30results about How to "Accurate recognition rate" patented technology

The invention provides an underwater target feature extraction method based on a convolutional neural network. 1. Divide the sampling sequence of the original radiation noise signal into 25 consecutive parts, and set 25 sampling points for each part; 2. Normalize and centralize the sampling samples of the j-th segment data signal; perform short-term Fourier transform to get the LoFAR image; 4. Assign the vector to the existing 3-dimensional tensor; 5. Input the obtained feature vector to the fully connected layer for classification and calculate the error with the label data, and check whether the loss error is lower than the error threshold. If it is lower than that, stop the network training, otherwise go to step 6; 6. Use the gradient descent method to adjust the parameters of the network layer by layer from the back to the front, and go to step 2. The recognition rate of the method of the present invention is compared with the traditional convolutional neural network algorithm, and the multi-dimensional weighting operation of the spatial information is performed on the feature layer to make up for the defect of the loss of spatial information caused by the one-dimensional vectorization of the fully connected layer. .

The invention provides an RFID (Radio Frequency Identification) reader with multiple antennae working simultaneously, and a radio frequency data signal identification method of the same. The RFID reader comprises an anti-collision unit, more than two radio frequency units and more than two antennae, wherein the radio frequency units are respectively connected with the anti-collision unit, and each antenna is connected with one radio frequency unit. The RFID reader can be used for realizing simultaneous working of multiple antennae of RFID, and solving the problem that the existing multi-antenna RFID system is prone to identification omission when only one antenna can work at the same time.

The invention provides a method for detecting file stealing Trojan based on a thread behavior, which comprises the following steps of: monitoring file operation and network operation of a thread; forming a behavior sequence buffer queue by the monitored thread, the process of the monitored thread, intercepted file operation, data read by a file, the network operation and data transmitted by a network; judging whether the file read by the thread is a file transmitted by the network or not according to the behavior sequence in the buffer queue; and if SO, checking whether the thread and the process of the thread have secrecy, if SO, judging whether a behavior that Trojan steals the file exists. The invention also provides a system for detecting the file stealing Trojan based on the thread behavior. According to the method and the system, provided by the invention, the misreport of the normal file transmission can be reduced, and the detection to the behavior that the Trojan steals the file is improved.

The invention belongs to the technical field of artificial intelligence, and discloses a behavior recognition method based on local joint point trajectory space-time volume in a skeleton sequence, which extracts the local joint point trajectory space-time volume from input RGB video data and skeleton joint point data; The pre-training model of the video data set extracts image features; constructs a codebook for each different feature of each joint point in the training set and encodes it separately, and concatenates the features of n joint points into a feature vector; uses the SVM classifier Behavior classification and identification. The present invention fuses manual features and deep learning features, and uses deep learning methods to extract local features, and the fusion of multiple features can achieve a stable and accurate recognition rate; the present invention uses the 2D human skeleton estimated by the pose estimation algorithm and the RGB video sequence. Extracting features has low cost and high accuracy, and it is of great significance to apply to real scenes.

The invention relates to a video tracking method based on self-adaptive block, which fully considers the difference of pixel values ​​in the target area, and the video tracking method based on particle filtering realizes efficient self-adaptive video tracking and improves the performance of video tracking. Accuracy and adaptability. For problems such as occlusion and interference in video tracking, the present invention can maintain a relatively accurate recognition rate, fully consider the content information in the video and image, and adaptively adjust the corresponding block strategy according to the characteristics of the tracked target area. To achieve highly intelligent, high-accuracy video tracking effect.

The invention discloses an improved training method for character recognition based on center loss. The method comprises the following steps: S1, performing training on a character recognition data set until convergence, and obtaining a pre-training model; S2, extracting the last The features of the layers are connected to obtain the feature center of the word, and the feature center of the acquired word is used to construct the feature center of a dictionary; S3, continue to train the pre-trained model until convergence. Beneficial effects: the center loss training model module adopted by the present invention can make the feature space of the same character more compact, and for similar characters, it can make them closer to their respective feature centers, making it easier for the similar characters to be recognized Differentiation can improve the accuracy of the model in character recognition without changing the model size and inference speed.