165 results about "Learning architecture" patented technology

The invention discloses a network flow fingerprint feature two-stage multi-classification Internet of Things device identification method, belongs to the technical field of Internet of Things device access control, and the algorithm extracts network flow features from network flow and matches and identifies accessed Internet of Things devices. The algorithm mainly comprises the following steps: firstly, acquiring N pieces of network message data when an Internet of Things device starts an access stage, and extracting features from three dimensions of sequence field contents, sequence protocolinformation and sequence statistical values to serve as device fingerprint features; using a one-to-many multi-classification machine learning architecture to perform preliminary identification on theto-be-detected Internet of Things device; and if a plurality of identification results appear in the preliminary identification, inputting the results into a maximum similarity comparison module forsecondary classification identification, and selecting the type with the highest similarity as a final identification result. According to the method, the problem that identification overlapping is easy to occur when the existing identification algorithm is used for identifying the Internet of Things device is solved, and the identification accuracy and uniqueness are improved.

The invention provides an image classification method based on a three-side filter and a stacked sparse autocoder. The method comprises: a smooth image is obtained by using a three-side filter, and Gaussian, speckle, pulse noises of a degraded image are filtered while a spectral-spatial feature of a pixel of the image is extracted; high-order feature extraction is carried out by using an improved stacked sparse autocoder; and supervised network fine tuning and classification are carried out by using a random forest classifier. According to the method provided by the invention, the improved stacked sparse autocoder and the random forest classifier are introduced into the hyperspectral data classification to form depth learning architecture; the improved stacked sparse autocoder can extract abstract and useful underlying features of spectroscopic data layer by layer, thereby improving the classification performance of the spectroscopic data. The method is not only suitable for hyperspectral image classification but also suitable for other image classification and has high transportability; and the image classification demand can be satisfied easily.

The invention discloses a land utilization change and carbon reserve quantitative estimation method based on remote sensing data. The method comprises the following steps: downloading an image; preprocessing the image; using and classifying land; calculating ground object carbon density according to ground survey data; making correlation analysis on the carbon reserves in the sample plots and the characteristic values, and selecting the characteristic values with significant correlation for modeling; and normalizing the screened characteristic values as an input layer of the convolutional neural network, putting the calculated carbon density of each sample plot into a network output layer, carrying out network training, and carrying out carbon reserve quantitative estimation on a to-be-studied region by utilizing a trained model. The invention is based on a hierarchical learning architecture of the multi-scale convolutional neural network, so that a land utilization classification result is better. On the basis of different feature values in the image and the carbon density obtained from ground survey data, the nonlinear relation between the feature variables and the carbon reserves is better fitted, and the final quantitative estimation result of the regional carbon reserves is improved.

A platform for training deep neural networks using push-to-corner preprocessing and adversarial training. A training engine adds a preprocessing layer before the input data is fed into a deep neural network at the input layer, for pushing the input data further to the corner of its domain.

A system and method of determining a neural network configuration may include receiving at least one neural network configuration, altering the received configuration for at least two iterations, calculating a first parameter of an altered configuration, calculating a second parameter of a consecutive altered configuration of the at least two iterations, comparing values of the calculated first parameter and second parameter, and determining a configuration having largest value of the calculated parameters.

The invention discloses a width learning-based long-term target tracking algorithm. The algorithm comprises the following steps of: establishing a width learning system; and carrying out tracking on the basis of the width learning system and establishing an acceleration steady feature algorithm-based whole-image detection mechanism. According to the algorithm, width learning system-based long-termtarget tracking is carried out, width learning architecture is relatively shallow in layer and relatively low requirement is put forward for calculation resources, so that deployment can be carried out on low-end equipment without losing great precision. According to the algorithm, an obtained target tracking model is high in training speed, small in reconstruction cost and low in time cost, andhas great superiority for detection of deformation, rotation and shielding in target tracking process. According to the algorithm, the SURF algorithm-based whole-image detection mechanism is applied,so that target information can be rapidly obtained when targets re-appear under the condition that the width learning system judges that the targets are lost, and the target positions can be updated so as to ensure that the tracking effect is more stable, robust and reliable.

Methods and systems for generating a targeted risk analysis model by using a knowledge transfer technique to enhance a generic risk analysis model are presented herein. The knowledge transfer may be temporal-based or domain-based. A first generic risk analysis model is generated to produce an outcome based on a set of input data related to a first set of features. The first generic risk analysis model is trained using a first set of training data having first characteristics. Based on a type of knowledge transfer requested, a second set of training data having second characteristics is obtained. The first generic risk analysis model is enhanced to produce a second targeted risk analysis model by retraining the first generic risk analysis model using the second set of training data.

The invention discloses a bionic robot peacock image identification method based on deep learning. The method comprises the following steps that: collecting a public face detection database as an image dataset for training and verification; designing deep learning architecture based on a convolutional neural network, and realizing a face detection function in the deep learning architecture; collecting a site image shot by a bionic robot peacock camera to fine tuning on the trained convolutional neural network to realize the face detection function under an indoor complex environment; and obtaining an empirical parameter to determine the dressing positioning of an audience, and carrying out statistics on the corresponding proportion of various colors. By use of the method, the accurate andefficient face detection and color identification of the recreational bionic robot under the complex environment can be realized, and robustness is high; in addition, for the site image, carrying outparameter fine tuning on the trained deep learning architecture; and finally, carrying out real-time face detection and dressing identification on the site image captured by the camera. The method canbe applied to science and technology museums, hotels and shops for tourists to visit and amuse.