45results about How to "Feature extraction is simple" patented technology

The invention provides a vehicle attribute identification method based on a multi-task convolutional neural network. The method comprises a training process and an identification process. Particularly the method comprises the steps of acquiring a picture of a to-be-identified vehicle, designing a multi-task convolutional neural network structure and training a network model vehicle attribute identification, identifying a vehicle model and returning vehicle window position coordinate of the vehicle, designing a vehicle image mask and generating a new vehicle image, extracting a multi-task convolutional neural network characteristic of the new vehicle image, training an SVM classification model, and identifying vehicle color. The vehicle attribute identification method is advantageous in that manual characteristic definition and re-classification by a user are not required; the multi-task convolutional neural network structure can simultaneously receive and process a plurality of tasks; and furthermore based on the multi-task convolutional neural network, structure information of the vehicle in the vehicle image is acquired for realizing an effective vehicle color identification method and improving identification accuracy, thereby supplying accurate basis for intelligent traffic.

The invention discloses a traffic light positioning method. The method comprises: expanding an object region where traffic lights are in in advance, obtaining a constraint region, and performing binaryzation on a grayscale image of the constraint region, extracting proportion of bright regions, the number of small area bright regions, the number of standard light area bright regions, the number of large area bright regions, and the average gray value of dark regions from binarization images, according to the features, determining day or night of an intersection image where the object region is in, if the image is a night image, detecting signal lights whose states are lightened through the binarization image, and according to the positions of the signal lights, adjusting positions of the traffic light; if the image is a day image, detecting a black frame in a traffic light structure through the binaryzation image, and according to the position of the black frame, adjusting the positions of the traffic light. The method is advantaged by low calculation complexity, fast operation speed, good environment adaptability, and simple realization.

The invention discloses a method for estimating the abundance of a hyperspectral image end member. The method for estimating the abundance of the hyperspectral image end member comprises a first step of extracting the image end from an image and selecting mixed pixel points, conducting linear decomposition and obtaining a corresponding abundance value, a second step of evaluating a normalized spectral characteristic value which corresponds to the end member, a third step of tracing points under a rectangular coordinate system, a fourth step of conducting curve fitting and obtaining a quadratic curve expression, a fifth step of obtaining the abundance of the rest points by means of mapping the spectral characteristic value, and a sixth step of evaluating the root mean square error RMSE between an estimated value and an actual value and judging whether the RMSE meets the evaluated precision. According to the method for estimating the abundance of the hyperspectral image end member, an abundance value is rapidly predicted through establishment of certain relationship between the spectral feature value and the abundance of the end member, the defect that the corresponding abundance can be obtained when the liner decomposition is conducted on all mixed pixel points is overcome. In actual application process, due to the fact that the linear decomposition is conducted on only a small amount of pixel points which are evenly distributed, the abundance value which corresponds to end members of all pixel points can be obtained, and the time of decomposition of the pixel points can be effectively shortened.

The invention discloses a graph-structured-data generation method based on deep convolutional generative adversarial networks (DCGAN), and belongs to the field of big-data technology. The method of the invention applies the deep convolutional generative adversarial networks to the generation field of graph-structured data. According to the method of the invention, firstly, actual graph-structureddata are converted into grid structure pictures with uniqueness; then learning of a generation model is carried out through the deep convolutional generative adversarial networks; and finally, specific-scale simulated graph-structured data are generated through the generation model, which is obtained by learning, according to input parameters and requirements. According to the graph-structured-data generation method based on the deep convolutional generative adversarial networks of the invention, features are extracted mainly from the actual graph-structured data to construct the model, and thus attributes which the generated simulated graph-structured data have are enabled to more accord with the actual graph-structured data.

The invention discloses a method for estimating pavement type based on depth convolution neural network without loss function, which comprises the following steps: step 1, collecting pavement workingcondition image, calibrating pavement type and establishing pavement working condition database; The depth convolutional neural network based on loss-free function is trained to obtain image features,and then binary hash coding and histogram processing are performed to obtain image feature output vectors. According to the feature output vector of the image and the corresponding road type, the support vector machine is trained and the parameters are selected, and the road type discrimination function is determined. 2, collect that working condition image of the pavement to be tested, obtainingthe characteristic output vector of the pavement to be tested according to the step 1, and determine the type of the pavement to be tested by using the trained support vector machine. It simplifies the feature extraction of convolution neural network depth learning model, and uses support vector machine to classify images, which greatly reduces the difficulty of convolution training and improvesthe efficiency of classification.

The invention belongs to the image processing technology field and relates to a pedestrian reidentification method based on a color texture distribution characteristic. The method comprises the following steps of inputting N image pairs to be matched including training data and test data and a corresponding label ln, wherein the n=1,..., N; extracting the color texture space distribution characteristic expression of input image data; acquiring the consistent characteristic expression of the color texture space distribution characteristic expression through multi-scale characteristic matching;and constructing two classifiers for the acquired consistent characteristic expression and outputting a probability expression describing a same target. The method has advantages that characteristic extraction is simple and a speed is fast; a training period is short; robustness is high; and precision is high.

The invention discloses a JPEG image double-compression automatic detection method, and the method comprises the steps: carrying out the preprocessing of a JPEG image; extracting a first effective number feature of a DCT coefficient and adjacent coefficient difference features of the JPEG image based on the Markov model, carrying out the fusion of the features, and carrying out the feature dimension reduction; and training and recognizing a support vector machine. The fusion features can be effectively used for various compression conditions (besides the common compression conditions, also comprising a condition that two compression qualities are equal and the second compression quality is 95). Moreover, after the extraction of the fusion features, the dimension reduction of a feature matrix is carried out, and the matrix after dimension reduction is taken as a feature vector for subsequent operation. The method reduces the calculation complexity, improves the algorithm efficiency, and is easy to use in reality.

The invention discloses a machine tool cutter vibration monitoring and analyzing method based on artificial intelligence and big data. According to the method, on the premise of big data, artificial intelligence methods such as machine learning and deep learning are combined to perform feature extraction, feature transformation and feature learning on vibration signals generated by key devices soas to monitor and analyze the key devices. The beneficial effects of the invention are that the vibration analysis is simple in feature extraction, is convenient to install and deploy, and does not damage the structure of a measured object body.

The invention discloses a classroom student behavior detection method, relates to image processing and recognition, and mainly solves the technical problems of low accuracy, low speed and poor robustness of a current full-image detection method. The classroom student behavior detection method comprises the following steps: converting an original image of a classroom student into a grey-scale map;inputting the grey-scale map into a deep learning model to carry out head and shoulder frame detection and prone table identification, and obtaining a head and shoulder frame; determining a hand raising detection area according to the head and shoulder frame, and performing hand raising detection in the hand raising detection area. The hand raising detection area is determined according to the head and shoulder frame, and hand raising detection is performed in the hand raising detection area so that the detection input image can be greatly reduced, the detection area can be greatly reduced, the hand detection efficiency can be effectively enhanced, the hand detection accuracy can also be greatly enhanced and the robustness is high.

The invention provides an aircraft voiceprint recognition method and system, electronic equipment and a storage medium, and belongs to the technical field of aircraft voiceprint data processing. The method comprises the following steps: acquiring sound signals in an environment in real time; sampling and quantizing the collected sound signals to obtain time domain features of the sound signals; converting the time domain feature into a frequency domain feature by using FFT (Fast Fourier Transform); and carrying out three-layer identification on the frequency domain characteristics to obtain an aircraft sound signal, and sending an indication signal. Based on the scheme, the problem that in the prior art, airplane signals are difficult to identify quickly in real time in an outdoor unmanned environment is solved.

The invention discloses a spatio-temporal interest point feature encoding method in human motion recognition. Human motion recognition in a video has broad application prospects in aspects of intelligent monitoring, video retrieval and the like. A human motion recognition method based on spatio-temporal interest point features has the advantages of simple extraction, strong anti-interference ability, good robustness and the like, and thus wide attention is got. However, during a process of encoding the spatio-temporal interest point features through vector quantization so as to acquire video representation vectors, problems of large representation errors, weak encoding discrimination ability and lost spatial position information exist. In order to solve the above problems, the invention discloses spatial regulation local constraint encoding algorithm. During a feature encoding process, local constraints are introduced for reducing representation errors, spatial regulation is introduced for enhancing discrimination of an encoding result and using the spatio-temporal interest point features for the spatial position information, and finally, the precision of human motion recognition is improved.

The invention belongs to the technical field of signal detection and estimation, and discloses a triangular linear frequency modulation continuous signal parameter estimation method based on differential envelope detection. A receiver samples an observed triangular linear frequency modulation continuous wave signal from the radar to obtain a sampling sequence; the receiver carries out differentialoperation, Hilbert transform and low-pass filtering on the sampling sequence to obtain a denoising envelope sequence; the receiver calculates according to the denoising envelope sequence to obtain signal characteristic parameters including a positive modulation frequency, a negative modulation frequency, a frequency modulation signal period, a minimum frequency in a sweep frequency interval and amaximum frequency in the sweep frequency interval. The triangular linear frequency modulation continuous signal parameter estimation method based on differential envelope detection has the advantagesof characteristic extraction and calculation in the time domain and lower time complexity, and can be used for solving the parameter estimation problem of the triangular linear frequency modulation continuous signal; the characteristic parameters of the triangular linear frequency modulation continuous signal can be obtained by utilizing the sequence difference and the envelope detection, and a parameter estimation value is calculated according to the envelope slope and the change time of the envelope slope.

The invention provides a noise frequency modulation signal identification method, and belongs to the technical field of communication. According to the method, the intrinsic characteristics of noise frequency modulation are extracted and judged by utilizing the characteristics of the noise frequency modulation in frequency spectrum and energy distribution, so that the purpose of accurately identifying the noise frequency modulation signal is achieved. Specifically, digital MPSK modulation signal judgment, digital MASK modulation signal judgment, digital MFSK modulation signal judgment and analog modulation judgment are sequentially carried out, and the residual frequency hopping signals are recursively sorted by utilizing the exclusiveness of two frequency hopping signals in the same period. Features are extracted through a signal time domain and a signal frequency domain, parameter estimation is carried out by means of spectral lines and frequency spectrums, noise frequency modulationsignal recognition is carried out according to a priority step-by-step confirmation method, and the method has the advantages of being simple in algorithm, small in calculated amount, easy in featureextraction and low in engineering implementation difficulty.