33 results about "Discriminative learning" patented technology

The invention discloses an image denoising method. The method comprises steps: an average block set of a to-be-denoised image is acquired; a corresponding mean value is subtracted from each average block in the average block set to obtain a noise block set; a first noise image in the noise block set is acquired; noise modeling is carried out on the noise block set according to a generative adversarial network to obtain a generator capable of generating the same kind of noise as the to-be-denoised image; a second noise image is acquired according to the generator; according to a noise-free image, the first noise image and the second noise image, a training set is constructed; according to the training set and a discriminative learning method, an image denoising network model is trained; andthe to-be-denoised image is inputted to the image denoising network model to acquire an image after denoising. The invention also provides an image denoising device, equipment and a storage medium. the denoising effects on unknown real noise in real life can be enhanced and the denoising efficiency is improved.

The invention discloses a target tracking method based on sparse discrimination learning. The method comprises the steps that a target appearance model is constructed for a first frame sampling target and the surrounding background; the two-dimensional image features of the target appearance model are extracted and are normalized to acquire an initial dictionary; a supervised discriminant dictionary learning method is introduced; a classification error term is added on the basis of a reconstruction error term to train a judgment dictionary; the minimum reconstruction error term is solved under sparsity constraint; an iteration switching optimization policy is used to update dictionary and sparse coding; Euclidean distance is used to measure the similarity between samples; and the most similar samples are used as a tracking target. According to the invention, when the target appearance model is established, the surrounding background in spatial correlation with the target is added as a clue template, and the processing of the target attitude change is robust.

The invention discloses a robot imitation learning method based on virtual scene training. The method comprises the following steps: designing a robot model and a virtual interaction environment according to a specific task; collecting and arranging an expert data set; determining a state value space S and an action value space A according to the specific task, and determining structures of the network of a strategy generator and the network of a discriminator according to the state value space S and the action value space A; sampling data from the strategy generator, designing a parameter updating strategy, and alternately training the strategy generator and the discriminator by combining the expert data set and adopting an adversarial training method until the discriminator converges toa saddle point; and testing a network model composed of the strategy generator and the discriminator obtained by training, and taking a real environment state as input of the strategy generator so asto obtain action output. According to the method, a value return function is judged and learned; a large number of complex intermediate steps of inverse reinforcement learning with high calculation amount are bypassed; and the learning process is simpler and more efficient.

The present invention discloses a real-time target tracking method based on multi-feature discriminative learning. The method comprises the steps of: 1) acquiring a gray-scale video frame in a video,and describing a brightness attribute of a tracking target by using a Cross-bin distribution field feature; 2) using the enhanced gradient histogram feature EHOG to carry out modeling on texture diversity of the tracking target; 3) extracting the color feature CN to maintain color consistency through a color video frame in the video; 4) projecting dimensional features obtained in the steps 1), 2),and 3) into the high dimensional feature space through the Hilbert space mapping to obtain the inner product mapping; and 5) placing an obtained confidence map in the CSK framework to track, findingout a tracking target location, and then updating the template to carry out target tracking. The method disclosed by the present invention can effectively solve the problems such as light change, background interference, occlusion, low real-time performance and the like existing in target tracking.

A system for providing automatic diagnosis and decision support includes: a medical image database; generative learning and modeling modules that build distributional appearance models and spatial relational models of organs or structures using images from the medical image database; a statistical whole-body atlas that includes one or more distributional appearance models and spatial relational models of organs or structure, in one or more whole-body imaging modalities, built by the generative learning and modeling modules; and discriminative learning and modeling modules that build two-class or multi-class classifiers for performing at least one of organ, structure or disease detection or segmentation.

A tracking and recognition system is provided. The system includes a computer vision-based identity recognition system configured to recognize one or more persons, without a priori knowledge of the respective persons, via an online discriminative learning of appearance signature models of the respective persons. The computer vision-based identity recognition system includes a memory physically encoding one or more routines, which when executed, cause the performance of constructing pairwise constraints between the unlabeled tracking samples. The computer vision-based identity recognition system also includes a processor configured to receive unlabeled tracking samples collected from one or more person trackers and to execute the routines stored in the memory via one or more algorithms to construct the pairwise constraints between the unlabeled tracking samples.

The invention provides a face recognition method and a face recognition system based on two-dimensional discriminant features. Discrimination learning is performed by using face training images; local intra-class divergence is tightened; and local inter-class divergence is separated. In order to effectively keep the topological structures and the intrinsic relevance between image pixels, a discriminant feature extracting system based on matrix description is designed, can be directly applied to a face image without destroying topological structures and intrinsic relevance between image pixels, thereby improving system performance. Induction of non-sample images is that a testing sample is mapped to an orthogonal projection output by the system so as to input extracted face image two-dimensional features to a nearest neighbor classifier for classification; a label of a training sample that has the greatest similarity or minimum distance to the testing sample is taken and used for distinguishing the classification of the face testing image, so as to obtain a most accurate face recognition result. In addition, according to the face recognition method and system provided by the invention, the feature extraction and classification can be directly carried out on the face image, so that the system efficiency is effectively improved, and the system expansibility is good.

The invention discloses a method for semantically annotating images on the basis of hybrid generative and discriminative learning models. The method includes generatively building models of the images by means of continuous PLSA (probabilistic latent semantic analysis) at generative learning stages, acquiring corresponding model parameters and subject distribution of each image, and utilizing the corresponding subject distribution as an intermediate representation vector of each image; constructing cluster classifier chains to discriminatively learn from the intermediate representation vectors of the images at discriminative learning stages, creating the classifier chains and integrating contextual information among annotation keywords; automatically extracting visual features of each given unknown image at annotation stages, acquiring representation of subject vectors of the given unknown images by the aid of a continuous PLSA parameter estimation algorithm, classifying the subject vectors by the aid of trained cluster classifier chains and semantically annotating the images by a plurality of semantic keywords with the highest confidence. The method has the advantage that the annotation and retrieval performance of the method are superior to the annotation and retrieval performance of most current typical methods for automatically annotating images.

The invention discloses a method for migrating a generative adversarial network by using adversarial learning and discriminative learning. The method comprises the following steps: S1, preparing a picture data set; s2, constructing a pre-trained GAN model; s3, constructing an ADT-GAN model through parameter migration; and S4, training the ADT-GANc. The ADT-GAN model initializes a generator and a discriminator through parameter transfer on the basis of a pre-trained GAN model trained by a source domain image data set by using transfer learning. A domain discriminator is added, a total objective function composed of an adversarial objective function and a domain discrimination objective function is optimized to drive a generator to generate image data of a target domain, and negative migration is avoided. Therefore, the training performance on a small target domain data set is improved, the number of iterations is reduced, and the image generation quality is improved.

Embodiments of the invention relate to a novel approach for determining the palatability, or preference, of foods, food stuffs, or veterinary biologics in animals. This approach provides a robust and reliable means of assessing palatability in animals using a cognitive palatability assessment protocol. The basic protocol utilizes a discrimination learning procedure, in which animals are presented with three simultaneous stimuli, such as objects, odors, sounds, digital pictures or projections, etc., and are allowed to respond to one, although any number of stimuli may be employed. The response is then recorded including metadata such as the rate of learning, response rate and response order for example. Depending on the animal's choice, the response will typically result in the animal receiving no reward, or either of two particular foods, food stuffs or veterinary biologics. The animal's response selection is indicative of a preference for one food, food stuff or veterinary biologic over the other and is indicative of the palatability of one food, food stuff or veterinary biologic in comparison to the other. This method requires the use of fewer animals than alternative approaches and may prove healthier for animals that do not self-limit their food consumption.

The invention discloses a feature extraction method for one-dimensional distance image of true and false targets with non-linear discriminant learning, and belongs to the technical field of radar target recognition. The invention first uses a non-linear function to map the one-dimensional distance image to a high-dimensional feature space, then obtains a projection transformation matrix through discriminant learning in the high-dimensional feature space, and then obtains feature vector of the one-dimensional distance image of any feature to be extracted based on the projection transformation matrix. The feature extraction method of the invention can well represent the non-linearity in the distribution of sample data, thereby improving the target recognition performance, overcoming the disadvantage that the existing conventional subspace method is not suitable for the non-linear data distribution, and effectively improving classification performance of radar true and false targets.

The invention discloses an intelligent business English polyseme discrimination learning device which comprises a first supporting plate, a second supporting plate and a back plate. The back plate comprises a first folding plate and a second folding plate, wherein the first folding plate and the second folding plate are hinged through a second rotating shaft. The second folding plate is hinged with the second supporting plate through a third rotating shaft. The first folding plate is hinged with the first supporting plate through a first rotating shaft. A flexible display screen is arranged above the front of the back plate. A power supply module and a processor chip are arranged inside the back plate. The processor chip is provided with a noise removing module, a speech recognition module, an information processing module and a wireless communication module, wherein the noise removing module, the speech recognition module, the information processing module and the wireless communication module are respectively electrically connected with the processor chip. A keyboard which is electrically connected with the processor chip is arranged on the right side of the first supporting plate. A raised platform is movably arranged on the front side of the second supporting plate. A concave table on the raised platform matches a mounting hole provided in the front side of the back plate. The height of the learning device can be properly adjusted. The display screen can be folded. The entire stored device takes up less space and is easy to carry.

The invention discloses an intelligent business English polyseme discrimination learning device which comprises a first supporting plate, a second supporting plate and a back plate. The back plate comprises a first folding plate and a second folding plate, wherein the first folding plate and the second folding plate are hinged through a second rotating shaft. The second folding plate is hinged with the second supporting plate through a third rotating shaft. The first folding plate is hinged with the first supporting plate through a first rotating shaft. A flexible display screen is arranged above the front of the back plate. A power supply module and a processor chip are arranged inside the back plate. The processor chip is provided with a noise removing module, a speech recognition module, an information processing module and a wireless communication module, wherein the noise removing module, the speech recognition module, the information processing module and the wireless communication module are respectively electrically connected with the processor chip. A keyboard which is electrically connected with the processor chip is arranged on the right side of the first supporting plate. A raised platform is movably arranged on the front side of the second supporting plate. A concave table on the raised platform matches a mounting hole provided in the front side of the back plate. The height of the learning device can be properly adjusted. The display screen can be folded. The entire stored device takes up less space and is easy to carry.

The invention provides a method for monitoring the flow state of a supersonic air inlet channel, which adopts a method of combining a time-frequency analysis technology and deep learning to monitor the flow state of the supersonic air inlet channel from a dynamic sensor signal. The method comprises the following steps: performing preliminary signal processing on a dynamic sensor signal by using continuous wavelet transform, converting the dynamic sensor signal into a time-frequency spectrogram, and then inputting the time-frequency spectrogram into a convolutional neural network (CNN) for classification. In order to reduce the classification error, the invention provides a DDL-CNN / TDL-CNN (Doublet / Triplet ConvolutionalNeural Network combined with Discriminative Learning) which simultaneously considers the loss of the cross entropy and the learning of the discriminative features. The proposed method encourages the convolution module to project the time-frequency spectrogram to a feature space, so that different flow states in the space become more separable. Experimental results show that compared with a traditional CNN, the DDL-CNN / TDL-CNN has better performance on multiple indexes.

The invention provides a moving target tracking method and system. The method comprises the steps of determining a plurality of target candidates on a current frame; determining a preset quantity of recently located targets as target samples, and performing sampling near a newest located target to obtain a series of background samples; according to the target samples and the background samples, constructing subspaces and performing training to obtain a linear classifier; reconstructing the target candidates through the subspaces, and obtaining classification reliability of the reconstructed target candidates by using the linear classifier, wherein the classification reliability serves as a factor of judging the possibilities that the target candidates become the targets; and determining the target candidate with the maximum possibility as the target. The advantages of subspace learning and discriminative learning methods are combined, so that the performance in multiple complicated tracking scenes can be improved.

The invention discloses a pedestrian re-identification method and device based on cross-scene continuous learning of consistency learning. The method comprises the steps of obtaining original features of data of each batch; inputting the original features into a pseudo task data conversion module, and respectively obtaining an old scene feature and a new scene feature corresponding to each original feature after scene style conversion; calculating intra-domain and inter-domain cross-scene consistency loss functions on old scene features, and performing pseudo task identity discrimination learning; calculating the pairwise similarity of the old scene feature sample and the new scene feature sample, and performing pseudo task knowledge distillation; after the new scene features are input into a classifier of the new scene, a cross entropy loss function is calculated, and identity discrimination learning is carried out; and calculating the distance between the old scene feature and the new scene feature corresponding to each sample, and carrying out cross-scene consistency learning. According to the method, the continuously updated and iterated models can be deployed in a plurality of successively learned scenes, so that the purpose of reducing the cost of model training and manual maintenance is achieved.

The invention relates to artificial intelligence, and provides a code scanning information detection method and device, equipment, and a storage medium. The method comprises the following steps: iteratively adjusting a generation learner and a discrimination learner based on code scanning samples, and obtaining a generation network and a discrimination network; obtaining to-be-analyzed code scanning samples during iteration adjustment; extracting target code scanning samples of the last round of the iteration adjustment; processing the target code scanning samples based on the generation network to obtain generation code scanning samples, and extracting samples to obtain training code scanning samples; generating discrimination accuracy according to a prediction result and an annotation result of code scanning data; if the discrimination accuracy is smaller than a preset threshold value, extracting to-be-increased samples according to the annotation result and the prediction result; adjusting network parameters based on the generation network and the to-be-increased samples to obtain a discrimination model; processing to-be-detected information based on the discrimination model to obtain a detection result. The accuracy of detection results can be improved; in addition, the invention also relates to blockchain technology, so that the detection results can be stored in a blockchain.

The invention discloses a diversity discrimination learning-based breast tumor recognition method. According to the method, a diversity classification learning model which is used for classifying diversity states of tumors is firstly created; a discriminative learning model which is used for classifying tumors greatly different from normal tumors is created; and a tumor is recognized through the diversity classification learning model and the discriminative learning model. Compared with the prior art, the diversity discrimination learning-based breast tumor recognition method is capable of effectively classifying diversity states of tumors so as to decrease influences, on performance, of diversities; normal tumors can be recognized by utilizing a traditional classification model; and abnormal tumors can be input into the discriminative learning model to be recognized, so that breast tumor can be correctly recognized and significance is provided for the early-stage auxiliary diagnosis of the breast tumor.

The invention belongs to the technical field of computer vision, and provides a weak supervision fine-grained image classification algorithm based on correlation-guided discriminant learning. An end-to-end correlation guided discriminant learning model is provided to fully mine and utilize the correlation of weak supervision fine-grained image classification to improve discrimination. First, a discriminative region packet sub-network is proposed, which first establishes correlations between regions and then enhances each region by weighted summary of all correlations from other regions to guide the network to discover more discriminative region groups. And finally, proposing a discriminative feature enhancement sub-network to mine and learn the internal spatial correlation between the feature vector elements of each patch, and improving the local discrimination capability of the patch by jointly enhancing the information elements and suppressing useless elements at the same time. A large number of experiments prove the effectiveness of DRG and DFS, and the most advanced performance is achieved.

The invention relates to the technical field of English learning devices, and discloses an intelligent business English polysemy discrimination learning device which comprises a base box, wherein an anti-skid pad is fixedly mounted at the bottom of the base box, and a first mounting groove and a second mounting groove are separately formed in the base box; a first connecting rod is fixedly mounted in the first mounting groove, a first compression spring and a first sliding block are movably mounted on the outer side of the first connecting rod, a first rotating rod is movably mounted in the top end of the first sliding block, a motor in a first pin shaft block is controlled to rotate, when the motor rotates, a second rotating shaft is driven to rotate through an output shaft of the motor, when the second rotating shaft rotates, a transmission belt is driven to rotate through a first fixed tooth block, and when the transmission belt rotates, a third rotating shaft is driven to rotate through a second fixed tooth block, and when the third rotating shaft rotates, the inclination angle effect of a display screen can be changed through the connecting block.

The invention discloses an image recognition method and device based on robust feature extraction. By performing discriminative learning on original training samples, the topology between image pixels can be effectively maintained while compacting local intra-class divergence and separating local inter-class divergence. structure, design a feature learning framework based on the 1-norm metric, which can output a projection matrix that can be used for in-sample and out-of-sample image feature extraction, and then obtain the two-dimensional robust features of the original training samples through the projection matrix, and construct a model that can be used to treat A nearest neighbor classifier for classifying test samples. It can be seen that the above scheme provided in this application does not need to convert the two-dimensional matrix corresponding to the image into a high-dimensional vector space, but can directly act on the image, thus not only effectively maintaining the topological structure and intrinsic correlation between image pixels , and can effectively reduce the complexity of the model calculation process. Furthermore, based on the 1-norm metric, robustness to noise is ensured during feature extraction.

The invention belongs to the technical field of computer vision, and provides a fine-grained image classification method based on discriminant learning. We propose a new end-to-end autoregressive localization with discriminative prior network model that learns to explore more accurate discriminative patch sizes and is able to classify images in real-time. Specifically, a multi-task discriminative learning network is designed, including an autoregressive localization sub-network and a discriminative prior sub-network with a bootstrap loss function and a consistency loss function to simultaneously learn the self- Regression coefficients and discriminative prior map. The autoregressive coefficients can reduce the noise information in the discriminative patch, and the discriminative prior map filters thousands of candidate patches into single-digit patches by learning the discriminative probability value. Extensive experiments show that the proposed SDN model achieves the state-of-the-art in both accuracy and efficiency.

The invention discloses an image anomaly detection and positioning method based on discriminant learning, and the method comprises the following steps: carrying out the feature extraction of an original image through a network trained based on an external data set; carrying out feature selection by adopting a gradient preference method; and sending the obtained features into a discrimination network with center constraint to carry out anomaly detection. According to the method and the device, the abnormity existing in the image can be accurately detected and positioned under the condition that abnormal data is not needed. The method has strong performance and theoretical advantages in image anomaly detection and positioning. And the detection efficiency and accuracy are greatly improved.

The embodiment of the present invention provides a vehicle re-identification and device based on deep joint discriminant learning. The method includes: building a system network architecture based on a deep convolutional neural network; performing an overall target loss function on the system network architecture through a video vehicle database Training, determining the parameters of the system network architecture; inputting the vehicle image to be recognized and the vehicle image within the effective geographical area into the system network architecture after the parameters are determined, extracting the feature vector of the image to be recognized and the The feature vector of the vehicle image within the effective geographical area, and compare the feature vector of the image to be recognized with the feature vector of the vehicle image within the effective geographical area, and search for the vehicle image within the effective geographical area A vehicle image having the highest similarity with the vehicle image to be recognized among the vehicle images. Through the embodiments of the present invention, the efficiency and accuracy of vehicle re-identification can be improved.

The invention discloses a target tracking method based on sparse discriminant learning, which includes: sampling the target and its surrounding background in the first frame to construct a target appearance model; extracting the two-dimensional image features of the target appearance model and performing normalization processing to obtain Initial dictionary; introduce a supervised discriminant dictionary learning method, increase the classification error item on the basis of the reconstruction error item, and train the discriminant dictionary; solve the minimum reconstruction error item under the sparsity constraint, and use the iterative exchange optimization strategy to update the dictionary and Sparse coding; Euclidean distance is used to measure the similarity between samples, and the sample with the highest similarity is used as the tracking target. Through the embodiment of the present invention, when establishing the appearance model of the target, the surrounding background that has a spatial correlation with the target is added as a clue template, and it is more robust to deal with the change of the target posture.