by completing end-to-end training and testing based on the deep neural network, the privacy image can be accurately distinguished with the privacy region located, facilitating the selective protection of the privacy information in the image.

The invention relates to a resume information extraction method and system. The resume information extraction method comprises the steps: A, obtaining resume data; B, converting the resume data into resume texts by utilizing a BERT Chinese pre-training model and a data augmentation technology, and classifying the resume texts according to sentence characteristics of the resume texts; C, conductingnamed entity recognition on the classified resume text sentences through a BERT + BiGRU + CNN + CRF model, and then extracting needed information elements; and D, storing the extracted information elements in a database, and outputting corresponding information in a structured manner. The resume information extraction system is mainly composed of a resume acquisition module, an input module, a classification module, an information element extraction module, a storage module and an output module. The resume information extraction method uses an incremental learning method, and uses a clause mode in data preprocessing of the classification model, so that the language model can adjust parameters by incrementally inputting new training data on the basis of inheriting past parameters, and theresume information extraction method has better continuity and generalization ability.

The invention belongs to the technical field of hyperspectral image processing, and discloses a spectral super-resolution adaptive weighted attention mechanism deep network data processing method, which comprises the steps of selecting two spectral reconstruction challenge data sets for verification; evaluation index selection: using a root mean square error RMSE and an average relative absolute value error MRAE as evaluation indexes, and then calculating the MRAE and the RMSE; constructing an adaptive weighted channel attention mechanism module; constructing a second-order non-local module based on partitioning; and constructing an adaptive weighted attention mechanism network and training. According to the invention, a brand-new adaptive weighted attention mechanism network is designed for spectral super-resolution, and a backbone network is formed by stacking a plurality of double-residual modules which realize double-residual learning through long and short connection. And a self-adaptive weighted channel attention module is embedded in the double-residual module, so that the channel characteristic response is recalibrated, and the characteristic expression capability of the network is enhanced.

The invention provides a multi-task learning deep network model training and target detection method and device, and relates to the technical field of computer vision and deep learning. The specific implementation scheme is as follows: obtaining training data of multi-task learning in a target detection scene; inputting the image sample into a backbone network in a multi-task learning deep network model based on Anchor-Free to obtain a feature map output by the backbone network; inputting the feature map into a feature pyramid network in a multi-task learning deep network model to obtain a multi-scale feature map output by the feature pyramid network; inputting the multi-scale feature map into a Head network in a multi-task learning deep network model to learn each task, and obtaining a prediction result corresponding to each task output by the Head network; and training a multi-task learning deep network model according to the prediction result corresponding to each task output by the Head network and the label on the image sample corresponding to each task. According to the invention, the representation capability of the network can be improved.