198results about How to "Improve generation effect" patented technology

The invention discloses a font style migration method based on a conditional cycle consistency generative adversarial network, and the method specifically comprises the steps: building a source font data set and a target font data set, and carrying out the normalization processing along with the data sets; building a font style migration network structure, wherein a network is composed of a generator based on a conditional generative adversarial network and a residual network and a discrimination network based on PatchGAN; inputting source fonts and label pictures into the generator, and generating target style fonts; enabling the discriminator network to perform authenticity discrimination on the generated target font and the real target font corresponding to the generated target font, optimizing the Chinese character migration network by means of the idea of adversarial network training to achieve mutual mapping migration of two font styles, reduce manual interference in the font generation process, and improve the automation degree of font generation.

The invention discloses an image super-resolution reconstruction method for generating antagonistic network based on feature fusion, which comprises the following steps: pre-processing ImageNet data set to obtain a reconstruction data set corresponding to high-resolution image and low-resolution image; The model of generating antagonistic network for training is constructed, and the interpolationreconstruction module is introduced into the model to construct the generating network and perceptual network for multi-feature fusion. The reconstructed data sets are sequentially input into the generated antagonistic network for model training. The images to be processed are normalized to obtain low-resolution images, which are input to the trained generation network to obtain the reconstructedhigh-resolution images. As that recursive residual network is utilize to extract multiple features such as edge and texture, highlight the edge and texture information of the generated image, so thatthat reconstruct image is clearer, The interpolation reconstruction module is used as a mediator to realize the residual discrimination, highlight the difference between the reconstructed and real high-resolution images and reduce the computational complexity of the discrimination network, and use VGG network to fuse multiple features to calculate the loss function, which makes the reconstructed image more effective and more consistent with the observation mode of human eyes.

Enhancement of the supercontinuum generation performance of a highly-nonlinear optical fiber (HNLF) is accomplished by performing at least one post-processing treatment on the HNLF. Particularly, UV exposure of the HNLF will modify its dispersion and effective area characteristics so as to increase its supercontinuum bandwidth, without resorting to techniques such as tapering or introducing unwanted reflections into the HNLF. The UV exposure can be uniform, slowly varying or aperiodic along the length of the HNLF, where the radiation will modify the nonlinear properties of the HNLF. Various other methods of altering these properties may be used. The output from the HNLF can be monitored and used to control the post-processing operation in order to achieve a set of desired features in the enhanced supercontinuum spectrum.

The invention discloses a visual dialogue generation method based on a context perceptual map neural network. The visual dialogue generation method comprises the following steps of 1, preprocessing the text input in a visual dialogue and constructing a word list; 2, extracting the features of a dialogue image and the features of a dialogue text; 3, obtaining a context feature vector of the historical dialogue; 4, constructing a context perceptual map; 5, iteratively updating the context perceptual map; 6, carrying out attention processing on the nodes of the context perceptual map based on a current problem; 7, performing multi-modal semantic fusion and decoding to generate an answer feature sequence; 8, generating the parameter optimization of a network model based on the visual dialogueof the context perceptual map neural network; 9, generating a prediction answer. According to the method, the context perceptual map neural network is constructed on the visual dialogue, and the implicit relationship between different objects in the image can be reasoned by using the text semantic information with finer granularity, so that the reasonability and accuracy of the answers generated by an intelligent agent for question prediction are improved.

The invention provides a deep adversarial diagnosis method for a fan bearing fault under a non-equilibrium small sample scene. The deep adversarial diagnosis method is characterized by comprising thefollowing steps of acquiring wind turbine bearing vibration signals, building an improved AC-GAN (Generative Adversarial Networks) model, building an improved AC-GAN sample, generating a wind turbinebearing vibration signal sample, and diagnosing wind turbine bearing faults under various scenes. The deep adversarial diagnosis method solves the problems of complex vibration signal noise interferences, fewer fault samples, and unbalanced number of samples between categories in the fan fault diagnosis based on the vibration signals, improves the fault identification accuracy under a small samplenon-equilibrium scene, has good fault identification accuracy under complex scenes such as high noise interferences, insufficient number of samples and nonequilibrium of different types of sample training set scales, has the advantages of being scientific and reasonable, strong in adaptability, high in practical value, and can provide references for fan research and development, wind farm operation and maintenance, wind turbine research and other related personnel.

The invention provides a Chinese character font migration system based on adversarial networks. The system includes: a Chinese character migration module, which maps source fonts into generation target fonts, of which contents are the same as contents of the source fonts but the fonts are different, through a fully convolutional neural network, and a discriminator module, which utilizes a deep neural network-based discriminator network to carry out true / false font discrimination on the generation target fonts and real target fonts, namely gold standards, corresponding thereto, and constantly optimizes a Chinese character migration network by virtue of thinking of adversarial network training to enable the same to output more realistic generation target fonts. On the basis of the adversarial networks, the system can migrate a Chinese character of any printed font to another printed font, or even to a font corresponding to a handwriting font of a certain person.

The invention relates to a mutagenized strain of streptomyces albus TUST2, and a method for using the same to produce epsilon-polylysine and a salt of the epsilon-epsilon. The strain is a high-yield mutagenized strain of epsilon-polylysine selected and bred by means of ultraviolet mutagenesis, ultraviolet-chemical mutagenesis, N ion injection mutagenesis, etc from a strain of streptomyces albus TUST2 which is capable of producing epsilon-polylysine and is screened from the earth in Hainan province of China; the mutagenized strain is resistant to 10mg / ml or above S-(2 Aminoethyl )-L-Cysteine and can generate 10 to 30 g / L acid under optimal conditions; zymotic fluid is removed of the streptomyces albus by configuration and is filtered by diatomaceous earth to obtain clear filtrate and to weak acid type ion exchange resin to obtain coarse epsilon-polylysine products which is subjected to ultrafiltration and purification to get epsilon-polylysine and the salt of the epsilon-epsilon. The distribution of molecular weight of epsilon-polylysine is between 4000 and 6500Da.

The invention provides a generation method and device for a word segmentation dictionary and a word segmentation processing method and device. The generation method for the word segmentation dictionary comprises the following steps: acquiring an original sentence and language material; segmenting the original sentence and language material to obtain segments, and filtering the segments to obtain a filter result, wherein the filtering step comprises at least one of the following steps: carrying out filtering based on word frequency and inverse frequency; carrying out filtering based on boundary; and carrying out filtering based on splicing; and generating the word segmentation dictionary according to the filter result. The method, which is independent of manual screening and a word segmentation apparatus, can be used for recognizing entries which are not logged in, so that the generation speed and effect of the word segmentation dictionary are improved.

The invention discloses a satellite image super-resolution method combining an adversarial network with an aerial image a priori. Firstly, an image pair composed of a 16-level noisy image and a corresponding 16-level non-noisy image is used for training a denoising model, and then the image super-resolution model is trained by using clear aerial data. Because there is no satellite image and aerialimage pairs, clear aerial images are used to construct an external priori dictionary of GMM model in the post-processing of the generated super-resolution images, and then the internal unclear satellite images are guided to be reconstructed. In order to further improve the image quality, Gaussian filter is used to sharpen the image after reconstruction. Finally, the high-resolution image of the original satellite image is obtained, and the visual quality of the image is improved based on the original satellite image. The effectiveness of the scheme can also be seen from the experiment. It provides an effective way to solve the problem of satellite image super-resolution and image quality improvement under the condition of conditional constraints in reality.

The invention provides a probe bench, and an organic thin-film device preparation and test integrated system and method. An organic thin-film device preparation apparatus (200) and a probe bench (100) are connected in a sealing mode, and an externally sealed sample transmission channel is formed between the two, a sample is transmitted between the organic thin-film device preparation apparatus (200) and the probe bench (100) via the sample transmission channel, and the whole course of a sample transmission process is isolated from the effects of external environmental atmosphere, so that the generation and test of an organic thin-film device are both isolated from external environment and are disposed in an enclosed and vacuum environment, thus the test accuracy is effectively improved.

The invention provides a priori semantic knowledge-guided high-resolution remote sensing image weak supervision building change detection method, which is used for carrying out building automatic change detection by utilizing a high-resolution remote sensing image, and comprises the following steps of: firstly, generating priori semantic knowledge by adopting a building extraction algorithm fusing domain self-adaption and a weak supervision strategy, improving the cross-domain expansion capability of priori knowledge to the maximum extent by fully utilizing two strategies of domain self-adaption and weak supervision, and reducing the number and difficulty of sample data set manufacturing; and then, designing a high-resolution remote sensing image weak supervision building change detection network guided by prior semantic knowledge, taking intermediate results of each stage of the building extraction network as prior knowledge, reducing the dependence of the network on change detection sample data to the greatest extent, and improving the effect of building change detection.

The invention discloses an image scene map generation method based on iteration refinement, and relates to the field of image scene maps. The method comprises the following steps of: inputting an image scene map data set and parameters thereof, and extracting visual features of an image; in combination with the image data and the scene map data in the training set, classifying the visual relationship between the entity pairs by using the visual information and the semantic information to generate an image scene map; generating a plurality of region bounding boxes and corresponding feature vectors of each region, and fusing the region feature vectors to obtain feature vectors of the image description text segments; inputting the feature vectors of the image description text segments into animage description text segment generator to generate image description text segments; constructing a refined graph, defining feature vectors and regional feature vectors of refined entity pairs, anditeratively executing image scene graph generation, image segmented image description file generation and refined feature vector until the maximum number of iterations is reached. According to the invention, the image scene map generation effect is improved.

The invention discloses an image super-resolution reconstruction method of a generative adversarial network based on Gaussian coding feedback, and the method comprises the steps: carrying out the preprocessing of an ImageNet data set, and manufacturing a reconstruction data set in one-to-one correspondence with a low-resolution image and a high-resolution image; Constructing a generative adversarial network model for training, and introducing a Gaussian coding feedback network into the model; Sequentially inputting the data sets obtained in the step A into a generative adversarial network formodel training; And inputting the low-resolution image to be processed into a trained generative adversarial network to obtain a high-resolution image. A generative adversarial network is formed by constructing a generation network and a discrimination network, a Gaussian coding feedback loop is added between the generation network and the discrimination network, more information is added to the generation network to guide the generation network to carry out training, and important features are added by improving a sub-pixel convolutional layer structure, so that useless information is reduced, and the reconstruction effect is improved.

The invention discloses a network intrusion detection method based on a conditional variation auto-encoder, belongs to the technical field of network space intrusion detection, and the method comprises the steps: firstly training a conditional variation auto-encoder which takes a logarithmic hyperbolic cosine function as a loss function for the expansion of a data set, and then training a classifier based on the convolutional neural network through the data set after data expansion to serve as an intrusion detector of the whole model. According to the method, the classification detection performance on a network space intrusion data set is improved by utilizing the generation capacity of conditional variation self-encoding and the excellent feature extraction capacity of the convolutionalnetwork for data features, and meanwhile, the problem of performance reduction caused by imbalance of data set samples is also solved.

The invention provides an automatic text generation method and device and a computer readable storage medium. The method comprises: training a Transformer encoding-decoding deep learning model, and carrying out text encoding and knowledge graph encoding on topic words to obtain corresponding hidden state representations, wherein a plurality of knowledge graph triples are added to each topic word in the knowledge graph encoding process; decoding according to the text code and the hidden state representation of the triple code of the knowledge graph to obtain a generated text; and automaticallygenerating a new text by utilizing the trained deep learning model according to the newly input topic words. The device is used for implementing method. By using a Transformer encoding-decoding deep learning model, the relationships between texts, between knowledge graph triples, and between texts and triples are fully learned by using Tansformers, and more abundant generation results are generated.

Various systems and methods are described for generating vacuum within an engine intake. An intake throttle including a hollow throttle plate having a plurality of perforations along its circumference is mounted on a hollow shaft, which in turn may be coupled to a vacuum consumption device. When vacuum is demanded by the vacuum consumption device, the opening of the throttle plate may be decreased and vacuum may be generated by flowing intake air past the perforated edge of the throttle plate.

The invention relates to a video description generation method and device based on a bidirectional time sequence diagram. The method comprises the following steps: extracting video frames from a videoand carrying out object detection, wherein each video frame is detected to obtain a plurality of objects; constructing a two-way time sequence graph for the video object, including a forward graph and a reverse graph, and performing calculating to obtain a two-way time sequence track of the object; extracting local features from the video frame and the object, constructing a feature aggregation model, and aggregating the local features to obtain aggregated features with high expression capability; and constructing a decoding model to generate natural language description, and adaptively distinguishing different video frames and different object instances by utilizing a hierarchical attention mechanism in the generation process. According to the method, the time sequence track of the videoobject can be modeled through the bidirectional time sequence diagram, the time sequence change information of the video object can be effectively expressed, the expression capability of video features is improved by utilizing local feature aggregation, and fine-grained video space-time information is modeled, so that the accuracy of video description generation is improved.

The invention discloses a random number generating method and device and relevant equipment. A client side sends a random entropy obtaining request to a service side, and receives entropy data corresponding to the random entropy request sent by the service side; the client side obtains initial entropy source data, and mixes the initial entropy source data with the entropy data sent by the client side to generate random numbers, wherein the initial entropy source data is local weak entropy source data obtained by the client side. By means of the random number generating method and device and the relevant equipment, on the condition that the client side does not have enough entropy sources, the weak entropy source can serve as salt figures for processing through the entropy data sent by theservice side, true random numbers are obtained, and then the high-quality random numbers can be generated.

Various systems and methods are described for generating vacuum within an engine intake. A system may comprise an intake throttle including a throttle plate, where the throttle plate may comprise a hollow interior passage, which in turn may be coupled to a vacuum consumption device. When vacuum is demanded by the vacuum consumption device, the throttle plate may be adjusted to increase airflow through the interior passage and vacuum may be generated at a constricted portion of the interior passage by flowing intake air through the interior passage of the throttle plate.

The invention relates to a visual scene graph generation technology, and discloses a visual scene graph generation system and method based on relation regularization, which can quickly and effectivelyjudge whether a relation exists between objects, and is beneficial to enhancing the detection effect of an object detection model. The system comprises an object detector, an object label refiner andan object relation generator. An object in the image is detected through an object detector to botain a label, an object frame feature and a joint frame feature of the object; And the labels of the objects is refined by using an object label refiner, the relationship between the objects is acquired by using an object relationship generator, and generating a final visual scene graph. The system and methods are suitable for generating the visual scene graph.

A push-pull type speaker device and a method of manufacturing the same are provided. The speaker device includes a frame, a permanent magnet installed within the frame using a fixing device, a first diaphragm and a second diaphragm installed in opposite end portions of the frame, a pair of electromagnets, respectively corresponding to the first and second diaphragm, each installed at a distance from the permanent magnet,; and a power source driver for supplying current to the pair of electromagnets, wherein a winding direction of each electromagnet is formed such that, when current is provided from the power source driver to each of the electromagnets, the pair of electromagnets have an opposite magnetic flux direction.

The invention relates to the field of beauty products, and particularly relates to a mineral far-infrared alkaline anion activated concentrate and a preparation method thereof. The mineral far-infrared alkaline anion activated concentrate is prepared by mixing water, a complexing agent and ore particles with diameters of 0.1-2mm, wherein the ore particles are prepared by mixing alkaline tourmaline, stone needle, germanite, nano antibacterial far-infrared balls, Gaussian anion balls, mineralized medical stone balls, Muyu-stone mineralization balls, nano antibacterial columns, negative potential purified water nano mixed balls, Gaussian magnetic ceramic balls and nano silver antibacterial balls; the complexing agent is prepared by mixing sodium carbonate, sodium bicarbonate, potassium carbonate, magnesium chloride, strontium carbonate, potassium iodide and soft sugar; and the addition amount of the water is 10 times the total addition amount of the ore particles. The mineral far-infrared alkaline anion activated concentrate can effectively enhance the cell activity of the body skin and slow down the aging and avoids side effects on human bodies; and the preparation method is simple and suitable for batch production.

Various systems and methods are described for generating vacuum within an engine intake. A system may comprise an intake throttle including a throttle plate, where the throttle plate may comprise a hollow interior passage, which in turn may be coupled to a vacuum consumption device. When vacuum is demanded by the vacuum consumption device, the throttle plate may be adjusted to increase airflow through the interior passage and vacuum may be generated at a constricted portion of the interior passage by flowing intake air through the interior passage of the throttle plate.

The invention, which belongs to the military laser test technology, provides an in-situ test system of a strong laser beam quality. The in-situ test system of the strong laser beam quality provided in the invention comprises: an optical measuring head unit, an information processing computer, a master-controlled electrical cabinet, a scattering receiving system and an indication laser. By using the system, the laser system beam quality of a plurality of platforms, such as a space-based laser system, a foundation, an airborne laser system, a shipbased laser system and vehicle-mounted, can be detected; an one-time rapid and comprehensive evaluation and an in-situ evaluation of the strong laser beam quality can be realized; generative capacity and anti-interference capability of the laser system can be raised. The system provides supports for the laser system to perform optimal efficiency.

Provided is a separation method, without causing generation of defects, such as transfer residues of liquid crystal high polymer, cracks or the separation and wrinkles or the like of a film, in a method for manufacturing of optical elements by transferring a liquid crystal polymer layer formed on an orientation substrate film from a long laminated film to a transmissive substrate film. An optical element is manufactured which has the liquid crystal polymer layer transferred to the transmissive substrate film by separating the oriented substrate film along a separation bar provided on the side of the oriented substrate film from the long laminated layer film, comprising the orientation substrate film / the liquid crystal polymer layer / a curing adhesive layer / the transmissive substrate film.

The invention discloses an image generation method based on a Gaussian mixture model prior variation auto-encoder. The method comprises the following steps: S11, presetting and generating an image training data set, wherein the training data set is composed of a plurality of batches of training data; s12, building a variation auto-encoder network of Gaussian mixture model prior; s13, uploading a plurality of preset batches of training data to the variational auto-encoder network, and determining posteriori distribution and prior distribution of the variational auto-encoder network; s14, determining a relationship between Gaussian components in the Gaussian mixture model to obtain a mapping function; s15, obtaining a reconstruction loss function and a KL divergence function by using the variational auto-encoder network and the obtained mapping function, calculating a posteriori distribution loss function and a priori distribution loss function of the variational auto-encoder network, and updating parameters of the variational auto-encoder network to generate an image; and S16, when an image is generated, taking the pseudo input as an input image and uploading the pseudo input to thevariational auto-encoder network to obtain a finally generated picture.

The invention discloses a text generation method. The method comprises the steps of obtaining a to-be-processed text; extracting keywords from the to-be-processed text, and performing diversity filling on the keywords to obtain a plurality of first texts; calculating a diversity score of each first text relative to the to-be-processed text; determining a second text from the plurality of first texts according to the diversity score; inputting the to-be-processed text and the second text into a pre-trained text generation model; and generating a third text with the meaning the same as that of the to-be-processed text. The invention further provides a text generation device, electronic equipment and a storage medium. According to the method, data enhancement is achieved, meanwhile, the textgeneration accuracy is improved, and the text generation effect is improved.

A method for manufacturing a semiconductor device includes the steps of forming a stress relief layer on a wafer such that the stress relief layer is away from at least part of electrodes formed on the wafer; forming a wiring layer in which lines that extend from the electrodes to the stress relief layer are formed; forming outer electrodes that are over the stress relief layer and are connected to the lines in the wiring layer; and forming a dielectric layer by applying dielectric liquid by an inkjet method to portions in the lines where the outer electrodes are connected to form a capacitor. The step of forming the dielectric layer is conducted after the step of forming the wiring layer.

The invention discloses a natural language generation method based on a time sequence topic model. The natural language generation method comprises the steps of obtaining a context word bag vector ofeach sentence in a document; generating a topic distribution vector of each sentence in the document by utilizing a time sequence topic model; inputting each word of each sentence and the corresponding topic distribution vector into a sequential language model to obtain each layer of hidden variable corresponding to each word; splicing the hidden variables of each layer together, and predicting the next word in the current sentence through a normalized exponential function; utilizing a stochastic gradient descent method to update encoder parameters in the time sequence language model and the time sequence theme model; and sampling and updating decoder parameters in the time sequence topic model. According to the method, the multi-layer topic model and the multi-layer language model are combined, the hierarchical semantic features and the hierarchical time sequence information in the text topic are extracted, the semantic range of the low-layer features is small, and the semantic rangeof the high-layer features is wider.