47 results about "Darknet" patented technology

A method of assessing a risk level of an enterprise using cloud-based services from one or more cloud service providers includes assessing provider risk scores associated with the one or more cloud service providers and in view of darknet intelligence data; assessing cloud service usage behavior and pattern of the enterprise; and generating a risk score for the enterprise based on the provider risk scores and on the cloud service usage behavior and pattern of the enterprise. The risk score is indicative of the risk of the enterprise relating to the use of the cloud-based services from the one or more cloud service providers.

The invention discloses a static gesture real-time recognition method based on YOLOv3, which comprises the following steps of making a training set, generating a migration Darknet-53 model, improveingcandidate frame parameters and carrying out the real-time gesture recognition, is based on the convolution neural network YOLOv3 model, uses the four types of image datasets collected by Kinect equipment to replace the common RGB image datasets, and the recognition results of four types of Kinect test images are fused to effectively improve the recognition accuracy. The K-Means clustering algorithm is used to improve the parameters of the initial candidate frame to improve the recognition speed effectively. And a transfer learning method is adopted to reduce the training time of the model.

The invention relates to a Yolov3-based personnel target detection method, which comprises the steps of obtaining an image, and setting Anchor parameters by using a K-Means algorithm when a Yolov3-based reference network is constructed; taking a Darknet-53 network as a backbone network; introducing a feature pyramid structure to perform feature extraction on the multi-scale target; calculating theloss of the prediction frame offset by using a cross entropy loss function; designing the scale of the Anchor according to the height-width ratio of the personnel target; replacing the Darknet-53 network with a MobileNet_v2 network; improving the feature pyramid structure by introducing hole convolution; and after-treatment optimization is carried out by introducing an IoU confidence coefficientand a soft-NMS algorithm, obtaining an improved Yolov3 network, and identifying and detecting a personnel target. Through optimization and improvement of the invention, faster and more accurate detection of personnel targets can be realized.

The invention discloses a real-time vehicle logo detection method based on multi-scale feature fusion and DCNN. The method comprises the following steps of collecting and screening pictures; making adata set, making a vehicle logo data set according to a deep learning standard VOC data set format; designing a network, taking a YOLO framework as a basis, taking an improved Darknet-20 network as abasic network, carrying out the channel fusion on feature maps with different depths, and building a network model; carrying out model training, training the vehicle logo data set by utilizing a network model, and carrying out parameter setting, data enhancement and multi-scale training during model training; and testing and evaluating the model. According to the invention, an end-to-end one-stagenon-cascade structure is used to process the vehicle logo detection as a regression problem, so that the improved network structure can better adapt to the detection of large and small vehicle logosand similar vehicle logos under various scenes, especially has very good robustness for the detection of small vehicle logo targets, and greatly improves the vehicle logo detection speed, recall ratioand accuracy.

The invention discloses an assembly robot part deep learning recognition method which comprises the following steps: firstly, obtaining an image of a to-be-recognized workpiece by using an industrialcamera, then, recognizing the image by using a YOLOv3 network, and outputting part category and position information; wherein the YOLOv3 network comprises five residual network blocks, and is characterized in that a CFENet module is introduced after each residual network block, and the CFENet module is integrated into the Darknet-53 feature extraction network for image feature extraction.. The method has the advantages that the workpiece under the normal pose can be recognized, the good detection effect is achieved on the parts under the complex conditions of camera overexposure, workpiece mutual shielding and the like, and the recognition accuracy is high.

The invention discloses an apple grading identification method based on deep learning. The method comprises the following steps: step 1, constructing an apple training data set: 1, crawling apple image data; 2, preprocessing images; step 2, detecting apple targets: 1, selecting an apple graph in the apple data set constructed in the step 1 as test data, and performing data training by using a Darknet framework; step 2, after training is completed, shooting apple photos by using a mobile phone, detecting apple positions of images, and labeling the positions step 3, detecting apple surface defects: taking single apple photos after screenshot as input images, independently extracting each positioned apple, and positioning the four surface defects; and step 4, grading and identifying apples. Compared with the prior art, the invention has the following advantages: 1, the weight is lighter; 2, the expansibility is high; and 3, the life requirements are better met.

The invention discloses an auxiliary obstacle perception method for visually impaired people based on an improved YOLO model, and the method comprises the steps: employing Darknet-YOLOv3 as a framework, and employing Darknet-53 as a feature extraction backbone network; according to the YOLOV3 algorithm, carrying out feature fusion by using a feature map up-sampling idea in a feature pyramid network FPN, so that the precision of small target detection is improved, and various common obstacles, including road cones, stone balls, isolation columns, forbidden cross bars, handrails, fire hydrants, plants, people, pits, water pits and the like, on sidewalks can be detected and identified. The method can identify various identifications and targets at traffic intersections, including zebra crossings, signal lamps, bicycles, motorcycles, vehicles, people and the like, and can also judge upstairs, downstairs, various steps and some other obstacle targets of unknown types.

A system for analyzing and clustering darknet traffic streams with word embeddings, comprising a data processing module which collects packets that are sent to non-existing IP addresses that belong to darknet's taps (blackholes) that are deployed over the internet; a port embedding module for performing port sequence embeddings by using a word embedding algorithm on the port sequences extracted from the data processing module while transforming the port sequences into a meaningful numerical feature vectors; a clustering module for performing temporal clustering of the feature vectors over time; and an alert logic and visualization module visualizes the data and provides alerts regarding a cluster that an analyst classified as malicious in the past.

The invention discloses a pedestrian detection method based on multi-layer convolution feature fusion, and the method comprises the steps: constructing a new feature extraction network Darknet-61 through the reconstruction of a residual network, enabling the feature extraction network Darknet-61 to have the capability of six times of down-sampling, and increasing the output of a YOLO output layerof a YOLOv3 algorithm from 3 layers to 5 layers through the new feature extraction network Darknet-61; subsequently, obtaining a target candidate box on the basis of a YOLOv3 algorithm with five-layeroutput through a k-mean algorithm, and carrying out subsequent processing on the current optimal candidate box in the target candidate box through an NMS method. According to the method, the Darknet-53 feature extraction network is improved, four residual networks and convolution layers are introduced, the down-sampling frequency is increased, a 7 * 7 feature map is output, the characterization capability of low-layer features is enhanced, and the precision of large-scale pedestrian detection is improved.

The invention provides a scene recognition and classification method based on Tiny-Darknet, and the method comprises the following steps: constructing a training sample set of a classification model;building a deep learning framework based on a Tiny-darknet network; configuring training parameters, and training a classification model; acquiring equipment monitoring image information; classifyingthe scene images; judging scene changes; completing scene recognition parameter configurations. The beneficial effects of the invention are that the method achieves the automatic switching of parameter configuration through scene recognition, is precise in scene recognition and detection, is high in universality, is suitable for monitoring equipment in any scene, guarantees the quality of a monitoring image of the monitoring equipment, improves the use performance of the equipment, and meets the actual demands.

The invention discloses a railway wagon loading video intelligent monitoring system. The system comprises a sensor unit, an image acquisition unit, a wagon number acquisition unit, a lamp control unit, a carriage segmentation unit, a transmission unit and an intelligent monitoring unit. The system comprises the following steps of positioning parts based on a Darknet deep learning framework and a Yolo neural network algorithm; carrying out abnormality detection on components by utilizing an abnormality detection algorithm, adopting different methods for different types of abnormality detection,adopting a customized high-definition color linear array camera, setting the sampling frequency of the camera in combination with the running speed of a train, obtaining a high-definition color imageof the train, and restoring real details. Meanwhile, the system is higher in environmental adaptability, and the imaging quality can be guaranteed under the conditions of rain, snow, night and the like which are unfavorable for operating personnel to come to the site by themselves; furthermore, the system is higher in intelligent degree, and abnormal detection of parts can be realized for multi-class state detection items.

The invention discloses a building construction target detection method based on a YOLO neural network. The method comprises the steps: 1, collecting original image data at a construction site, dividing the collected original image data into a test set and a training set, and carrying out the preprocessing of the training set; step 2, training a Darkne-53-based target recognition model of the YOLOneural network; 3, testing the target recognition model based on Darknet-53 by using the test set to obtain a test result; 4, analyzing a test result obtained in the step 3; and step 5, acquiring animage in the construction site, and detecting the building construction target in the acquired image by using a Darknet-53-based target recognition model. According to the method, the defect that theexisting YOLO algorithm cannot quickly and accurately identify the problems of deep target layer and small scale in the building construction site image is solved.

The invention relates to an AI intelligent identification method for numbers of a water meter. The AI intelligent identification method adopts Darknet-19 as a basic network, and is realized based on YOLOv2: numbers in each box in a data source-data annotation and format conversion-creation.Names file-creation.data file-modification.Cfg configuration file-training model-testing model-model is presented on a use interface in a continuous Arabic number combination form. According to the image identification and detection method, water meter number identification is achieved based on YOLOv2, and Darknet-19 is adopted as a basic network, so that the effect is improved without deepening or widening the network while the network is simplified, and obvious advantages are achieved in the aspects ofthe recall rate and positioning; and the precision is high, and finally the numbers can be presented on the use interface in a continuous Arabic numeral combination mode, and complete artificial intelligence is achieved, and a user does not need to analyze and guess, so that the AI intelligent identification method is visual and convenient.

The invention discloses a small target detection method based on improved YOLOv3, and the method comprises the steps: constructing a feature extraction network of the improved YOLOv3, and replacing the DarkNet-53 of the YOLOv3 with the feature extraction network which consists of residual modules containing a channel attention mechanism; constructing a feature fusion network of the improved YOLOv3, and adding a structure from bottom to top on the basis of a feature pyramid network of the YOLOv3; adopting a K-means + + clustering algorithm to replace the K-means algorithm to cluster the self-made data set to generate an anchor box; constructing an improved YOLOv3 detection network, and carrying out detection on feature maps with the sizes of 52 * 52 and 26 * 26; and constructing a loss function of the improved YOLOv3, and the CIoU loss function being used as the loss of the target box. Compared with an existing YOLOv3 algorithm, the method has the advantages that the detection precisionof the small target is improved by 9.96%, and the detection precision is obviously improved.

The invention relates to a defect detection method and system for a metal three-dimensional lattice structure. The method comprises the following steps: acquiring a sectional image of a metal three-dimensional lattice structure; performing feature extraction on the tomographic image by adopting a feature extraction network in a darknet-53 network model to obtain multi-scale prediction informationcorresponding to the tomographic image; determining a defect category corresponding to each prediction frame according to the prediction information; according to the clustering center in the darknet-53 network model, correcting the position information of each prediction box by using a yolol layer to obtain the actual position of each prediction box; determining a final prediction box by adoptingan NMS algorithm according to the actual position of the prediction box; and determining the position and the defect type of the defect in the metal three-dimensional lattice structure according to the final prediction box. The method can be suitable for a metal three-dimensional lattice structure, and then the defect detection accuracy is improved.

The invention discloses an infrared target detection method based on improved YOLOv3, and the method comprises the steps: taking Darknet-53 as a network detection framework, removing a convolution layer between the network detection framework and a prediction module, adding multi-scale fusion prediction, and fusing repeated blocks of low-level features through a residual layer; adding attention modules to the bottoms of the repeated blocks, and adding a residual pyramid transition network between the repeated blocks, wherein the number of channels of the repeated blocks is increased progressively along with the number of times of repetition. The method has the advantages of being good in feature extraction capacity and information transition level when the target of the infrared image is detected.

Disclosed is a food material identification method suitable for embedded equipment, a backbone network Darknet-53 of YOLOv3 is replaced with a lightweight network MobileNet, a DIOU-NMS algorithm is adopted to replace a traditional NMS algorithm, the detection speed is greatly increased while multi-class detection of food materials is met, and the real-time detection requirement of the food materials is met; and while the real-time detection requirement is met, the detection accuracy is improved, and the detection accuracy requirement is met. The rapid transplantation of the detection model between the server and the embedded platform is realized, so that the detection performance of the detection model on miniaturized equipment (especially mobile equipment) based on an embedded system can be effectively ensured, and the method has a wider application prospect.

The invention provides a video-based people number anomaly detection method, comprising the steps: S1, collecting a sample, and training a personnel detection model; training a yolo model under a darknet framework by labeling personnel in a sample; S2, setting a monitoring area and allowable people number abnormal duration; S3, preprocessing the video image of the monitoring area, and removing image noise points; and S4, detecting the image by using a deep learning Yolo model to obtain the position and confidence of the personnel in the area. The video-based people number anomaly detection method achieves the balance of indexes in effect and performance by tailoring the original yolo network, and solves the problem of poor performance of the original network is solved. By utilizing a largenumber of samples and sample enhancement, the video-based people number anomaly detection method increases diversity of the samples, improves the detection effects of various people, and reduces theinfluence of false detection on people number abnormality alarm by setting a calculation method for people number abnormality duration.

The invention provides a power equipment identification method and system, a medium and electronic equipment. The method comprises the following steps: acquiring a to-be-identified image; according to the obtained image and a preset convolutional neural network model, obtaining a positioning identification result of the electric power external insulation equipment, wherein the preset convolutional neural network model adopts a YOLO-V3 model, a standard convolutional structure in a basic network Darknet-53 of the YOLO-V3 model is replaced by a deep separable convolutional structure, and a full connection layer and a Softmax layer of the Darknet-53 are removed; according to the method and the device, the detection capability of the model on a small target is enhanced while the real-time performance is high, and real-time and efficient target detection can be better realized on embedded terminal equipment.

The invention discloses a driverless environment target detection method based on embedded equipment. The method comprises the steps of firstly completing the recognition of a target through employing an improved YOLOv3 algorithm, carrying out the improvement of a dimension clustering algorithm K-Means of the YOLOv3, obtaining an optimal anchor frame scale adaptive to a data set through employing an improved K-Means + + algorithm, and improving the clustering precision; secondly, for a Darknet-53 backbone network in YOLOv3, shortening the forward reasoning time of the model by combining a BN layer and a convolutional layer of the backbone network; and finally, realizing great reduction of model parameters through a model compression technology so as to be applied to embedded equipment, analyzing real-time video data input by a camera, and judging a target category in a current scene.

The invention discloses a blind person auxiliary walking method and system based on deep learning target detection. A DarkNet-19 target detection model optimized based on a YoLov2 model is constructed, a training set is utilized to train a DarkNet-19 target detection network, a development set is utilized to test the trained model, and the model is continuously optimized according to test result optimization; a model weight file is obtained according to the trained and optimized DarkNet-19 target detection network, the model weight file is read into an ImageAI library fused with a non-maximum suppression algorithm to obtain an auxiliary blind person walking detection model, and the auxiliary blind person walking detection model is verified by using a test set until a convergence condition is reached; a to-be-detected monitoring video is obtained, a blind person walking assisting detection model is adopted to detect the road surface condition in each real-time monitoring picture in the monitoring video, various obstacles encountered on the road by the blind person, the traffic condition and the road surface information are fed back to the blind person, and the technical problems that an existing blind person walking assisting tool is low in intelligent degree, and cannot be safely and accurately assist the blind in walking are solved.