53results about How to "No need for manual labeling" patented technology

A technical scheme of the invention discloses a method for obtaining a text classifier used for automatic corpus tagging, and the text classifier. The method comprises: determining a concept set, using a concept keyword corresponding to each concept to match a non-tagging corpus text and perform automatic tagging processing, the concept keyword being in a concept keyword set; for each concept, when the number of texts in a tagging corpus text set corresponding to the concept meets a threshold condition, training a corresponding text classification model for the concept, to obtain a corresponding text classifier, and finally a text classifier set corresponding to the concept is obtained, all text numbers meeting the threshold condition. The invention provides an algorithm structure having universality, and a classification system is flexibly changed, calculation time and resources are saved. The text classifier just needs few initial corpus texts, and the text classifier automatically tags without manual tagging, so as to further save time and cost.

The invention discloses a video abstract generation method and a related device, which are applied to a machine learning technology of artificial intelligence. The method comprises: acquiring a targetvideo; extracting image features of each video frame to generate an image feature sequence; dividing a video frame sequence into a plurality of target segment sequences to generate a paragraph feature sequence; fusing the image feature sequence and the paragraph feature sequence to obtain fused features; and carrying out reinforcement learning training based on the fusion features to generate a video abstract. According to the invention, an unsupervised video abstract generation process is realized, the video abstract generation quality is ensured by combining the image features of the picture level and the paragraph features of the depth level, and the unsupervised video abstract generation process does not need manual annotation, so that the video abstract generation efficiency and accuracy are improved.

The invention discloses a new energy lithium battery surface defect detection method based on adaptive deep learning. The method comprises the following steps: carrying out nonlinear mapping on a lithium battery surface grayscale image; transforming the decoupled irradiation component and reflection component to a frequency domain; performing filtering, inverse Fourier transform and exponential transform on the frequency domain data to obtain a reconstructed lithium battery image; based on morphological processing and background differencing, enhancing gray scale response at the defect; carrying out image segmentation and connected domain analysis and screening processing, and taking a result as a labeled image; designing an operator to simulate illumination details, and carrying out sample enhancement operation on the surface grayscale image of the lithium battery; training a deep convolutional neural network based on the enhanced sample image set and the labeled image; and achievinglithium battery surface defect detection based on the trained network. By utilizing the method, the detection efficiency can be improved and the false detection rate can be reduced in a lithium battery surface defect detection scene.

The invention discloses an extraction type document automatic abstracting method based on context semantic perception. The extraction type document automatic abstracting method mainly solves the problem that a traditional algorithm lacks the recognition degree of sentences in different contexts. The method comprises: firstly, using an LDA topic model for calculating topic probability distributionin a document, and then determining the similarity between each sentence and a topic word; extracting semantic features of sentences by using a CNN model, further calculating the similarity between each sentence and the features, finally adding values of topic similarity and feature similarity of each sentence to obtain a final sentence score, and taking a proper number of sentences as abstracts according to score ranking. According to the method, a topic model and a deep learning model are introduced, a topic abstracting method is determined, sentence meanings in different contexts can be analyzed more accurately, and a calculation reference method is provided for other automatic document abstracting methods.

One embodiment of the invention provides an audio data processing method. The audio data processing method includes the steps: acquiring a spectrum corresponding to audio data; dividing the spectrum corresponding to the audio data into a homophonic tone spectrum and a non-homophonic tone spectrum, wherein the frequency in the homophonic tone spectrum is an integer multiple of the frequency of thefundamental tone; and determining the emotional information expressed by the audio data according to the characteristic information of the homophonic tone spectrum and the characteristic information of the non-homophonic tone spectrum. The audio data processing method can accurately identify the emotion expressed by the audio so as to enable a user to search the related audio of expressing the corresponding emotion according to the emotion dimension so as to bring a better experience for the user, by dividing the spectrum corresponding to the audio data into the homophonic tone spectrum and the non-homophonic tone spectrum, and determining the emotion information expressed by the audio data according to the characteristic information of the homophonic tone spectrum and the characteristic information of the non-homophonic tone spectrum. Besides, the embodiment of the invention provides an audio data processing device, a medium and a computing device.

The invention provides a gang fraud phone recognition method based on suspicion measurement, which adopts an unsupervised mode, utilizes call record information between phone numbers to define suspicion measurement for each phone number, and adopts a risk level quantification mode to recognize a gang fraud phone. According to the method, a subjective marking result of a user does not need to be introduced, defects caused by subjective judgment can be avoided, model training and online use can be completed within a short time, a simple and practical recognition method is provided for the user to suffer from group fraud, and the problem caused by lagging is effectively relieved.

The invention discloses a retrieval method of color images based on quaternion Harmonic-Fourier moments. The method includes expanding Harmonic-Fourier moments according to the principle that quaternions contain color information, expanding the Harmonic-Fourier moments to quaternion level, acquiring a new image-describing feature vector according to the character that the Harmonic-Fourier moments can describe shape features of images, and designing a new retrieval method of color images based on quaternion Harmonic-Fourier moments. Experimental results show that the system has such features that the system can effectively represent the colors and shape features of the images, the design is simple, implementing is easy, and image feature extraction needs no manual annotation; the manual annotation process tedious and error prone is avoided, and retrieval precision and efficiency is greatly improved.

The invention discloses an intelligent speech recognition method and device based on classification identification and related equipment, and the method comprises the steps: carrying out the endpoint detection and feature extraction of original speech data, so as to obtain speech feature data; identifying the industry attribute of the original voice data through the attached industry attribute identifier when the original voice data is input; recognizing the voice feature data through a trained universal voice recognition model to obtain a recognition result; judging whether the confidence coefficient in the recognition result is lower than a preset value or not; inputting the corresponding text of which the confidence coefficient is lower than a preset value into the corresponding industry vertical module for optimization identification; and selecting the text corresponding to the maximum confidence coefficient as a final recognition result according to the confidence coefficient in the recognition result and the optimized confidence coefficient. The industry attribute identifier is attached when the original voice data is input, so that a corresponding industry vertical module can be conveniently called subsequently to carry out optimization identification; manual labeling is not needed, and high accuracy is maintained through secondary verification, recognition and correction.