37 results about "Score fusion" patented technology

The invention discloses a method and a system for detecting voice spoofing attack of speakers on the basis of deep learning. The method includes constructing audio-frequency training sets, initializing deep feed-forward neural networks and deep recurrent neural networks and respectively training the deep feed-forward neural networks and the deep recurrent neural networks by the aid of multi-frame feature vectors and single-frame vector sequences of the training sets; respectively leading frame level and sequence level feature vectors of to-be-tested audio frequencies into two trained linear differential analysis models in test phases, weighting two obtained result grades to obtain scores and comparing the scores to predefined threshold values so as to discriminate voice spoofing. The method and the system have the advantages that local features can be captured, and global information can be grasped; the linear differential analysis models are used as classifiers in identification and verification phases, the voice spoofing attack can be judged by means of grade fusion, and accordingly the voice spoofing detection accuracy can be greatly improved.

The invention discloses a face recognition method and system based on adaptive score fusion and deep learning. The system comprises a near infrared face and visible light face acquisition unit for jointly acquiring pictures of the same face by cooperation of cameras of two modes, a face feature extraction unit for respectively extracting features from the acquired near infrared face picture and visible light face picture via a deep convolution model, an adaptive score fusion unit for solving the similarity (that is, the score of the pictures) of the two pictures via an Euclidean distance and then fusing the face features extracted by using a deep learning algorithm via an adaptive score fusion algorithm on the score level to solve the final score of the user, and a result output unit for outputting the final score. The method and the system can solve the influence of the intensity of near infrared light on face imaging, and can also solve the problem that the details of the near infrared picture disappear by using the visible light picture.

The invention discloses a voiceprint recognition method based on gender, nationality and emotion information. The method specifically comprises the following steps: firstly, data preprocessing, secondly, feature extraction, and secondly, neural network parameter training: according to the specific structure of a neural network, in the training process, the input sequence of training sentences is disrupted, then 128 sentences are randomly selected as a training batch, and the number of data iterations is 80; wherein a training file required by the scoring fusion tool is a development set resultand a test set result of each system; wherein VOXCELEB1 test is used in the test set; wherein the development set is a test file which is generated based on 1211 training speaker statements and comprises 40,000 test pairs; and a final test set scoring result is obtained through 100 times of iteration. According to the invention, the recognition rate is improved.

A score fusion method and apparatus. The score fusion method includes receiving a plurality of scores respectively obtained by a plurality of classifiers, and fusing the received scores using a likelihood ratio. The score fusion apparatus includes a fusion unit which receives a plurality of scores respectively obtained by a plurality of classifiers and fuses the received scores using a likelihood ratio, and a parameter providing unit which provides the fusion unit with a plurality of parameters needed for fusing the received scores.

The invention discloses a behavior recognition method based on a space-time attention enhancement feature fusion network, and belongs to the field of machine vision. According to the method, a networkarchitecture based on an appearance flow and motion flow double-flow network is adopted, and is called as a space-time attention enhancement feature fusion network. Aiming at a traditional double-flow network, simple feature or score fusion is adopted for different branches, an attention-enhanced multi-layer feature fusion flow is constructed to serve as a third branch to supplement a double-flowstructure. Meanwhile, aiming at the problem that the traditional deep network neglects modeling of the channel characteristics and cannot fully utilize the mutual relation between the channels, the channel attention modules of different levels are introduced to establish the mutual relation between the channels to enhance the expression capability of the channel characteristics. In addition, thetime sequence information plays an important role in segmentation fusion, and the representativeness of important time sequence features is enhanced by performing time sequence modeling on the frame sequence. Finally, the classification scores of different branches are subjected to weighted fusion.

Embodiments of the invention disclose an offline identity authentication method and apparatus, which is used for solving the problems that dependence on a human face database of the public security ministry exists and a chip head portrait, a certificate surface head portrait and a certificate holder head portrait are difficultly identified in consistency in the prior art. The method disclosed by the embodiment of the invention comprises the steps of obtaining two or more images used for identity authentication to form a to-be-authenticated multivariate image group; extracting a concatenated PCA convolutional characteristic of each image in the to-be-authenticated multivariate image group to obtain a corresponding eigenvector; performing fusion on image information of the to-be-authenticated multivariate image group according to the eigenvector and a score fusion policy with a supervisory signal to obtain a corresponding fusion vector; and inputting the fusion vector into a pre-trained SVM classifier to perform authentication judgment on whether images in the to-be-authenticated multivariate image group are consistent or not so as to obtain an identity authentication result.

The invention relates to the technical field of computer vision, in particular to a face attribute classification system based on a bidirectional Ladder structure, which aims to solve the problem of how to fully utilize characteristics of different levels in a deep network and corresponding relations between characteristics of different levels and different face attributes so as to improve the accuracy of face attribute classification. For this purpose, the face attribute classification system based on the bidirectional Ladder structure provided by the invention comprises a bidirectional Ladder self-encoder module, a self-adaptive attention module and a self-adaptive score fusion module, the bidirectional Ladder auto-encoder module comprises an encoder module and a decoder module, the self-adaptive attention module comprises a plurality of attention sub-modules, and the self-adaptive score fusion module is configured to obtain a face attribute classification result of the face image tobe detected according to an output result of the encoder module and a result output by the attention sub-module. Based on the structure, the coding features and the decoding features of different levels can be fully utilized, and the accuracy of face attribute classification is improved.

The invention relates to a stereoscopic video quality evaluation method based on a 3D convolution neural network. The stereoscopic video quality evaluation method based on the 3D convolution neural network comprises the following steps: preprocessing data; training the 3D convolution neural network; and performing quality score fusion: dividing the whole test video into two parts randomly, whereinone part is used for the training of a 3D CNN model, and the other part is used for model test; and obtaining a prediction score of each input video block from the test stereoscopic video after the training process of the 3D CNN model. In order to obtain the overall evaluation score of the video, a quality score fusion strategy considering global time information is adopted: firstly, performing integration on the cube-level scores on the spatial dimension by means of the average pooling; defining the weight of each segment calculated based on the motion intensity to simulate global time information; then calculating the weight of the motion intensity of each time dimension in the total motion intensity of the stereoscopic video; and finally, summarizing the video-level prediction scores as the weighted sum of each time dimension to obtain stereoscopic video fusion quality score.

The embodiment of the invention provides a conversion rate evaluation method and a device, and the method comprises the steps: obtaining feature data according to a commodity attribute table and a user attribute table; inputting the obtained feature data into a preset click Wide & Deep model, outputting the click probability of the user to the commodity link, inputting the obtained feature data into a sharing Wide & Deep model, and outputting the sharing probability of the user to the commodity link; according to the click probability and the sharing probability, after score fusion is carriedout, obtaining the conversion probability of the commodity by the user; wherein the click or share Wi-Fi & Deep model is obtained by training according to the sample commodity attribute table and thesample user attribute table of the determined click or share result. According to the method, the click probability and the sharing probability of the user on the commodity can be accurately obtained,and the determined conversion probability is more accurate through fusion of the two fine ranking scores and the recall score.

The invention provides a voice keyword detection method based on complementary model score fusion. The method comprises the following steps of 1), introducing keyword modeling based on i-vector on thebasis of keyword modeling in an audio feature space, (2) self-adaptive segmented window shift: for a voice sample to be detected, intercepting a voice segment from the starting signal, obtaining thedistribution expression of the current segment in the voice feature space, calculating the similarity between the distribution expression and the keyword class attributes to obtain a class score sequence of the current segment, obtaining the window shift of the next segment according to the score of the current segment, processing segment by segment until the signal is ended, and dividing a voicesample to be detected into K segments, and 3) conducting score fusion by using the position of a keyword candidate point. According to the method, a keyword detection algorithm with certain complementarity is realized by adopting two different models, the score results of the two models are fused, the problem of voice keyword detection under the condition of a small training sample amount can be solved, and meanwhile, the keyword detection accuracy can be improved.

The invention provides a language identification and classification method based on a noise reduction automatic encoder. The method includes: step 1) extracting a to-be-identified speech signal from ato-be-identified speech segment to obtain underlying acoustic characteristics; step 2) extracting an original i-vector from the underlying acoustic characteristics obtained in step 1); step 3) calculating and obtaining a phoneme vector pc (u); step 4) splicing the original i-vector with the phoneme vector pc (u), and then inputting to a DAE-based i-vector compensation network to obtain a compensated i-vector; step 5) inputting the original i-vector obtained in step 2) and the compensated i-vector obtained in step 4) respectively into a pre-trained logistic regression classifier to obtain corresponding score vectors; and step 6) performing score fusion on the corresponding score vectors obtained in step 5) to obtain a final score vector, further obtaining a probability of each language category, and determining the language category to which it belongs.

Various embodiments of the present invention provide systems and methods for determining a personalized fusion score. In certain embodiments, the systems and methods are configured for calculating preliminary fused scores for consumers at least in part by applying a first score fusion technique across the sample of consumer data. Segmentation scores are then calculated based at least in part upon the preliminary fused scores. In those and other embodiments, the segmentation scores enable creation of a plurality of cluster subsets within the sample of consumer data. In certain embodiments cluster subsets are defined at least in part by a particular score mix, while in other embodiments subsets are defined at least in part by respective score fusion techniques that prove optimal for each subset. Further, in various embodiments, application of multiple score fusion techniques across respective cluster subsets provides personalized fusion scores for the consumers in each respective cluster subset.

The invention discloses a comparison score fusion method and device for multi-modal biological recognition, a medium and equipment, and belongs to the field of biological recognition. The method comprises the following steps: acquiring comparison scores of various biological features; normalizing the comparison score of each biological feature according to a plurality of set thresholds, wherein the plurality of thresholds are corresponding comparison thresholds when the FAR of the biological feature on the training set is a plurality of specific values; and performing weighted summation on the normalized comparison score according to a weighting coefficient to obtain a fused comparison score, wherein the weighting coefficient of the comparison score of each biological feature is calculated and obtained according to the corresponding FRR when the FAR of the biological feature on the training set is a plurality of specific values. According to the method and device, high precision can be achieved on data sets of banks and the like.

The invention belongs to the technical field of end-to-end speech recognition, and discloses a fusion method based on an end-to-end speech recognition model and a language model, and the method comprises the following steps: S1, training an end-to-end speech recognition model through employing a speech and text pair, and training an external language model through employing text data; s2, independently taking out a decoder part of the trained speech recognition model and forming an independent model; s3, independently training the independent model by using training data to a text, and obtaining an estimation model of the internal language model after convergence; and S4, the score fusion of the estimation models of the speech recognition model, the external language model and the internal language model is decoded to obtain a decoding result. The algorithm can improve the recognition accuracy after the speech recognition model and the language model are fused, and has a wide application prospect in the field of speech recognition.

The invention discloses a human body behavior identification method based on a motion cooperation space. The method comprises the following steps: S1, respectively carrying out key frame extraction based on a motion state measurement coefficient on an initial skeleton sequence and a depth map sequence; s2, extracting motion coordination space vectors from the skeleton sequence processed in the step S1, and splicing the motion coordination space vectors into motion coordination space features; extracting DMM features from the depth map sequence processed in the step S1 to obtain a depth motion map; and S3, inputting the deep motion map and the motion collaborative spatial features into a deep network at the same time, and performing score fusion. Based on the idea of multi-modal fusion, skeleton data and depth data are combined, so that the data are more complete, complementation of various heterogeneous information is realized, redundancy among modals is eliminated, a new behavior recognition method system is established, and a new thought and theoretical basis are provided for research and application of a human body behavior recognition method.

The invention discloses a biological feature multi-modal fusion recognition method and device, a storage medium and equipment, and belongs to the field of biological recognition. Adjusting a comparison threshold value of the latter modal according to the quality score of the biological characteristic image of the latter modal in the multi-modal identification; when the comparison score of the latter mode is not smaller than the adjusted comparison threshold value of the latter mode and the quality score of the biological feature image of the latter mode is not smaller than the decision threshold value of the latter mode, judging that the recognition is passed or performing the recognition of the subsequent modes; otherwise, fusing the comparison scores of all the identified modals to obtain a fused score; and according to the fusion score, judging whether the identification is passed or carrying out subsequent modal identification. According to the method, the comparison threshold value is dynamically adjusted according to the quality score of the biological feature image, auxiliary decision making is carried out on the recognition result through the quality score of the biological feature image, meanwhile, the probability of identity authentication is gradually increased in combination with comparison score fusion, and the safety and reliability of biological identity recognition are improved.

The invention discloses a biometric feature recognition multi-modal fusion method and device, a computer readable storage medium and equipment, and belongs to the field of biometric recognition. In the process of registration and verification/recognition of biological characteristics, various biological characteristic samples are collected, the mass fraction of the biological characteristic samples are obtained through sample quality judgment, the mass fraction and preset fusion threshold control parameters are compared according to a fusion strategy, and if fusion conditions are met, fusion processing is carried out. The fusion comprises sample data fusion, feature data fusion, comparison score fusion and comparison result fusion. Whether the fusion condition is met or not is judged according to the mass fraction of the multiple biological characteristic samples, fusion of the multiple biological characteristics is achieved, the limitation existing when identity recognition authentication is carried out through a single biological characteristic in a traditional method can be overcome, biological characteristic fusion processing is more reasonable and more effective, and a more comprehensive and safer guarantee is provided for identity identification and authentication.

The invention discloses a multi-modal score fusion method and device, a computer readable storage medium and equipment, and belongs to the field of biological recognition. The method comprises the following steps: acquiring a first modal comparison score and a second modal comparison score, and normalizing the first modal comparison score and the second modal comparison score by using LFAR transformation; forming a score pair by the normalized first modal comparison score and the normalized second modal comparison score, and performing polynomial kernel mapping on the score pair to obtain high-dimensional data; and inputting the high-dimensional data into the trained logistic regression model to obtain a fused comparison score. The method can give full play to the advantages of different modes, and achieves good recognition precision.

Various embodiments of the present invention provide systems, methods, and computer-program products for fusing at least two scores. In various embodiments, at least two scores are received in which each score predicts the probability of an outcome associated with a particular unit. In particular embodiments, A mass and a distance are calculated between two objects based on the at least two scores in which the first of the two objects is a constant and the second of the two objects comprises one or more characteristics of the particular unit. Further, in particular embodiments, a gravitational force between the two objects is calculated based on the mass and the distance and this gravitational force is used as a fused score for the at least two scores.