58 results about "Triplet loss" patented technology

The embodiment of the invention discloses a face recognition method and a device. The method comprises the steps of extracting the Haar feature of a current to-be-recognized face image, and detecting the human face area of the to-be-recognized face image by adopting an ADaBoost classifier so as to obtain a face region image; performing the multi-scale feature extraction on the face region image by utilizing a convolution neural network model, and obtaining a feature vector of the face region image; inputting the feature vector, a pre-built legal face database and a preset user similarity threshold value into a multi-task learning model pre-constructed according to a Softmax loss function and a Triplet loss function, and judging whether the to-be-recognized face image is a legal user or not according to the output value of the multi-task learning model. The extracted feature is good in robustness and good in generalization ability. Therefore, not only the face recognition rate improved, but also the accuracy of face recognition is improved. The safety of identity authentication is improved.

The present invention relates to a face verification method and device based on the Triplet Loss, a computer device and a storage medium. The method comprises the steps of: based on a face verification request, obtaining a certificate photo and a figure scene photo; performing face detection, key point positioning and image preprocessing of the scene photo and the certificate photo, and obtaininga scene face image corresponding to the scene photo and a certificate face image corresponding to the certificate photo; inputting the scene face image and the certificate face image into a pre-trained convolutional neural network model used for face verification, obtaining a first feature vector, corresponding to the scene face image, output by the convolutional neural network model and a secondfeature vector, corresponding to the certificate face image, output by the convolutional neural network model; calculating a cosine distance of the first feature vector and the second feature vector;and comparing the cosine distance with a preset threshold, and determining a face verification result according to a comparison result. The reliability of face verification is improved by the method.

The invention discloses a triple loss-based improved neural network pedestrian re-identification method. The method comprises the following steps of constructing a sample database, establishing positive and negative sample libraries based on the sample database, and randomly selecting two positive samples and one negative sample to form a triple; constructing a triple loss-based neural network, and performing training, wherein the neural network is formed by connecting three parallel convolution neural networks with a triple loss layer; inputting a to-be-tested picture and each sample picture in the expanded sample database, which serve as a group of inputs, to the trained neural network in sequence, wherein another input of the neural network is zero or zero input; and calculating a distance of eigenvectors of two input pictures output by the neural network by utilizing a Euclidean distance, and querying and arranging first 20 Euclidean distances in an ascending order, and then performing simple manual screening to obtain a final identification result. The method has the beneficial effects that the identification method can be suitable for a picture scene with a relatively great change, can ensure robustness, and has relatively high identification accuracy.

The invention discloses a fine-grained vehicle type identification method based on weak supervision positioning and subclass similarity measurement, which comprises the following steps: 1) weak supervision positioning: using a pre-trained VGG-NET network to locate the image object, processing the mask map of convolution layer, and obtaining the boundary frame of the object. 2) constructing a fuzzysimilarity matrix: features of the pictures in the training set after the positioning of the picture are extracted by using B-CNN, and a fuzzy similarity matrix to measure the similarity of each subclass is obtained according to softmax classification results; 3) performing sampling to form a triple set: sampling to form a triple set on that basis of a fuzzy similarity matrix; 4) joint learning of the improved loss function: the improved triplet loss and the weighted softmax loss are used to restrict the distance between samples of the same sub-category and increase the distance between samples of different sub-categories. Compared with the original model, the invention is more accurate in positioning, and the classification accuracy is obviously improved, so that the vehicle target can be positioned well.

A system and method for using triplet loss for proposal free instance-wise semantic segmentation for lane detection are disclosed. A particular embodiment includes: receiving image data from an image generating device mounted on an autonomous vehicle; performing a semantic segmentation operation or other object detection on the received image data to identify and label objects in the image data with object category labels on a per-pixel basis and producing corresponding semantic segmentation prediction data; performing a triplet loss calculation operation using the semantic segmentation prediction data to identify different instances of objects with similar object category labels found in the image data; and determining an appropriate vehicle control action for the autonomous vehicle based on the different instances of objects identified in the image data.

The invention discloses a pedestrian re-identification method based on depth learning, includes Step 1: pre-training the CNN model: including pedestrian feature extraction and feature measurement, wherein pedestrian feature extraction adopts the method of combining global feature and local feature, Euclidean distance is used as similarity measure, and loss function based on measure matrix is established under the constraint condition of distance of eigenvector. The loss function is optimized by adding constraint function on the basis of traditional Triplet Loss. 2, testing the data set: inputting the image of the teste data set into the CNN model trained in the step 1, obtaining the image features, calculate the similarity between the target pedestrian image and the reference pedestrian image by the Euclidean distance, and finally arranging the reference pedestrian image according to the similarity magnitude to obtain the pedestrian re-recognition result. The method is suitable for there-recognition of people in complex scenes, and has strong transplantability for scene changes, stable algorithm, fast speed and strong practicability.

The invention relates to a fast surveillance site image vehicle retrieval method and system based on deep learning. According to the method, a deep neural network is adopted to extract vehicle feature information, vehicle features are extracted on the basis of the inception_resnet_v2 network, network weight sharing is realized, a large amount of repeated calculation is effectively avoided, a lossfunction thereof is a triplet loss function, and is adopted to carry out triplet sample training, and 128-dimension vectors are directly generated. In an image retrieval stage, the method adopts, toestablish an index for the features, a manner of clustering the features, and increases a query speed. According to the method, an extraction speed of image features can be accelerated, and partial illegal vehicles of fake-plate vehicles and clone vehicles can be effectively inspected, traced and arrested while fast real-time responding is carried out.

The invention discloses a face recognition method based on distinguishability feature fusion, comprising the following steps: A, cutting a global image and at least two local images in each training sample image; B, carrying out model training on each intercepted image by adopting a multi-loss function to obtain a corresponding model, wherein the multi-loss function is obtained by combining the a-softmax loss function with the center loss function. C, carrying out fusion and dimension reduction of each model obtain by training by using a ternary loss triplet loss function, and obtaining the final depth feature of the training sample image. The technical proposal disclosed in the application can solve the problem of data fusion, face posture and model fusion in the process of face recognition by using CNN, and achieve better face recognition effect.

A method and apparatus of training an acoustic feature extracting model, a device and a computer storage medium. The method comprises: considering a first acoustic feature extracted respectively from speech data corresponding to user identifiers as training data; training an initial model based on a deep neural network based on a criterion of a minimum classification error, until a preset first stop condition is reached; using a triplet loss layer to replace a Softmax layer in the initial model to constitute an acoustic feature extracting model, and continuing to train the acoustic feature extracting model until a preset second stop condition is reached, the acoustic feature extracting model being used to output a second acoustic feature of the speech data; wherein the triplet loss layer is used to maximize similarity between the second acoustic features of the same user, and minimize similarity between the second acoustic features of different users.

The invention relates to a point cloud registration model and method combining an attention mechanism and a three-dimensional graph convolutional network, and the model is a three-branch Siamese architecture, and comprises a Dejector model and a Descriptor model. The Detector model is used for extracting attention features of points and constructing an attention mechanism; the Descriptor model isused for generating an expression of a three-dimensional depth feature to represent the three-dimensional depth feature of the point, and learning and judging the depth feature of the point cloud. Themethod comprises the following steps: carrying out model training, and constructing a loss function to train a model by using a failure align triplet loss, so as to effectively extract attention features and descriptor features from a point cloud; after model training, carrying out the point cloud registration. According to the method, the key points and the three-dimensional depth features of each key point can be automatically extracted, in the three-dimensional graph convolutional network, the multi-layer perceptron MLP is combined with the graph convolutional network GCN, a new point cloud feature extraction module is designed, more point cloud features with identification significance can be extracted, and the accuracy of point cloud registration is improved.

The present invention discloses a method and device of vehicle reidentification based on multiple objective function deep learning. The method comprises: employing first-class vehicle training data generated based on a vehicle image data set and a preset Softmax objective function to train the preset network parameters in a deep learning network frame, obtaining an optimal depth learning network frame, employing second-class vehicle training data generated based on the vehicle image data and a preset Triplet Loss ternary objective function to train the optimal network parameters in the depth learning network frame, obtaining a target depth learning network frame, and performing reidentification of the vehicle image to be identified according to the target depth learning network frame. Through adoption of the target depth learning network frame, the image features with high robustness and high reliability can be obtained to facilitate vehicle reidentification and effectively satisfy the requirement of the vehicle reidentification in the monitoring video.

A BERT-based customer service question and answer system belongs to the technical field of data calculation and comprises a receiving module, a preprocessing module, an intention module and a template engine module. The receiving module is used for receiving questions proposed by a user side; the preprocessing module is used for processing the received problem; the intention module is used for analyzing and acquiring the intention of the acquired problem; the template engine module is used for matching the obtained questions with standard questions to obtain question methods; an answer configuration module is used for generating answers for the questions provided by the system. According to the system, a BERT model is adopted for feature vector extraction; monitoring is carried out based on a triplet loss function of the Euclidean distance; compared with the adoption of a dichotomy cross entropy loss function, the generated vectors are more natural and reasonable in similarity distance calculation, and compared with a conventional training model, the triplet net simultaneously trains positive and negative samples, so that the model convergence is faster; meanwhile, the data in the system is in a closed-loop state, the modification period is shortened, and the accuracy of the system is improved.

The invention discloses a voice and face composite identity authentication method based on an end-to-end deep neural network. The method comprises the steps that a text-related voice signal and a facial video signal are collected, voice vocal print features and multi-frame face features are extracted, the voice vocal print features and the multi-frame face features are connected to obtain an identity feature vector, dimension reduction processing is carried out on the identity feature vector, and personal identity verification is carried out by adopting a Triplets Loss method. The voice and face composite identity authentication method based on the end-to-end deep neural network has the advantages that identity authentication is carried out by combining the voice and face features, the deficiency caused by a single feature is made up for, the voice and face composite identity authentication method can be applied to occasions such as entrance guard and check on work attendance, man-machine interaction and the like, and the identification accuracy is effectively improved.

The invention relates to a speaker recognition method based on Triplet-Loss. The speaker recognition method comprises the following steps: S1, acquiring voice signals, comprising three groups of samples, respectively a group of voice sequence of a speaker, another group of voice sequence of the same speaker and a group of voice sequence of a different speaker; S2, carrying out preprocessing on thevoice signals, so that channel noises generated during the voice collecting process are removed; S3, carrying out voice feature parameter extraction on the voice signals after denoising; S4, constructing an RNN neural network based on an LSTM neural network; S5, training the RNN neural network by adopting 90% of the extracted three groups of voice feature parameters as the input of the RNN neuralnetwork; and S6, after the RNN neural network is trained, carrying out speaker recognition by taking the rest 10% of the three groups of voice feature parameters as the input of the RNN neural network. The speaker recognition method has the advantages that the accuracy is high, the recognition effect is good, the reliability high, etc.

The invention discloses a ship plate recognition method based on deep learning feature comparison, and the method employs a deep learning convolutional neural network technology to construct a ship plate detection model and a ship plate character recognition model, is fast in calculation speed and high in precision, and has very high robustness for a variety of illumination, backgrounds, environments, ship appearance changes and the like. Variability and diversity of Chinese characters in ship plate character recognition are fully considered; the number recognition and the Chinese character recognition of the ship plate characters are separately processed; a staged training method is adopted, wherein the method comprises the following steps: firstly, performing training on a ship license digital data set based on logistic loss and cense entropy loss; and carrying out training on the ship plate Chinese character data set based on the logistic loss and the triplet loss, so that the training efficiency and the convergence speed are ensured. In addition, based on triplet loss training, the situations of more types and uniform distribution of ship plate Chinese character data sets can be effectively handled, the inter-class difference is increased while the distance in the class is reduced, and the recognition effect is improved.

A system for cross-modal data retrieval is provided which includes a neural network having a time series encoder and text encoder jointly trained based on a triplet loss relating to two different modalities of (i) time series and (ii) free-form text comments. A database stores training sets with feature vectors extracted from encodings of the training sets. The encodings are obtained by encoding the time series using the time series encoder and encoding the text comments using the text encoder. A processor retrieves the feature vectors corresponding to at least one of the modalities from the database for insertion into a feature space together with a feature vector corresponding to a testing input relating to at least one of a testing time series and a testing free-form text comment, determines a set of nearest neighbors from among the feature vectors based on distance criteria, and outputs testing results.

The invention relates to a multi-view gait recognition method based on deep learning, and belongs to the technical field of computer vision and machine learning, wherein the method comprises the steps: S1, collecting the gait of a person, randomly extracting n frames from each segment of video sequence, and extracting a gait contour diagram through a background subtraction method; S2, preprocessing the extracted n target contour diagrams; S3, representing periodic change by the amplitude of human motion, and detecting a contour diagram in a gait period; S4, mapping frames in the gait sequenceto different channels on the basis of the gait energy diagram of the single channel, and determining the boundary of the channel according to the amplitude of the frames; S5, combining the obtained multi-channel gait templates into a set for input, extracting features of each template, and then aggregating the features; and S6, selecting a metric learning method, proposing a Triplet Loss functionwith enhanced constraints to learn optimal features, and identifying a feature aggregation vector obtained by identification.

A system and method for using triplet loss for proposal free instance-wise semantic segmentation for lane detection are disclosed. A particular embodiment includes: receiving image data from an image generating device mounted on an autonomous vehicle; performing a semantic segmentation operation or other object detection on the received image data to identify and label objects in the image data with object category labels on a per-pixel basis and producing corresponding semantic segmentation prediction data; performing a triplet loss calculation operation using the semantic segmentation prediction data to identify different instances of objects with similar object category labels found in the image data; and determining an appropriate vehicle control action for the autonomous vehicle based on the different instances of objects identified in the image data.

The invention discloses an aerial image matching method based on local deep hashing. The method comprises the following steps: 1) calculating interval number N of an image to be matched according to aerial image overlapping ratio, and according to overlapping rate requirements, estimating a locally matched region; 2) carrying out optimization on the local region through a normalization cross-correlation algorithm; 3) constructing a Triplet network structure as a feature extraction network based on a VGG-F network, and replacing an output layer by a Hash layer to construct a Hash network; 4) carrying out improvement based on traditional Triplet loss, increasing absolute distance constraints and quantifying error loss as a loss function optimization network; and 5) detecting feature points according to a DoG algorithm and constructing feature point neighborhood as input of the network, obtaining a binary hash code of each image block through a trained network, and finishing matching in ahamming space through approximate nearest neighbor searching. The aerial image matching method based on local deep hashing has higher accuracy under the condition of meeting real-time performance.

The invention provides a multi-task Triplet loss function learning method based on a semantic hierarchy. The multi-task Triplet loss function learning method comprises the following steps: constructing a semantic hierarchy network for a database, performing triplets sampling on the semantic hierarchy, training the multi-task Triplet network, and performing multi-task classification by using a treeclassifier. The multi-task Triplet loss function learning method proposes a loss function combining a semantic hierarchy network with Triplet for solving the problem of multi-level training of the Triplet network, utilizes semantic knowledge to guide a network to differentiate sample structures hierarchically, learns a Triplet feature which contains semantic hierarchy information and is higher ingeneralization, effectively applies the Triplet feature to multi-task learning, and improves feature separability under different semantic hierarchies. Meanwhile, a new hierarchical sampling method is studied, so that the network can mine more effective hard triplets, and finally, the performance of the network is improved.

The invention relates to intelligent comprehensive overall quality evaluation in the field of intelligent environmental protection, and particularly discloses a training method of a neural network for intelligent comprehensive overall quality evaluation, which adopts an encoder-decoder architecture to evaluate the overall water quality, and particularly, in the training process, detection result information of a water quality sample and information of sampling time and sampling depth are fully utilized and fused through a multi-layer encoder structure to carry out encoding so as to obtain an encoded feature vector, and a triple loss function value of the encoded feature vector is calculated and is used for training the classifier so as to obtain a classification result of the water quality sample. The accuracy of comprehensive overall evaluation of the water quality is improved.

The invention discloses an image retrieval method based on sketches. The image retrieval method comprises the following steps of S1, tranining classification models of two CNNs corresponding to the sketches and photos respectively; S2, constructing a retrieval model by using the classification model obtained in the step S1, and training the retrieval model based on quadruplet loss; preprocessing the images in the image library; retrieving a single model; fusing results obtained by the plurality of retrieval models to obtain a final retrieval result. The method is based on the theory that the feature vector spacing corresponding to the sketch and the similar image is reduced, and the feature vector spacing corresponding to the sketch and the heterogeneous image is increased at the same time. Compared with triplet loss, the loss is reduced; quadruplet loss limits the distance between the sketch and the image and pays attention to the heterogeneous spacing of the image at the same time, so that the distribution of different types of images in the final feature space has higher class discrimination, i.e., a larger inter-class distance and a relatively smaller intra-class distance are generated, and thus the retrieval model has better performance.

The invention discloses a legal text representation method based on decoupling legal elements, which comprises: definition and labeling of the legal elements, data construction, model training, specific application of downstream tasks and the like, and combines field knowledge of the legal elements with a deep learning model. For a text containing a certain legal element, a triple is constructed by means of a positive sample containing the legal element and a negative sample not containing the legal element, and vector representation of the legal text about the legal element is learned based on triplet loss. For a plurality of legal elements, a triple is constructed for each legal element and a corresponding model is trained, and model training among different legal elements is mutually independent and does not influence each other. According to the method, the nuisance among the legal elements can be well captured, meanwhile, the influence of other irrelevant information is reduced, and a new solution is provided for text representation in the legal field.

The invention provides a finger vein indexing method based on multi-scale attention mechanism deep hash. The method comprises: obtaining a training set including finger vein feature data; traversing and expanding the training set to obtain an expanded training set; extracting finger vein depth features of a multi-scale attention mechanism based on the extended training set and a pre-constructed multi-scale attention feature extraction model; and based on the multi-scale attention mechanism finger vein depth features and a pre-constructed classification supervision model, hash coding model andretrieval task supervision model, performing collaborative training of triple loss functions to generate a final training model. According to the method, a neural network method is mainly adopted, a multi-scale attention mechanism is added, compared with a traditional feature extraction algorithm, the robustness of the features is improved to a certain extent, meanwhile, the model is end-to-end, the integrity of hash coding and feature extraction can be guaranteed, and the coding quality and the coding efficiency are guaranteed.

The invention provides a method for identification and localization of images using triplet loss and predicted regions and a medium. A method and a system for classifying image features using a neural network and system is provided. The method includes training the neural network using triplet loss processes including receiving an anchor image, selecting a positive image and a negative image, generating a image embedding associated with each of the anchor image, the positive image, and the negative image, classifying image features extracted from the anchor image based on the image embedding of the anchor image, determining an image label location associated with the classified image features, extracting features associated with the determined image label location, and classifying the features associated with the determined image label location; and combining the multi-label loss with localized image classification loss and the triplet loss using a weighted loss sum.

The invention provides a domain self-adaption method based on a triple and difference measurement, which comprises the following steps of: randomly extracting samples from a target domain to form a target domain batch, and inputting a target domain batch to obtain sample features; inputting the sample features into a multi-classifier, and performing entropy minimization processing; inputting the sample features into a multi-binary classifier at the same time, and determining k critical samples and k pairs of similar classes according to the output; then, screening effective samples by utilizing triplet loss to construct a source domain batch, and training a multi-binary classifier and a multi-classifier through an extracted source domain batch sample; and finally, sending the target domain batch and the source domain batch into the domain adversarial network, and carrying out domain alignment operation. According to the method, a triple loss function is used, the margin between positive and negative sample pairs in the loss is reasonably designed, and domain alignment is carried out by using a domain adversarial network, so that sample distribution of a source domain and a target domain tends to be consistent, and samples, close to a classification boundary, of the target domain are indirectly far away from the boundary; therefore, the samples of which the target domains are close to the classification boundary can be correctly classified.

The invention provides a fingerprint identification model construction method based on Resnet and Triplet Loss, and the method comprises the steps: constructing N groups of initial samples; training Ngroups of initial samples by using Triplet Loss to obtain N groups of training samples; inputting the N groups of training samples into the initial Resnet model to train the initial Resnet model to obtain a to-be-tested target Resnet model; inputting the two groups of test samples into a to-be-tested target Resnet model for calculation to obtain two groups of image characteristic quantities; calculating the maximum cosine similarity between the two groups of image feature quantities by using a preset algorithm; judging whether the maximum cosine similarity is greater than or equal to a presetvalue or not; generating a target Resnet model according to the to-be-tested target Resnet model; or constructing M groups of initial samples and inputting the M groups of training samples into the to-be-tested target Resnet model to train the to-be-tested target Resnet model. The invention further provides a storage medium and computer equipment. According to the method, the problem of narrow-edge fingerprint matching identification is effectively solved through the fingerprint identification model constructed by Resnet and Triplet Loss.

Methods are provided for generating training triplets that can he used to train multidimensional embeddings to represent the semantic content of non-speech sounds present in a corpus of audio recordings. These training triplets can be used with a triplet loss function to train the multidimensional embeddings such that the embeddings can be used to cluster the contents of a corpus of audio recordings, to facilitate a query-by-example lookup from the corpus, to allow a small number of manually-labeled audio recordings to be generalized, or to facilitate some other audio classification task. Thetriplet sampling methods may be used individually or collectively, and each represent a respective heuristic about the semantic structure of audio recordings.

The invention relates to intelligent prediction in the field of intelligent mining, and particularly discloses a training method of a neural network for intelligent prediction of a caving state, which is used for training the neural network for intelligent prediction of the caving state based on a triple loss function. Specifically, in a training process, partial randomness is introduced into a high-dimensional feature space of a surface image of a current mine wall to obtain a second feature map, and a distance loss function value between the second feature map and down-sampling of the high-dimensional space of the current image is calculated; a cross entropy loss function value is calculated with the second feature map by combining the weight of rock and sandy soil stacked on the current vertical plate to acquired a classification loss function value from the second feature map through a classifier, and a neural network is trained through triple loss function values, so that a classification result can be predicted to a certain extent. Meanwhile, the training speed and precision of the model can be enhanced.

The invention belongs to the field of image retrieval, and particularly relates to a hand-drawn image real-time retrieval method based on multi-granularity associative learning, which comprises the following steps of: training an improved deep neural network model by adopting a triplet loss function and a multi-granularity associative learning method, extracting an embedded vector of a sketch branch by the trained deep neural network model, sending the sketch branch to a discriminator to judge the grade of the sketch branch, sending the sketch branch to a dimension reduction layer corresponding to the grade, calculating the Euclidean distance between the sketch branch and the image, and returning the retrieved top-k images according to the Euclidean distance; according to the method, a multi-stage model is designed, diversity confusion of incomplete sketches is avoided, and a progressive multi-granularity association learning method for the incomplete sketches is provided, so that the embedding space of each incomplete sketch approaches the embedding space of a subsequent sketch and a corresponding target picture, and the target picture is retrieved with the fewest sketch strokes as much as possible.