36results about How to "Enhanced Feature Learning" patented technology

The invention discloses a multimodal speech emotion recognition method based on an enhanced depth residual neural network, which relates to the technical fields of video stream image processing, speech signal analysis and the like, and solves the emotion recognition problem of human-computer interaction. The invention mainly comprises the following steps: extracting the feature expression of video(sequence data) and speech, including converting the speech data into the corresponding speech spectrum expression, and encoding the time sequence data; Convolution neural network is used to extractthe emotional features of the original data for classification. The model accepts multiple inputs and the input dimensions are different. The cross-convolution layer is proposed to fuse the data features of different modes, and the overall network structure of the model is enhanced depth residual neural network. After the model is initialized, the multi-classification model is trained with speechspectrum, sequence video information and corresponding affective tagging. After the training, the unlabeled speech and video are predicted to obtain the probability value of affective prediction, andthe maximum value of probability is selected as the affective category of the multi-modal data. The invention improves the recognition accuracy on the problem of multi-modal emotion recognition.

The invention provides a deep learning remote sensing cloud detection method based on Gabor transformation and an attention mechanism, and solves the problem of insufficient feature extraction in remote sensing image cloud detection. The method comprises the following implementation steps: establishing a remote sensing image database and a corresponding mask map; constructing a convolutional neural network comprising a Gabor transformation module and an attention module; determining a loss function of the network; inputting a training sample in the training image library into the convolutionalneural network, and iteratively updating the loss function through a gradient descent method until the loss function converges to obtain a trained convolutional neural network; and inputting the datain the test database into a convolutional neural network to obtain a detection result of the cloud area. According to the invention, the image feature extraction technology based on Gabor transformation and an attention mechanism is adopted, a deep learning method is used for cloud detection of the remote sensing image, feature extraction is sufficient, detection precision is high, and the methodis used for the preprocessing process of the remote sensing image.

The invention belongs to the field of image processing, and particularly relates to a CT image segmentation method based on an improved AU-Net network, and the method comprises the steps: obtaining a to-be-segmented brain CT image, and carrying out the preprocessing of the obtained brain CT image; inputting the processed image into the trained improved AU-Net network for image recognition and segmentation to obtain a segmented CT image; identifying a cerebral hemorrhage area according to the segmented brain CT image. The improved AU-Net network comprises an encoder, a decoder and a hopping connection part. Aiming at the problem of low segmentation precision caused by large size and shape difference of the hemorrhage part of the cerebral hemorrhage CT image, the invention provides a coding-decoding-based structure, and a residual octave convolution module is designed in the structure, so that the model can segment and identify the image more accurately.

The invention relates to a three-dimensional target detection system based on laser point cloud; the system comprises a voxel size division module, a feature coding module, a feature extraction and fusion module, a target regression and detection module and a laser radar. The output end of the laser radar is connected with the input end of the target regression and detection module through the voxel size division module, the feature coding module and the feature extraction and fusion module in sequence, and during use, firstly, the voxel size division module performs voxel division on a three-dimensional target point cloud obtained from the laser radar by adopting different voxel scales; a plurality of voxelized point clouds are obtained, then feature coding is performed on the plurality of voxelized point clouds by a feature coding module; feature extraction and fusion are performed on the coded voxelized point clouds by a feature extraction and fusion module to obtain a final feature map; and finally, a three-dimensional target detection box is obtained by a target regression and detection module according to the final feature map. The design can guarantee that the structural features of the point cloud are not lost, and improves the detection precision of the three-dimensional target.

The invention relates to a hyperspectral image classification method combining a 3D/2D convolution network and adaptive spectral unmixing, and the method comprises the steps: building a network modelthrough employing a 3D/2D dense connection network and a plurality of intermediate classifiers, and enabling the adaptive spectral unmixing to serve as the supplement of a network classification result. The design of multiple intermediate classifiers with early exit mechanisms enables the model to facilitate classification by using adaptive spectral unmixing, which brings considerable benefits tocomputational complexity and final classification performance. Besides, the invention further provides a 3D/2D convolution based on the spatial spectrum characteristics, so that the three-dimensionalconvolution can contain less three-dimensional convolution, and meanwhile, more spectral information is obtained by utilizing the two-dimensional convolution to enhance characteristic learning, so that the training complexity is reduced. Compared with an existing hyperspectral image classification method based on deep learning, the hyperspectral image classification method is higher in calculationefficiency and higher in precision.

The invention discloses a stereoscopic image visual salience extraction method based on a dual-learning network, which forms a training set of a human gaze map, a left viewpoint color image and a leftparallax image of the stereoscopic image. Then, based on the training set, the depth learning model is constructed by using the feature extraction technology of VGG network model. Secondly, the depthlearning model is trained by using the human gaze map as the monitor and the left viewpoint color image and left parallax image as the input parameters. Then, the left view color image and the left parallax image of the stereoscopic image to be significantly extracted from the vision are taken as input parameters and input into the model obtained from the training to obtain the visually significant image of the stereoscopic image to be significantly extracted from the vision; the advantage is that it can run quickly, and has strong robustness and prediction accuracy.

The invention provides an online and offline fused medical image quality interpretation self-adaption method and system, and a storage medium. According to the online and offline fused medical image quality interpretation method, ''online'' operation is reflected in that a doctor can acquire a medical image quality interpretation result in real time at the image acquisition end; ''off-line'' operation is reflected in correction of prediction results by an image storage end and a medical cloud server and retraining and upgrading processes of an AI model; and ''self-adaption'' is reflected in that a closed loop is formed in the whole online and offline processes, so that online real-time return and offline feedback correction are continuously iterated and optimized, and online and offline fusion is realized. The online and offline fused medical image quality interpretation method ensures the real-time performance of acquiring the quality interpretation result by the doctor, continuouslyimproves the generalization and accuracy degree of the model by retraining and upgrading the AI model, has great significance in clinical application, can greatly reduce the time cost of manual interpretation of doctors, improves the accuracy and efficiency of auxiliary diagnosis of intelligent medical image quality interpretation, and achieves early discovery and early modification of low-qualityimages to the maximum extent.

The invention is suitable for the technical field of machine learning, and provides a neural network model training method, which comprises the steps of inputting a training sample set into a preset student neural network model for processing to obtain first feature information, and calculating a first loss value according to the first feature information and a first loss function; inputting the training sample set into a preset teacher neural network model for processing to obtain fourth feature information, and calculating a second loss value according to the first feature information, the fourth feature information and a second loss function; determining a target loss value according to the first loss value and the second loss value; if the target loss value does not meet the preset stopping condition, updating the preset student neural network according to the target loss value; and if the target loss value meets the preset stopping condition, outputting the trained student neural network model. According to the invention, the training convergence condition of the neural network model can be effectively improved, so that the neural network can better learn features, and the recognition precision and generalization ability of the neural network are improved.

The invention discloses a cross-field text classification method based on a self-adaptive noise reduction encoder. The method includes: using a feature selection method suitable for use in a cross-field task to filter out feature words with low appearing frequency and meaninglessness in samples in a source-field data set and a target-field data set, calculating a better noise interference coefficient in a self-adaptive manner according to distribution difference between the samples in the source-field set and the target-field set, using the better noise interference coefficient to interfere with feature space, using a stacking edge noise reduction encoder method to construct new feature space, and constructing a classifier. According to the method, relationships between potential featuresbetween fields can be better mined, field difference can be reduced, and thus correctness of classification can be improved.

The invention discloses a large-scale point cloud semantic segmentation method based on a lightweight neural network. The main body of the method is a semantic segmentation neural network model which comprises a feature extraction network and a feature propagation network. The feature extraction network is used for extracting global features and local features of the three-dimensional point cloud, the feature propagation network propagates the features to original points, and an output feature map corresponds to the probability that each point belongs to each semantic category. According to the method, a non-local module is designed in local and non-local modules, feature learning of the local module is enhanced, network weight parameters are updated by using a focus loss function, the weight of a simple sample is automatically reduced in the training process, and the model is quickly concentrated on a difficult sample. In terms of performance, the training time and the memory space are greatly reduced, and meanwhile, the method has very high point cloud semantic segmentation precision.

The invention relates to a microblog comment viewpoint object classification method based on an event graph convolutional neural network, and belongs to the technical field of natural language processing. The method comprises the following steps: taking a microblog text and comments as document nodes, explicitly taking a keyword co-occurrence relationship, a reply relationship and document similarity as weights of edges of the document nodes, and constructing an adjacent matrix of a graph convolutional neural network on the basis of the weights; giving more weights to the initial features of the document nodes and the word nodes closely related to the keywords; and finally, under the supervision of a small number of labels, learning expressions of word nodes and document nodes so as to finish classification. Experimental results on two event microblog data sets show that compared with other reference models, EventGCN can significantly improve the classification performance of viewpoint objects.

The invention discloses a remote sensing scene classification method and system based on homogeneity and heterogeneity Transformers. An input remote sensing scene picture is divided into a heterogeneous Patch and a homogeneous Patch by using a direct division method and a superpixel segmentation method respectively; a heterogeneous feature learning branch and a homogeneous feature learning branch are used to simultaneously extract heterogeneous information and homogeneous information in the divided heterogeneous Patch and the homogeneous Patch, and global information, local knowledge and related context information in a remote sensing scene are extracted based on a Transform structure to obtain a heterogeneous feature he-token0 and a homogeneous feature ho-token0; the heterogeneous feature he-token0 and the homogeneous feature ho-token0 are fused, the feature discrimination is enhanced by combining a metric learning mechanism, and finally remote sensing scene classification is completed according to the fused features. The method comprehensively extracts the feature representation of the remote sensing image, and has better remote sensing scene classification performance.

The invention discloses an urban fine-grained flow prediction method and system based on limited data resources. The method comprises the following steps: acquiring flow distribution data and corresponding external factor data of a to-be-predicted area within a certain time; obtaining a fine granularity flow distribution diagram and a coarse granularity flow distribution diagram according to a set coarse and fine granularity scaling factor; performing down-sampling on the coarse-grained flow distribution diagram to obtain a down-sampled coarse-grained flow diagram; and training a spatial inference encoder from a down-sampling coarse-grained flow diagram to a coarse-grained flow distribution diagram, wherein the spatial inference encoder is used for predicting the fine-grained flow of the region. Under the condition that training resources are limited, an inference network is provided to simplify the urban fine-grained traffic prediction problem, and the traffic prediction precision and reliability of the model are improved.

The invention relates to a brain image classification method and device based on magnetic resonance imaging and deep learning, and the method comprises the following steps: obtaining original MRI scanning image data to be classified, and carrying out the preprocessing, and obtaining a three-dimensional MRI image; taking the three-dimensional MRI image as an input of a trained classification network model to obtain a final classification result; wherein the classification network model carries out 3D to 2D feature dimension conversion on the three-dimensional MRI image along an axial plane, a sagittal plane and a coronal plane respectively, corresponding initial classification results are obtained through an axial locus base classifier, a sagittal locus base classifier and a coronal locus base classifier respectively, and integrating and fusing the multiple initial classification results to acquire the final classification result. Compared with the prior art, the invention has the advantages of high classification precision, low model complexity and the like.

The invention discloses a text classification method based on a semi-supervised graph convolutional neural network, and the method comprises the steps: coding texts into fixed vectors through employing a BERT model in order to construct a semantic relation between the texts, analyzing a similar relation between the texts, and constructing a side relation between documents. The feature representation of the text can depend on similar document features, the neighbor node features of the document nodes are aggregated by using the graph convolutional neural network to perform feature learning, and the feature representation of the target document nodes is enhanced. By adopting the GMMM model, not only can feature learning of the nodes be promoted, but also label information spreading can be performed, and the problem of sparse label data is effectively solved.

The invention discloses a pedestrian re-identification method and system based on mixed supervision. The method comprises the following steps: selecting a single-camera labeled pedestrian image and an unlabeled pedestrian image as training image samples; clustering the unlabeled pedestrian image, segmenting the unlabeled pedestrian image into a clustered pedestrian image and a non-clustered abnormal value, and judging the clustering reliability; providing supervision signals of a single-camera labeled pedestrian image, an unlabeled clustered pedestrian image and a non-clustered abnormal value, and carrying out joint training on the supervision signals and an encoder; and dynamically updating supervision signals of the single-camera labeled pedestrian image, the unlabeled clustered pedestrian image and the non-clustered abnormal value. According to the invention, joint feature learning is carried out through all available information of the single-camera labeled data and the unlabeled data, and training is carried out by using the most reliable clustering of the unlabeled data, so that a more reliable learning target is provided.

The invention relates to the field of danger perception, in particular to an intelligent helmet for danger perception and a perception system, and the helmet comprises a helmet body, a sound collection device, a processor and a danger early warning device; the processor obtains sound feature information and the distance between the helmet and the sound source based on the sound signals obtained by the sound collection device, processes the sound feature information and the distance, and obtains the danger degree and the response time needed when people perceive the danger; the danger early warning device executes danger early warning when the personnel do not react within the reaction time; the sensing system comprises a data acquisition module used for acquiring sound feature information and the distance between the helmet and the sound source; the data processing module is used for processing the sound feature information and the distance to obtain the danger degree and the response time required when people perceive the danger; and the danger early warning module is used for generating a danger early warning prompt when the personnel do not react within the reaction time.