33results about How to "Low classification error rate" patented technology

The invention discloses a method for key text detection and classification training of certificate pictures. The method includes the following steps: S1, constructing a certificate template, and generating training samples based on the certificate template; S2, constructing an algorithm model integrating text detection and classification, and Output the text area and its classification results based on the detection feature map and classification feature map output by the network; S3, import the training samples into the integrated algorithm model of text detection and classification for training, and respectively analyze the classification feature map and detection prediction feature map during the training process Carry out two types of negative sampling; S4, sequentially calculate the classification feature map loss value in the classification feature map and the detection prediction feature map loss value in the detection prediction feature map, and fuse the classification feature map loss value and the detection prediction feature map loss value to reverse Optimize the training text detection and classification integrated algorithm model to the propagation. Beneficial effect: the training of deep neural network can be supported by collecting a small number of samples.

The invention belongs to the technical field of intelligent automobile running environment perception, and discloses an automobile running working condition judgment system and a judgment method of the automobile running working condition judgment system. The method comprises the steps that the parameters of the speed and pedal opening degree of an automobile are selected to be treated, character parameters in speed signals and pedal signals representing the working condition are extracted, the character parameters are firstly reduced for the first time through a correlational analysis method, the first fifteen parameters with largest correlation and the working condition are selected for correlation analysis, ten character parameters are selected from large to small according to the correlation as the standard for dividing the working condition, then the ten character parameters are reduced for the second time through kernel principal component analysis, character parameter values are reduced to seven, and finally the running working condition of the automobile is classified based on the semi-supervised nucleus fuzzy C-mean value clustering analysis of an objective function. According to the system and method, the judgment speed is high and accurately, the fuel consumption and tail gas emissions in the automobile running process can be reduced through the judgment result, and the system and method have great significance in the study of various new automobile type development, dynamic property matching of the whole automobile and the like.

The invention discloses a multi-country banknote serial number character recognition method based on image fingerprints, which belongs to the technical field of banknote pattern recognition. Based on the template set and sample set of multi-country banknote serial number character pictures, the perceptual hash algorithm is used to generate training sample data, and the classification model based on SVM classifier is constructed, and the training sample data is used to train the classification model, and the trained classification model is used Identify the serial number character picture of the banknote to be processed. The method of the invention converts a single feature in the banknote serial number character recognition process into multiple features, and adopts a method based on an SVM classifier for recognition, which has a small demand for samples, fast training speed, fast recognition speed, and good performance Anti-interference ability can improve the error tolerance rate of prediction.

The invention discloses a hardware Trojan detection method based on a lifting algorithm. The first and second controllability values ​​and the observability values ​​of each node of the integrated circuit are obtained by processing according to the topology of the integrated circuit; the first and second controllability values ​​and the observability values ​​of each node of the integrated circuit are Based on density clustering, a matrix with class distances is obtained; a weak classifier is used for training classification, and the weak classifiers are combined to obtain a final classifier, and then a matrix X' with class distances obtained by processing the integrated circuit to be tested is used by the final classifier Processing is performed to obtain the hardware Trojan detection result of the integrated circuit to be tested. The present invention has higher accuracy than the bypass analysis method under the condition of large sample.

The invention discloses a method for automatically acquiring genetic knowledge for product concept design: 1. Using the first-order predicate logic method to analyze the function and behavior information of the target product, and using the decomposition and reconstruction principle to decompose the functional information into functional units; 2. From the product instance Find product instances of the same family as the target product in the library to form a sample set; 3. Cluster the samples into K categories, the value of K is set according to the type of target product, and calculate the Euclidean distance between the target product and the center point of each category, The class of the center point with the smallest distance is the class to which the target product belongs; 4. Extract the attribute information related to the functional unit from the samples of the class to which the target product belongs; 5. Use the K-fold cross-validation method to select Sixth, the selected m attributes and behavior information are stored in the product gene bank in the form of coding chains to obtain the genetic knowledge of the target product. The present invention can automatically obtain the information that affects the conceptual design process. key knowledge.

A regularization method for detecting line segment edges in an image of a non-textured metal component. The method comprises: calculating gradient values of respective pixels in an input image of a component; employing an LSD algorithm to detect preliminary line segments; and using distances and angle relationships between the line segments to connect discontinuous line segments. Thus, the invention detects complete line segments and avoids the problem of broken line segments when performing detection. The method is an improvement on LSD algorithms, and can be used to output complete line segments while also maintaining the fast speed of the LSD algorithms. Moreover, the method is applicable to both RGB images and grayscale images and so can meet the requirements of actual applications.

The invention discloses an image fingerprint-based recognition method for multi-national banknotes and belongs to the technical field of banknote pattern recognition. Based on the template set and sample set of the multi-national banknote background image, the perceptual hash algorithm is used to generate training sample data, and the classification model based on the SVM classifier is constructed, and the training sample data is used to train the classification model, and the trained classification model is used for processing Banknote images are used to predict and identify currency types. The method of the invention converts a single feature in the banknote identification process into multiple features, and adopts a method based on an SVM classifier for identification, which requires a small amount of samples, fast training speed, fast recognition speed, and has good anti-interference ability , which can improve the error tolerance rate of prediction.

The invention discloses a deep neural network structure design method inspired by an optimization algorithm. For a classical feedforward network structure in which all layers share the same linear and nonlinear transformations, the forward process in the feedforward network is equivalent to using The gradient descent method minimizes the iterative process of a function F(x); further adopts the re-ball method with faster convergence speed and the Nesterov acceleration algorithm to minimize the function F(x), thus obtaining a new network structure with better performance ; Can be used in artificial intelligence, computer vision and other application fields. Adopting the technical scheme of the present invention and designing the neural network structure based on the optimization algorithm can improve the traditional design method of relying on experience and experimental search, and obtain a more efficient neural network structure, thereby saving a lot of time and computing resources.

The invention provides a method for extracting and modeling the emotional information of a Chinese sound; the extracting method for the emotional information of the Chinese sound is that: formulate the specification of a emotional speech database, which includes the pronouncer specification, the recording play book design specification and the naming specification of audio files and so on; collect the emotional speech data; evaluate the validity of the emotional speech, namely, at least ten evaluators apart from a speaker carry out a subjective listen evaluation experiment on the emotional speech data. The modeling method of the emotional information of the Chinese sound is that: extract the emotional characteristics of the sound, define and distinguish the characteristic combination of each emotion type; adopt different characteristic combinations to train the SVM model of a multilevel sound emotion recognition system; verify the identification effect of the classifying models, namely, verify the classification effect of the multilevel classification models of sound emotion in a situation unrelated to the speaker by adopting a cross leave-one-out method. The method solves the problems that the domestic emotional speech databases are less in emotion type and the number of the domestic emotional speech database is very limited; at the same time, the method realizes an efficientspeech emotion identification system.

The invention discloses an image classification method of an equivariant convolution network model based on a partial differential operator. For the input layer and the middle layer of the convolution network model, the equivariant convolution and The equivariant convolution of the middle layer constructs the equivariant convolution network model PDO-eConv and performs model training; the input of the model PDO-eConv is image data, and the output is the prediction and classification of images, so as to realize efficient image classification, recognition and visual analysis . The present invention can provide a better parameter sharing mechanism and achieve a lower error rate of image classification.