32results about How to "Increase the number of features" patented technology

The invention relates to an encrypted traffic feature extraction method based on feature fusion, and belongs to the technical field of machine learning, network service security and traffic identification. The method comprises the following steps: step 1, extracting characteristic values of different dimensions of an encrypted data packet in an encrypted stream; step 2, calculating and normalizinga feature contribution degree, performing feature selection based on the feature contribution degree, selecting an optimal feature number n participating in fusion, and selecting the first n featuresas optimal feature quantities participating in fusion; and step 3, classifying features of different dimensions based on the optimal fusion feature number n, carrying out dimension raising and fusionon the optimal features selected in the step 2 by using a kernel function, and outputting a final feature set participating in classification. According to the method, the encrypted network traffic fingerprint can be better depicted; the relation between different characteristics can be represented; the number of features participating in fusion can be rapidly determined, and the feature fusion efficiency is improved; higher accuracy is realized.

The invention discloses an image texture feature extraction method based on a four-quadrant block mode of central window variation. For any image area of ​​2M×2N pixels, the sum of M×N pixel values ​​in the central part is firstly calculated as the threshold , and then decompose the image area of ​​2M×2N pixels into 4 sub-regions of M×N pixels, and then calculate the sum of the pixel values ​​of these 4 sub-regions and compare them with the threshold, if the sum of the pixel values ​​is greater than or equal to the threshold, Set the position coding value of the sub-region to 1, otherwise it is 0; finally, multiply the position coding values ​​of the four sub-regions by the corresponding weights and sum them up to obtain the texture feature value of the 2M×2N pixel image region. The invention only needs to use 4 bits to effectively represent the local texture information of the object, and can be applied to target detection, identification and tracking.

The invention proposes a human action recognition method based on convolutional neural network feature coding, which mainly solves the problems of complex calculation and low accuracy in the prior art. The implementation plan is: use TV-L1 to obtain the video optical flow graph; respectively perform convolutional neural network, local feature accumulation encoding, dimensionality reduction and whitening processing, and VLAD vector processing from the video space direction and the optical flow action direction to obtain the spatial direction VLAD vector and Action direction VLAD vector; combine the two direction information of video space and optical flow action to obtain human action classification data, and then perform classification processing. The invention performs local feature accumulation encoding on the convolution features, which improves the recognition rate and reduces the amount of calculation when processing complex background data, and the features obtained by fusing video and optical flow VLAD vectors have higher robustness to environmental changes , can be used in areas such as residential areas, shopping malls, and confidential places to detect and recognize human movements in surveillance videos.

A method for classification and recognition of surface EMG signals for individual users, setting the initial gesture data set D u , and collect user data to get user gesture dataset D p ; Carry out data preprocessing, perform classification labeling and denoising processing, and convert the data into a three-dimensional format matrix; build a training set, sub-order the matrices in the data set, and expand the user gesture data set D p , the new contains D u and D p A dataset of gesture data D all As the training set required for training the neural network; input the training set into the neural network of the LeNet‑5 model structure for training; obtain the recognition result through the output of the trained neural network, and save the network weight of the neural network, and call it directly next time , without retraining. The method has high accuracy, stable training results and short training period.

The invention discloses an auto financial risk control system containing structured and unstructured data. The implementation steps of the auto financial risk control system containing structured and unstructured data are as follows: data preprocessing, word segmentation, and n-grams algorithm , PCA algorithm, training, PCA reverse algorithm, set operation, loop, merge, the design of the auto financial risk control system containing structured and unstructured data is reasonable, and can effectively increase the number of features input by the risk control model, so that it can Directly improve the accuracy of risk control model prediction.