42results about How to "Optimize network parameters" patented technology

The invention discloses an end-to-end semantic instant-positioning and graph building method based on deep learning. A continuous original image sequence and an original three-dimensional point cloudsequence corresponding to the image sequence are respectively collected and obtained through a color camera and lidar, and pose transformation information, depth information and semantic segmentationinformation of continuous five-image sequences are obtained by processing; and a multi-task deep neural network with branches is constructed, the sequences are input into the multi-task deep neural network, and the multi-task deep neural network is trained for obtaining parameters, and the trained multi-task deep neural network is adopted to process a consecutive to-be-tested five-image sequence to obtain pose transformation information, depth information and semantic segmentation information among images. Compared with a traditional ORB-SLAM algorithm and methods similarly based on deep learning, the method of the invention has better performance.

The invention discloses an SAR image target classification method based on NSCT double CNN channels and a selective attention mechanism. The method comprises steps that training sample sets D1 and D2 for target detection and classification are acquired; the D1 and D2 are expanded to acquire sample sets D3 and D4; models M1 and M2 for target detection and classification are respectively trained; significance detection and morphological processing on test images are carried out, connected domain marking is further carried out, target candidate areas corresponding to a connected domain mass center are extracted, translation in multiple surrounding pixel points is carried out, and the target candidate areas are generated; classification determination of the target candidate areas is carried out through utilizing the M1, and accurate positioning of a target is acquired; a final class of the target is determined through voting decision after M2 classification. The method is advantaged in that a non-down-sampling contour wave layer is added, low frequency and high frequency characteristic images are inputted to a double channel CNN to form the NSCT double channel CNN, the selective attention mechanism is applied to SAR image classification, SAR image target detection classification accuracy is improved, and a problem of low target classification accuracy in the prior art is solved.

The invention belongs to the field of computer vision and specifically relates to a space-time bone characteristic and depth belief network-based human body behavior identification method. According to the method, a bone sequence is obtained via a depth camera based on each motion; bone point time sequence characteristics are extracted from the bone sequence and comprise displacement characteristics, speed characteristics and acceleration characteristics of each bone point; a bone reference point is chosen, relative distance between each bone point and the bone reference point is calculated, space characteristics of a motion bone sequence is obtained, the bone point time sequence characteristics and the space characteristics of a motion bone sequence are subjected to clustering operation,a characteristic length is fixed, a global motion description operator is established and subjected to dimensionality reducing operation, and motion classification is finished based on cross validation and a classifier.

The embodiment of the invention discloses an edge detection network optimization method, a pavement disease identification method and a pavement disease identification system. An edge detection network is used for identifying the specified pavement disease. The optimization method comprises the steps of segmenting each sample pavement picture into a plurality of first grids; obtaining a first matrix corresponding to each sample pavement picture according to the fact whether the first grids possess the specified pavement disease; inputting each sample pavement picture to the edge detection network for identification of the specified pavement disease, and outputting a second matrix corresponding to each sample pavement picture; calculating to obtain a loss function of each sample pavement picture according to vectors of the first matrix and the second matrix corresponding to each sample pavement picture; and optimizing a network parameter of the edge detection network until the second matrix output by the edge detection network with optimized network parameter makes an arithmetic average value of the calculated loss functions of all sample pavement pictures minimum. According to theembodiment, an identification result of the edge detection network is accurate and efficient.

A method for identifying a continuous stirred tank reaction process based on a deep neural network comprises the following steps: step (1) of obtaining process variable data during continuous stirredtank reactor operation; step (2) of performing data pre-processing on the collected process variable data: first needing to standardize the data; selecting time lag, and organizing the process variable into a three-dimensional input form; and finally, dividing the data into a training set, a verification set and a test set; step (3) of establishing, a three-dimensional long-term and short-term memory nerve network, an identification model and training: utilizing a memory unit to establish a three-dimensional long-term and short-term memory nerve network model, and determining a network structure and hyperparameters; utilizing an adaptive moment estimation algorithm to optimize network parameters on the training set, selecting the hyperparameters of the network model on the verification set, and establishing the identification model based on the three-dimensional long-term and short-term memory nerve network and performing training. According to the method, online monitoring on a process state is performed, and the precise identification of the product concentration is achieved.

The invention provides a method and a system for locating a random access failure reason of User Equipment (UE). The method comprises the following steps that: the UE reports a random access parameter employed by the UE in a failed random access process to the network side; and the network side locates the reason of the random access failure of the UE according to the random access parameter reported by the UE. The system comprises the User Equipment (UE) and a network side apparatus, wherein the UE is used for reporting a random access parameter employed by the UE in a failed random access process to the network side apparatus; and the network side apparatus is used for locating the reason of the random access failure of the UE according to the random access parameter reported by the UE. By employing the method and the system provided by the invention, the reason of the random access failure of the UE can be located conveniently, so that the network performance can be effectively and quickly optimized by the network side.

The invention discloses a self-supervised learning fine-grained image classification method based on a twin network. According to the method, firstly, an attention encoder is used for extracting deep convolution features of an image, an attention graph containing semantic information is obtained, and image features are encoded in a bilinear pooling mode; secondly, positioning a salient region where a high response value is located on the attention map from the original image, cutting and erasing the salient region so as to form views of different view angles, and learning view angle invariance features in a self-supervision mode; and finally, combining the center loss function and the consistency loss function to display and constrain different view angle characteristics, and keeping the intra-class consistency of the different view angle characteristics. The method enables the network to obtain significant performance gain, and can significantly improve the classification accuracy on the baseline of fine-grained image classification.

The invention discloses a method for optimizing grid-related parameters for a power generator unit based on a double-stranded quantum genetic algorithm. The method comprises the following steps: selecting the grid-related parameters required for the optimization of the power generator unit in a power grid system; calculating constraint conditions of the grid-related parameters required for the optimization of the power generator unit and total objective functions with optimal transient stability; determining fitness functions of the grid-related parameters for the power generator unit; initializing a double-stranded quantum population; judging whether a current chromosome is mutant or not by utilizing the mutation probability and performing quantum bit non-gate mutation in case of mutation; converting the probability amplitude of each quantum bit of the current chromosome into a solution space, determining a value-substituted transient stability calculation program obtained by conversion into a total objective function value with optimal transient stability and performing fitness evaluation to determine individual fitness and store a global optimal solution; calculating a quantum rotation angle advance step length and updating a quantum gate to obtain a next generation of chromosome; by optimizing the grid-related parameters for the power generator unit, the transient stability of the system is improved, and the machine-grid coordination is realized.

The invention provides a rain image rain stripe removing method and system based on image filtering and CNN. According to the method, a rain image data set is expanded and trained by using a rain image data set enhancement method, multiple groups of rain images and label images are acquired, the rain images are subjected to image filtering to obtain rain image high-frequency parts, the rain imagehigh-frequency parts are input into a rain removal network to obtain rain removal images, and an SSIM loss function is added to optimize the rain removal network; the rain removal image is input intothe H-G discrimination network to obtain a discrimination result, the discrimination result is fed back to the rain removal network to further improve the rain removal quality, and the problems that an existing rain removal method is difficult to accurately describe a rain model, rain stripes cannot be completely removed, and details are easy to lose are solved.

The embodiment of the invention discloses a neural network training method, an image processing method and device, equipment and a medium. The method comprises steps of carrying out the splicing of sample images, and obtaining a spliced sample image; extracting feature data of the spliced sample images through a first neural network; splitting the extracted feature data according to a rule of splicing the sample images to obtain feature data of each spliced sample image; determining a prediction result of each image processing task through a first neural network according to the feature data of each sample image; according to the determined prediction result of each image processing task and the corresponding annotation result of each sample image, obtaining an annotation result. Accordingto the method, the network parameters of the first neural network are adjusted, training is carried out through the mixed data set, the network parameters of the first neural network are optimized through different data sets, and the first neural network with the optimized parameters can process a plurality of image processing tasks at the same time.

The invention discloses a multi-source ciphertext image retrieval method based on federated learning and secret sharing. The method comprises the following steps that: S1, model training is performed on a convolutional neural network of a double-cloud platform by taking an image owner joining the double-cloud platform as a federated member based on federated learning; and S2, the authorized user completes ciphertext image retrieval by means of the double-cloud platform based on additive secret sharing. According to the scheme, a multi-source ciphertext image retrieval scheme based on federated learning and secret sharing is provided, a neural network model structure used by federated learning simplification retrieval is utilized to obtain better network parameters, better neural network parameters and a simpler network model structure are obtained through overhead of the image owner end, a better convolutional neural network can be used in ciphertext image retrieval, and a user can obtain a better result faster during retrieval.

The invention relates to a self-inspection method for a camera mechanism of a robot, electronic equipment and a storage medium, and the method comprises the steps of collecting an image when the camera mechanism is shielded in a preset scene, and carrying out the preprocessing of the image, thereby forming a training set; constructing a self-checking network model of the camera mechanism, wherein the self-checking network model of the camera mechanism comprises multiple layers of sensing units; and training a camera mechanism self-inspection network model by adopting images of the training set, optimizing network parameters of the camera mechanism self-inspection network model, thereby adopting a robot camera shielding self-inspection strategy based on a neural network algorithm, and meanwhile, a mode of combining a convolution module and a multi-layer perceptron are used for simplifying the calculation complexity of a shielding detection module, and in combination with business scene characteristics, the shielding self-checking module is dynamically called, so that robot resources are saved.

The invention discloses a finger vein verification method, electronic equipment and a storage medium, which are applied to a U-Net network architecture, and the method comprises the steps: obtaining afinger vein image, and carrying out the ROI extraction of the finger vein image; extracting vein lines through a method for detecting the local maximum curvature of the cross section of the finger vein, and obtaining a segmented image; carrying out differential calculation processing on any two segmented images to obtain a differential image so as to carry out data expansion; carrying out channelcopying on the difference image and then taking the difference image as input of a pre-training stage, and reserving an optimal weight of the pre-training stage as a pre-training weight of a cascadeoptimization stage pair; and performing channel superposition on the differential image and the two original segmented images to serve as input of the cascade optimization stage, optimizing parametersof a pre-trained network, and obtaining an optimized finger vein verification model.