515results about How to "Improve discrimination ability" patented technology

The invention discloses a deep learning-based vehicle detection method which is used for conducting vehicle detection in a complex environment by combining Edge Boxes and an improved Faster R-CNN model. The method comprises the steps that an image is processed by using the Edge Boxes, and an accurate vehicle candidate area is extracted primarily; and then the candidate area is input to the improved Faster R-CNN model to achieve precise positioning on vehicles, and a final detection result is obtained through classification and discrimination. In order to enhance the detection capacity on the small-size vehicles by the model and the discrimination capacity of the model, different layers of convolutional features are combined, some detail information of the vehicles is supplemented, a difficult negative sample mining strategy is added in the training stage, therefore, the model can pay more attention on difficult samples, and the vehicles can be well distinguished with the backgrounds ofsuspected vehicles.

The invention provides a vehicle model recognition model construction method based on depth learning. According to the method, a deep convolution neural network structure is designed for vehicle model recognition; softmax classification loss and sorting loss which is built through a vehicle model type hierarchical structure and is based on tetrad are used together to supervise the deep convolution neural network training; and intra-class differences and inter-class differences of vehicle models are constrained at the same time to learn character representation with rich semantic information and a discriminating classifier. In order to speed up network convergence, an online difficult sample mining policy and an improved gradient backtracking optimization algorithm are used, which greatly reduces the network training time. The invention further provides a vehicle model recognition method based on depth learning. The vehicle model recognition method uses the model constructed by the model construction method to recognize the vehicle type in a vehicle picture, which effectively improves the vehicle model classification accuracy.

The invention provides a remote sensing image semantic segmentation method based on context information and an attention mechanism. For a to-be-labeled high-resolution remote sensing image, data preprocessing and data amplification are firstly carried out, and then the amplified data is used for training a remote sensing image semantic segmentation model. In the training process, preliminary features are extracted from the image through a backbone network, then multi-scale features are extracted and fused through a multi-scale context information module, then features of different levels are combined with an attention fusion module, and finally bilinear interpolation up-sampling is directly utilized to obtain a final segmentation result. For an unlabeled image, firstly the unlabeled imageis segmented into small image blocks, the small image blocks are input into a semantic segmentation network to obtain a corresponding result, and then the small image blocks are spliced to obtain a final large-size segmentation result graph.

The invention discloses a human skeleton behavior recognition method and device based on deep reinforcement learning. The method comprises: uniform sampling is carried out on each video segment in a training set to obtain a video with a fixed frame number, thereby training a graphic convolutional neural network; after parameter fixation of the graphic convolutional neural network, an extraction frame network is trained by using the graphic convolutional neural network to obtain a representative frame meeting a preset condition; the graphic convolutional neural network is updated by using the representative frame meeting the preset condition; a target video is obtained and uniform sampling is carried out on the target video, so that a frame obtained by sampling is sent to the extraction frame network to obtain a key frame; and the key frame is sent to the updated graphic convolutional neural network to obtain a final type of the behavior. Therefore, the discriminability of the selectedframe is enhanced; redundant information is removed; the recognition performance is improved; and the calculation amount at the test phase is reduced. Besides, with full utilization of the topologicalrelationship of the human skeletons, the performance of the behavior recognition is improved.

The invention discloses a semi-supervised mechanical fault diagnosis method based on an adaptive migration neural network, and the method comprises the steps: firstly obtaining a plurality of source domain fault data sets composed of source domain fault training samples and corresponding tags, and a plurality of target domain fault data sets composed of target domain fault data without tags, wherein the target domain fault data is divided into a target domain fault training sample and target domain fault test data; normalizing the data; constructing an adaptive migration neural network diagnosis model, supervising the training model and constructing a classifier loss function by using the source domain fault data set, constructing a classifier discrimination loss function, and performing adversarial training on the feature extractor and the classifier by using the target domain fault training sample; inputting the target domain fault test data into the trained model, and summing and averaging the two output probability values to obtain a final classification diagnosis result. The method can improve the discrimination capability of the fault data of the target domain, and effectively improves the intelligent fault diagnosis task under the actual variable working condition.

The invention provides a behavior recognition method of depth supervised convolution neural network based on training feature fusion, belonging to the artificial intelligence computer vision field. This method extracts multi-layer convolution features of target video, designs local evolutionary pooling layer, and maps video convolution features to a vector containing time information by using local evolutionary pooling layer, thus extracts local evolutionary descriptors of target video. The local evolutionary descriptors of target video are extracted by using local evolutionary pooling layer.By using VLAD coding method, multiple local evolutionary descriptors are coded into meta-action based video level representations. Based on the complementarity of the information among the multiple levels of convolution network, the final classification results are obtained by integrating the results of the multiple levels of convolution network. The invention fully utilizes the time information to construct the video level representation, and effectively improves the accuracy of the video behavior recognition. At the same time, the performance of the whole network is improved by integrating the multi-level prediction results to improve the discriminability of the middle layer of the network.

The invention discloses a hyper-spectral image classification method based on recurrent neural network with the object to solving the problems that in prior art, the input characteristic determination ability is weak and that the extraction of local spatial characteristics is not complete. The method comprises the following steps: 1) extracting the spatial texture characteristics and the sparse representation characteristics of a hyper-spectral image and piling and combining them as the low-level characteristics; 2) extracting from the low-level characteristics the sample local spatial sequence characteristics; 3) according to the local spatial sequence characteristics, creating a recurrent neural network model; and utilizing the training sample local spatial sequence characteristics to train the recurrent neural network model parameters; and 4) inputting the testing sample local spatial sequence characteristics into the well-trained recurrent neural network model; obtaining the highly abstract high-level semantic characteristics and obtaining the classification information of the testing sample. According to the deep learning method of the invention, the correct efficiency for hyper-spectral image classification is increased and the method can be used for vegetation investigation, disaster monitoring, map making and intelligence obtaining.

The invention provides a target tracking method based on multi-characteristic adaptive fusion and kernelized correlation filtering technology. The method comprises steps of according to target position and the dimension of the previous frame tracking, acquiring a candidate region of target motion; extracting histogram characteristics and color characteristics in the gradient direction of the candidate region, fusing the two kinds of characteristics, carrying out Fourier transform so as to obtain a characteristic spectrum and then calculating kernelized correlation; determining the position and the dimension of the target at the current frame, and acquiring a target region; extracting histogram characteristics and color characteristics in the gradient direction of the target region, fusing the two kinds of characteristics, carrying out Fourier transform so as to obtain a characteristic spectrum and then calculating kernelized self-correlation; designing the adaptive target correlation and training a position filter model and a dimension filter model; and using a linear interpolation method to update the characteristic spectrums and the related filters. According to the invention, the discrimination capability of the models is improved; robustness of the target tracking of the target in a complex scene and the appearance change is improved; calculation complexity is reduced; and tracking timeliness is improved.

The invention discloses a cross-modal retrieval method based on adversarial learning and asymmetric hashing. The cross-modal retrieval method comprises the following steps: 1) selecting a cross-modaldata set and dividing the cross-modal data set into a training set and a test set; training a model, wherein the model comprises a loss module, a feature extraction unit and a Hash code learning unit,wherein the hash code learning unit comprises a hash code learning network and a shared multi-label binary matrix, and the loss module comprises the confrontation loss of the feature extraction unit,the ternary margin loss of the hash code learning unit, the cosine quantization loss and the asymmetric hash loss; 2) generating a binary hash code of each modal data in the cross-modal data set by using the trained optimal model; and 3) for a given query data, firstly generating a binary hash code, and then calculating a Hamming distance between the binary hash code of the query data and the binary hash code of the modal data in the cross-modal data set, which is different from the modal of the query data, so as to obtain a cross-modal data instance meeting the condition.

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 anti-fraud modeling modeling method and an anti-fraud monitoring method based on machine learning. The anti-fraud model modeling method based on machine learning comprises the following steps: extracting sample data required for modeling from a database, and carrying out labeling processing on each sample data; matching the association information of each sample data from the database, using the results of labeling processing to establish the multi-dimensional credit data based on the user, and processing and dividing the credit data into training set data and test set data; training and adjusting the parameters of the anti-fraud model by using the training set data; using the test set data to test the anti-fraud model, obtaining the fraud probability value thatthe test set data is fraudulent users. The obtained fraud probability value is compared with the corresponding actual sample situation, and the stability of the anti-fraud model is judged according tothe comparison result, and the anti-fraud statistical threshold value is established. The method can effectively reduce the risk of fraud through label processing and supervised machine learning.

The invention provides a cross-modal hash retrieval method based on triple deep networks. The method is used to solve the technical problem of low retrieval precision existing in existing cross-modalhash retrieval methods, and includes the realization steps of: preprocessing data, and dividing the data into training data and query data; acquiring hash codes of image training data and text training data; using triple supervisory information to establish an objective loss function; carrying out orderly iterative optimization on the objective loss function; calculating hash codes of image querydata and text query data; and acquiring retrieval results of the query data. According to the solution provided by the invention, the triple information is used to construct the objective loss function, semantic information is increased, an intra-modal loss function is added at the same time, discriminability of the method is improved, and precision of cross-modal retrieval can be effectively improved. The method can be used for Internet-of-things information retrieval and image and text mutual-searching services of e-commerce, mobile equipment and the like.

The invention provides an unsupervised field adaptation method and system based on adversarial learning and a medium, and the method comprises the steps: a feature extraction step: employing a featureextraction network to extract the features of images in a source field and a target field, and obtaining the image features of the source field and the image features of the target field; a categoryprediction step: predicting the probability that the image belongs to each category according to the obtained image feature of the source field and the image feature of the target field, and obtaininga category prediction probability; and a field discrimination step: predicting the probability that the image feature comes from the source field and the target field through a field discrimination network according to the obtained image feature of the source field and the image feature of the target field, and obtaining a field prediction probability. According to the method, the characteristicsthat the image extraction fields in the source field and the target field are invariable and have relatively high discrimination capability can be realized, so that the unsupervised field adaptationis realized.

The invention discloses a scene text recognition method based on sparse coding characteristics, and relates to computer vision and pattern recognition. The method includes the steps: inputting a natural scene text image to be recognized; by the aid of a multi-scale sliding window method, detecting and recognizing a window area in the image by a character classifier, for each character class, determining a large output area of the classifier as a candidate character area, determining a small output area as a background area, finding the candidate character area in the image, retaining the area with the largest output value of the classifier and the corresponding character class for the area with a large overlapping ratio by the aid of a non-maximum suppression method, and removing the repetitive and redundant candidate character area to obtain a character detection result; combining detected characters into a word or text line; outputting a scene text recognition result. Structural characteristics of the characters can be more effectively expressed and extracted, so that the recognition rate of a scene text is increased.

The invention discloses a facial expression identification method based on multi-task convolutional neural network. The expression identification method comprises the following steps: firstly, designing a multi-task convolutional neural network structure, and sequentially extracting low-level semantic features shared by all expressions and a plurality of single-expression distinguishing characteristics in the network; then adopting multi-task learning and simultaneously learning learning tasks of the plurality of single-expression distinguishing characteristics and multi-expression identification tasks; monitoring the all tasks of the network by using combined loss, and balancing the loss of the network by using the two loss weights; finally, acquiring a final facial expression identification result from a maximum flexible classification layer arranged at the last of a model according to the trained network model. Characteristic extraction and expression classification are put in an end-to-end framework to be learned, the distinguishing characteristics are extracted from input images, and expression identification on the input images are reliably carried out. Experimental analysisshows that the algorithm is excellent in performance, complicated facial expressions can be effectively distinguished, and good identification performance on a plurality of published data sets can beachieved.

The invention discloses a license plate character recognizing method of a support vector machine. The method comprises a training process of the support vector machine and a process of recognizing a license plate by use of the support vector machine. In the license plate character recognizing method disclosed by the invention, the character features are extracted by an LPP (locality preserving projections) algorithm, and the obtained character features can be better classified for judgment, thereby improving the judgment effect; and a character / number classifier is established by layers, and the characteristics of the license plate can be recognized in a small class range, thus the recognition resolution is improved, the complexity of the recognition algorithm and the training parameter storing resources are reduced, and application in an embedded system is suitable. As video stream statistical identification based on confidence is adopted, the shortcoming of bad recognition effect caused by a bad single-frame shooting effect or bad license plate preprocessing effect is avoided, and the recognition rate can be further improved.

The invention discloses a voice generation method based on a generative adversarial network. According to the method, randomly-generated noise data meeting Gaussian distribution is converted into a simulation sample through a generative model; as the simulation sample does not have the language content, when the generative model and a discrimination model are circularly updated, generative capacities required to be learned by the generative model and discrimination capacities required to be learned by the discrimination model are correspondingly increased, and accordingly the generative capacities of the generative model and the discrimination capacities of the discrimination model are improved; when a contrast value between a training sample and the simulation sample is smaller than or equal to a preset threshold value, it is thought that the generative model has the capacity of generating real data; a voice database generated through the generative model has enough reality, and the recognition rate can be increased when the generative model is applied to identity recognition. Correspondingly, the voice generation method, a voice generation device and voice generation equipment based on the generative adversarial network and a computer readable storage medium have the same advantages.

Provided is an inversion method for the biophysical property of a forest based on a LiDAR comprehensive waveform model. Crown segmentation is carried out, waveform information within the individual tree range is gathered and then gathered on a standing forest scale, and thus transformation of the inversion scale is more flexible and reasonable; the ground return pulse is determined by means of a digital landform and a height error threshold, so that the determination effect of return pulses of a crown layer is indirectly improved; and characteristic information of comprehensive waveforms is extracted in multiple dimensions, so that rich forest crown layer information is extracted from the waveform data in a deepened manner. The inversion method improves the precision in estimating the biophysical property of the forest, provides theoretical and method supports for explanation of the mechanism, and effectively avoids the defect that value of the waveform data is greatly reduced due to abandonment of rich continuous vertically-distributed information of the forest recorded in original waveform data.

The present invention provides a method for integrating crowdsourced annotations. According to the method provided by the present invention, a generalized anti-Gaussian distribution is defined by using a regularization hyperparameter, a spacing distance hyperparameter, an annotator voting weight, and a difference between the number of times that annotators annotate a current prediction item as a corresponding estimate and the number of times that annotators annotate the current prediction item as a subcategory; sampling is performed to obtain an auxiliary parameter; and annotator weights are updated by using the auxiliary parameter, thereby significantly enhancing a discrimination capability of a model. Then, a traditional annotation integration majority voting model and a confusion matrix model are integrated, thereby achieving an objective of more comprehensively describing a data generation process. In addition, an accurate prediction item update value is obtained through sampling, and moreover, running efficiency is also improved.

The invention relates to a method for segmenting a liver and a focus thereof in a medical image, and the method comprises the steps: firstly carrying out the screening and integration preprocessing ofabdominal CT image data, dividing the abdominal CT image data into a plurality of data sets with different purposes, building a new neural network, and carrying out the initial training through employing small image data; then, storing the trained model, carrying out secondary training by using the original image and a new data enhancement mode, carrying out expansion and corrosion processing onthe predicted image, and evaluating by using a medical evaluation index; through the model prediction results trained by DL, GDL and TL loss functions, adding and averaging the prediction results of the three loss models to form a fusion feature; finally, modifying the network, wherein the three loss models are fused in a single network for training prediction. End-to-end training test can be carried out, liver and focus can be identified at the same time with high precision and high speed, doctors are effectively helped to identify CT images, time and energy consumed by doctors are greatly reduced, and the probability of misdiagnosis is reduced.

The invention belongs to the technical field of computer vision, and discloses a target tracking method based on internal clipping and multi-layer feature information fusion, which comprises the following steps: acquiring a video sequence data set, and forming a training set according to the video sequence data set; constructing a twin network, wherein a basic backbone of the twin network adopts acombination of a ResNet18 feature extraction network improved by an internal clipping unit and an RPN network; based on the training set, training a twin network, and obtaining a training convergenttwin network model; and performing online tracking by using the twin network model. The method solves the problems that a deep learning tracking method in the prior art is poor in feature expression ability and model generalization ability, and a tracker cannot effectively cope with severe space-time change of target appearance.

The invention discloses a fine-grained image classification method and system, computer equipment and a storage medium. The fine-grained image classification method comprises the steps of: establishing a fine-grained image classification network which is a double-branch network for attention inhibition and attention enhancement and comprises a residual network and an attention layer; acquiring a training set, wherein the training set is composed of a plurality of training images; training the fine-grained image classification network by using the training set, and acquiring a fine-grained image classification model by using a gradient-propelled maximum value and minimum value cross entropy loss function; acquiring a to-be-classifiedimage; and inputting the to-be-classified image into the fine-grained image classification model, so that the to-be-classified image flows in the residual network and does not pass through the attention layer, and a category prediction result is obtained. The fine-grained image classification method and the system is implemented based on weak supervised learning and an attention mechanism, and the fine-grained image classification model obtained throughtraining can realize a good fine-grained image classification effect.

The invention relates to the field of computer visions, and proposes a face recognition method based on deep learning in a monitoring environment. The method comprises the following steps: building aface feature database, and performing feature extraction and feature comparison on faces in a video, thus face detection and face recognition are achieved based on an MTCNN network and a Sphereface network respectively; and by adding affine transformation, increasing the diversity of a data set, network clipping, network sparsification and layer-by-layer quantization, the problems of low recognition accuracy of a traditional face recognition method, differences in recognition accuracy between East and West faces, and excessive bandwidth required for the face recognition technology based on deep learning are mainly solved. Experiments show that the face recognition method proposed by the invention can be used in the monitoring environment and has higher recognition accuracy.

The invention discloses a method and device of detecting human vital signs based on ultra-wideband radar. The method comprises the steps of acquiring multiple first echo signals of normal human breathand / or heartbeat and multiple second echo signals of abnormal human breath and / or heartbeat through ultra-wideband radar, preprocessing the signals, and extracting to obtain a normal detection sequence and an abnormal detection sequence; setting a convolutional neural network; training the convolutional neural network based on the normal detection sequence and the abnormal detection sequence; detecting breath and / or heartbeat of a current detection subject through the trained convolutional neural network, and outputting the results of whether the heartbeat is normal. The device herein comprises a signal processing module, a convolutional neural network setting module, a training module and a detection module. The method and device have the advantages that deep learning helps improve the accuracy of detecting heartbeat frequency, the capacity to distinguish normal heartbeat and abnormal heartbeat is improved, and real-time detection is distinguished accurately.

The invention belongs to the technical field of computed tomography image processing. According to the specific technical scheme, the CT image super-resolution reconstruction method based on the generative adversarial network comprises the following specific steps: 1, establishing a dense connection relationship among different residual blocks based on a multi-stage dense residual block generatornetwork; 2, adding a bottleneck layer to the front end of each dense residual block; 3, optimizing the global network by adopting the Wasserstein distance loss and the VGG feature matching loss; 4, arranging a multi-path generator based on the sequence from thick to thin; 5, generating an image based on conditional expression generative adversarial learning; 6, reconstructing a CT image super-resolution reconstruction framework of the generative adversarial network based on multiple paths of conditions from coarse to fine; 7, reconstructing a loss function. According to the method, network redundancy is reduced, feature multiplexing among different residual blocks is realized, the maximum information transmission of the network is realized, the feature utilization rate is improved, and thereconstructed image quality is greatly improved.

The invention discloses a correlation filtering video adaptive tracking method combined artificially with depth features, which comprises the following steps: pre-training a depth neural network; Depth convolution feature extraction; Manual feature extraction; Multi-feature fusion computation; Scale Adaptive Target Detection and Response Computation; Correlation filter model updating; Output current frame tracking results. As that depth convolution feature are adopted, The gradient direction histogram and color name feature are fused to model the appearance of the object, the background response is suppressed by the spatial context information of the object, the discriminant ability of the model is enhanced by the adaptive scale estimation method, and the computational complexity is reduced by the fast Fourier transform in the frequency domain. The invention improves the robustness of tracking under complex scenes, and can be widely applied to the fields of video surveillance, human-computer interaction, robot technology, road scene understanding and the like.

The invention discloses an intelligent driving local track fault tolerance planning method based on lane lines and GPS following. The method includes steps: firstly initializing a following mode and establishing an intelligent driving vehicle coordinate system; secondly discriminating the recognition states of GPS data and the lane lines according to GPS and lane line information; then calculating a fault tolerance deviation according to the discriminated recognition states, and updating the following mode; and finally performing local path and track planning based on the new following mode. According to the method, the validity of various data is determined, the accuracy of subsequent calculation is improved, the fault tolerance deviation is designed based on various data states, the fault tolerance deviation is dynamically updated in real time, the system complexity is simplified, the practical application is easy, and the robustness of data processing is improved; and real-time state transition of the following mode is performed according to multi-sensing data including the GPS and the lane lines, continuous and smooth control of multiple following states is realized, and the comfort and the stability of intelligent-driving vehicles are improved.

The invention discloses a large-scale image retrieval method. The method comprises the steps of image feature extraction, Hash function projection vector learning, Hash function offset learning, image feature dimensionality reduction, image feature encoding and image retrieval. By adopting the method, a large scale of images can be retrieved rapidly. Firstly, the discrimination among codes is enhanced by learning a discriminant Hash function, thereby better distinguishing different types of image features; secondly, the image features are subjected to dimensionality reduction and encoding by using a Hash function, thereby lowering the storage demand of the image features and the computation overhead of a retrieving process. By adopting the large-scale image retrieval method, large-scale image retrieving is realized efficiently and accurately, thereby achieving a high application valve.

The invention discloses a similarity-weight-semi-supervised-dictionary-learning-based human behavior identification method. With the method, a problem of low human behavior identification rate of the existing supervision method in the prior art can be solved. The identification method comprises: (1), an inputted data set is divided into test samples and training samples; (2), local feature detection is carried out on all samples and local features with the labeled samples are selected randomly to obtain an initialized dictionary; (3), according to the initialized dictionary, dictionary learning is carried out by using a semi-supervised method; (4), group sparse coding is carried out on all samples by using the learned dictionary to obtain a coding matrix of each sample; (5), vectorization is carried out on the coding matrix of each sample to obtain a final expression; and (6), testing sample classification is carried out by using the final expression of each sample and a sparse representation classification method to complete human behavior identification in the testing samples. Therefore, discrimination of dictionary learning is enhanced; the human behavior identification rate is improved; and the method can be used for target detection in a video.

The invention discloses a supervised fast discrete multi-modal hash retrieval method and system. The method includes receiving a multi-modal training data set, wherein each sample contains a pair of multi-modal data features; projecting the multi-modal training dataset to a joint multi-modal intermediate representation by using a joint multi-modal feature map; for the joint multimodal intermediaterepresentation of multimodal training datasets, constructing a supervised fast discrete multimodal hash objective function; solving the objective function to obtain a hash function; receiving multimodal retrieval data set and multimodal test data set, projecting samples into joint multimodal middle representation, and then projecting them into Hamming space to obtain hash code according to hash function; based on hash codes, retrieving samples from multimodal test datasets in multimodal retrieval datasets. The invention learns discrete hash codes for heterogeneous multi-modal data, and ensures learning efficiency and retrieval precision at the same time.