32results about How to "Rich expressive ability" patented technology

The invention relates to a semantic segmentation method and system based on edge dense reconstruction for streetscape understanding, and the method comprises the steps: carrying out the preprocessingof an input image of a training set, enabling the image to be standardized, and obtaining preprocessed images with the same size; extracting general features by using a convolutional network, then obtaining three-level context space pyramid fusion features, and extracting coding features by using the two parts of cascade connection as a coding network; acquiring semi-input size encoding features by using the encoding features, acquiring edge features based on a convolutional network, and reconstructing image resolution by taking a dense network fused with the edge features as a decoding network in combination with the semi-input size encoding features, and acquiring decoding features; calculating semantic segmentation loss and auxiliary supervision edge loss, and training the deep neural network by taking minimization of weighted sum loss of the semantic segmentation loss and the auxiliary supervision edge loss as a target; and performing semantic segmentation on the to-be-segmented image by using the deep neural network model, and outputting a segmentation result. The method and the system are beneficial to improving the accuracy and robustness of image semantic segmentation.

The production method of optical variable image includes the following steps: using a light source and utilizing imaging system to image the shape diaphragm on homographic optical element to producinga a pair of diffraction light beams; collecting the diffraction light beams on the recording material to produce interference fringe point with a certain shape; moving the position of recording material and recording next interference fringe point, repeating above-mentioned operatino until said method is completed. Said optical variable image phototype setting system includes optical circuit system formed from light source, imaging system before diffraction, homographic optical element and imaging system after diffraction, working table and control portion. Said invention can effectively utilize energy of incident light to implement production of homographic diffraction grating.

An image processing apparatus for finding reflectivity based on a BRDF model expressing a ratio of reflection of light incident upon one point of a surface of an object to be drawn at the object surface, the image processing apparatus having an operation unit for calculating the reflectivity based on a BRDF model calculated by a quadratic-form matrix expression including a vector comprised of a light source direction vector, a viewpoint direction vector, and a normal direction vector and a matrix determining the characteristics of the BRDF model, thereby able to achieve both a variety of expression power and good operation efficiency when mounted in a programmable pixel shader and further having enough of a compactness to easily cope with a shift-variant BRDF, and a method of the same.

The invention discloses a method for improving the traffic sign recognition precision in extreme weather and environment, which is based on a YoV5 target detection model, integrates a focusing module, a cross-stage local fusion module and a spatial pyramid pooling structure, and can better extract feature map information from local features for traffic sign images with poor light, and the feature map more accurately expresses the image. For a small number of training data, the expressions of the traffic signs in different environments are simulated by using Gaussian noise, adding salt and pepper noise, reducing brightness, sharpening an image, reducing the size and the like in proportion, and the traffic signs are copied to a target-free picture by using a copying-pasting method, so that a data set is greatly enriched. By using the method provided by the invention, different image modes under different resolutions can be captured more easily, and the features of the target can be extracted and fused to the greatest extent; and meanwhile, convergence is quicker and more accurate, fewer positioning errors exist, and more accurate prediction is generated.

The invention discloses a design method for a multi-pose human face detector based MSNRD features. The design method comprises a training stage and a detection stage. The training stage comprises the steps of: determining size of a template at first, and preparing a training data set; then constructing a featurepool; and training a strong classifier under an Adaboost framework. The detection stage comprises the steps of: traversing subimages of each scale at each position of an input image at first, judging whether the image is a human face; forming a set composed of all human face images, reserving elements with large output values of the classifier if a ratio of an overlapping area of any two elements in the set to a minimal element area exceeds 0.3, repeating the step till the ratio of the overlapping area of any two elements in the set to the area of a smaller one does not exceed 0.3, and regarding the set as a final result of detection for output. The design method provided by the invention enriches the feature expression capability, simplifies the feature calculation process, avoids the training of detectors respectively for different poses, reduces training workload, and increases the detection efficiency.

The invention discloses a human face recognizing method based on multi-level local obvious mode characteristic counting. The method comprises the steps of preprocessing human face image; computing the local differential mode characteristic vectors of different orders in local adjacent domain where each pixel of the normalized human face image is positioned; coding each order of local differential mode characteristic vector of each pixel of the human face image into corresponding local obvious mode characteristic; performing block-dividing on the local obvious mode characteristic image of each order of the human face image and performing space histogram counting; splicing all local obvious mode characteristic histograms of each order of the human face image and enhancing by utilizing the whitened main component analysis; computing corresponding weight according to each order of the enhanced local obvious mode histogram characteristics; and measuring the characteristic similarity of two human face images according to the weighed cosine distance. The human face recognizing method based on the multi-level local obvious mode characteristic counting is used in the human face recognizing system on low-power consumption mobile equipment, and is lower in both time computing complexity and space computing complexity.

The invention discloses a software data flow analysis method based on intermediate language and stain analysis. The method comprises the following steps of S1 defining instruction format and expression format; uniformly describing the general instruction types of the intermediate languages, and construting the temporary variable expressions, register expressions and stain marker expressions for the data representation in stain analysis process; S2 constructing a stain propagation rule based on an intermediate language, wherein the stain mark is represented by a taint_label; S3 dynamic tracking an analysis flow based on a program data flow of the intermediate language; S4 when the program is executed, constructing the data flow relationship between the stain source and the global variableTG, the local variable TL, and the system call function parameter TF based on the tracked stain information. The method of the invention has the advantages of better accuracy, stronger comprehensiveness and richer information.

The invention discloses a large-scale MIMO fingerprint positioning method based on a complex neural network. The large-scale MIMO fingerprint positioning method mainly comprises an offline stage and an online stage. In the offline stage, firstly, sample points are divided at equal intervals in a positioning area, a base station end collects position fingerprint information of a user on each sample point in the positioning area, and a position fingerprint database is constructed; the position fingerprint information of each sample point is taken as the input of a complex neural network, the position of the corresponding sample point is taken as the output label of the complex neural network, the complex neural network is constructed, and the complex neural network is trained through the fingerprint database. In the online stage, the base station end uses the complex neural network trained in the offline stage, and uses the trained complex neural network to calculate and obtain the position coordinates of the user based on position fingerprints of the user received in real time, thereby realizing relatively high-precision user positioning.

The invention discloses a VGG deep network-based visual target tracking method. The method comprises the following steps of S1, compiling a running environment of MatConvNet; S2, establishing a VGG deep neural network; S3, performing video frame input, and judging whether an input frame is an initial frame or not, if the input frame is a non initial frame, entering the step S4, and if the input frame is the initial frame, skipping the step S4 and entering the step S5; S4, performing estimation of a new state of a target, and entering the step S5; and S5, performing online update of a filter model. Compared with a conventional visual target tracking method, more semantic information is comprised in features, and higher tracking precision can be achieved. Compared with a visual target tracking method using high-layer depth features, used low-layer data can reduce the calculation loss and does not lack the semantic information. Therefore, the tracking precision and the tracking speed areweighed, and excellent tracking performance is obtained.

The invention discloses a GAN enhanced magnetic induction imaging method and system based on complex valued convolution, and the method comprises the steps: S1, collecting voltage sequence data, constructing a complex valued neural network model, inputting the voltage sequence data into the complex valued neural network model for training, and obtaining a preliminary conductivity distribution image; s2, constructing a generative adversarial network model, and inputting the initial conductivity distribution image into the generative adversarial network model for training to obtain a generator for image enhancement; and S3, inputting the initial conductivity distribution image into the generator to obtain a high-precision target conductivity distribution diagram. According to the invention, the adversarial generative network model is used as an image optimization module to carry out image enhancement on the output of the complex valued convolutional network, the complex valued characteristics of the voltage sequence data are fully utilized, the training efficiency of the neural network and the accuracy of conductivity reconstruction are improved, and the resolution and precision of the final image are further improved.

The invention discloses an infrared image convolutional neural network super-resolution method based on visible light image enhancement. An infrared image and a visible light image of a scene are obtained through shooting of an infrared and visible light dual-resolution camera, an infrared-visible light image pair is formed, and a training set is obtained through processing; and the initialized convolutional neural network model is iteratively trained by using the training set until the number of iterations reaches a preset number of times and the convolutional neural network model is trained,and an infrared image shot by the infrared camera is input into the trained convolutional neural network model to obtain a super-resolution infrared image. According to the method, the information ofthe visible light image is utilized, the problem that the details of the infrared image are not rich in the super-resolution process is solved, the super-resolution infrared image has better detail expression ability, and the robustness of the convolutional neural network model is high.

The invention relates to the field of medical science and technology, and particularly discloses a cervix uteri abnormal cell detection device and method. The device comprises: an acquisition module for acquiring a cervical cell pathological slide image; a segmentation module for segmenting the cervix uteri cell pathological slide image into a plurality of cervix uteri cell pathological slide image blocks; a preprocessing module for preprocessing each cervix uteri cell pathological slide image block to obtain a plurality of target cervix uteri cell pathological slide image blocks; a processingmodule for processing each target cervix uteri cell pathological slide image block to obtain cervix uteri cell features, wherein the cervix uteri cell features are obtained through weighted fusion ofbidirectional features from high-order features to low-order features and from low-order features to high-order features of cervix uteri cells; and a determination module for determining the abnormalprobability of the cervical cells in the cervical cell pathological slide image according to the cervical cell characteristics. The embodiment of the invention is beneficial to improving the detection precision of the cervical abnormal cells.

The invention discloses a multifunctional musical instrument, comprising a piano head (1), a neck (12), a body (13), strings (14), a music resonant cavity (2), a lyre body resonant cavity (3), a combined resonator (4) and a nozzle (5). A cavity of the music resonant cavity (2) is provided with a through hole (22) and a mask cover (23). The music resonant cavity (3) is detachably and movably connected with the combined resonant body (4). When the resonant cavity (3) of the body of the piano is connected with the combined resonant body (4), the sound outlet (31) of the body of the piano is communicated with the sound outlet (41) of the combined cavity. The multi-function of "one instrument with multi-functions" formed by the organic and integral structural design of "sharing the same body" greatly increases the musical expression power of the whole instrument through the replacement and communication structure of the outlet between different resonant cavities and the co-current function.

The invention discloses a human body three-dimensional modeling data acquisition and reconstruction method and a human body three-dimensional modeling data acquisition and reconstruction system based on a single mobile phone. In the aspect of data acquisition, only a single smart phone is used, and an augmented reality technology is utilized to guide a user to acquire high-quality video data input for a reconstruction algorithm; therefore, a high-quality three-dimensional human body model can be stably obtained by a subsequent human body reconstruction algorithm. In the aspect of a reconstruction algorithm, a deformable implicit nerve radiation field is designed. The implicit space deformation field estimation model is used to solve the problem of tiny motion of a shot in the process of collecting data by a single mobile phone; the implicit distance field with symbols is used for representing the geometry of the human body, the expression ability is rich, and the reconstruction precision of the three-dimensional human body model is improved. By integrating data acquisition and reconstruction algorithms, reliable human body high-quality three-dimensional modeling data acquisition and reconstruction based on the single mobile phone are realized.

The invention provides a face key point detection method and device, electronic equipment and a storage medium. The method comprises: obtaining a to-be-detected face image; and inputting the face image to be detected into a face key point detection model to obtain a key point detection result output by the face key point detection model, wherein the face key point detection model is obtained by training based on a sample face image, and a sample face UV image, a sample face mask image and real coordinates of sample key points corresponding to the sample face image, and the face key point detection model is used for performing spatial self-attention enhancement on face features of the to-be-detected face image based on a face UV image and a face mask image of the to-be-detected face image to obtain face enhancement features, and performing face key point detection based on the face enhancement features. According to the method and the device provided by the invention, the accuracy of face key point detection is improved.

The invention discloses a transmission tower identification and positioning method based on a high-resolution remote sensing image. The detection of a large-range transmission tower can be realized based on the high-resolution remote sensing image. On the basis, an improved YOLOv3 target rapid detection algorithm is applied to transmission tower detection and positioning of a high-resolution remote sensing image, aiming at different shapes of the transmission tower in the remote sensing image, the size of a prior frame is reset through K-means, a CIoU loss function is introduced in frame regression, DIoU and NMS are combined, and the problem of missing detection of dense small targets by YOLOv3 is solved. In order to solve the problems that the number of layers of feature extraction deepens details of a target and position information is reduced, an SPP module is added to enrich the expression ability of a final feature map. In addition, the method solves the problem that YOLOv3 fails to detect a super-large image target, achieves the automatic recognition and positioning of the transmission tower, and provides a guarantee for the safe operation of a power grid.