39results about How to "Enhanced feature information" patented technology

The invention discloses a sentiment classification method and system based on multi-modal context semantic features. The method comprises the following steps: segmenting a short video into semantic units with the same number by taking an utterance as a unit, generating corresponding video, voice and text samples, and extracting three characterization features, namely an expression feature, a spectrogram and a sentence vector; respectively inputting the three extracted characterization features into expression, voice and text emotion feature encoders, and extracting corresponding emotion semantic features; constructing corresponding adjacent matrixes by using context relationships of emotion semantic features of expressions, voices and texts; and inputting the expression emotion semantic features, the voice emotion semantic features, the text emotion semantic features and the corresponding adjacent matrixes into corresponding graph convolutional neural networks, extracting corresponding context emotion semantic features, and fusing the context emotion semantic features to obtain multi-modal emotion features for emotion classification and recognition. According to the method, the context relationship of the emotion semantic features is better utilized through the graph convolutional neural network, and the accuracy of emotion classification can be effectively improved.

The invention discloses an interpretable CNN image classification model-based optical remote sensing image classification method. An explainable CNN ResNet model composed of six groups of 51 basic convolution layers, three full connection layers and one Softmax layer in total is built; the method comprises the following steps of: performing down-sampling by utilizing ResNet; obtaining a feature map containing context information; and then interpretable modification is performed on the ResNet model to obtain a new interpretable CNN network based on the ResNet model, and the ResNet model extracts features through multiple groups of convolution-pooling layers with residual modules and finally inputs the features to a full connection layer to classify the images. The interpretability of an existing deep learning model can be enhanced, and the performance of the model is further improved.

The invention discloses a mycobacterium tuberculosis image segmentation identification method based on an optimization watershed and an apparatus thereof. The method comprises the following steps of acquiring a mycobacterium tuberculosis RGB image, converting a mycobacterium tuberculosis image RGB mode into a Lab mode, and calculating a die image g of a gradient of an a component; according to a value distribution condition of the die image g of the gradient, deleting a background and reserving a mycobacterium tuberculosis boundary, carrying out boundary line segment connection at a mycobacterium tuberculosis boundary fracture position, carrying out a watershed algorithm on an acquired image so as to acquire a watershed image L; and successively determining each area of the watershed image L, according to color information and mycobacterium tuberculosis shape information, deleting an impurity area and reserving the mycobacterium tuberculosis. In the invention, mycobacterium tuberculosis classification identification precision and a speed under microimaging can be effectively increased.

The invention relates to a dynamic video tactile feature extraction and rendering method, and belongs to the field of virtual reality and human-computer interaction. The method comprises the steps ofdecompressing a received video, preprocessing the video, segmenting shots based on inter-frame color histogram features, extracting saliency maps fused with time-space domain tactile saliency featuresfrom all the segmented frames in each shot, and performing pixel-level tactile rendering according to the saliency maps of the video frames. According to the method, the video content is divided intothe salient region and the non-salient region by extracting the video frame salient features fused with the space-time characteristics. And pixel-level tactile stimulation is applied to the video frame by adopting a one-to-one mapping relationship between the visual channels and the tactile channels. And real-time tactile feedback is generated through the terminal, so that the realistic experience of watching the video by the user is enriched. The method can be widely applied to video education, multimedia entertainment and human-computer interaction.

The invention discloses a low-overhead household garbage classification method based on a dense convolutional network, and the method comprises the steps: (1), preprocessing data, (2), building the dense convolutional network, (3), dividing a data set into a training set, a verification set and a test set, and (4), selecting a proper optimizer and a loss function in a training process, setting hyper-parameters, and setting an evaluation index as the accuracy rate. According to the method, optimization is carried out in preprocessing of input data, matrix fusion is carried out on a three-channel color image and an edge detection image to serve as input of a model, and feature information is enhanced; a dense convolutional network structure is constructed; a Dropout layer is additionally arranged; a learning rate self-adjusting and hyper-parameter adjusting method is used.According to the method, the model has enough feature extraction capability; the feature mapping of the model is usedas the input of a subsequent layer; the problem of gradient disappearance caused by a deep network is relieved; good balance is realized on the aspects of low overhead and high precision; and 90.8% of precision and 5.08 M of file size are realized.

A distribution line pin defect detection method based on improved ALI and Faster-RCNN comprises the following steps: collecting pin defect images of a distribution line, manually marking the pin defect image data, and constructing a training sample set; building network structures, wherein the first network structure is an adversarial learning inference network, the basic structure of the first network structure is composed of an inference network, a generative network and a judger, and the second network structure is a Faster-RCNN network; carrying out detection model training according to the obtained training sample, and completing training after training for a specified number of steps; and inputting an image to be detected into the trained adversarial learning inference network, outputting a high-quality reconstructed pin image, and finally completing defect identification through the trained Faster-RCNN network. According to the method, detail information such as local textures and edges of the distribution line pin image can be enhanced to improve image quality, and accurate features are extracted in combination with a target detection algorithm to realize intelligent detection of pin defects.

The invention relates to a double-flow face forgery detection method based on Swin Transformer, which is used for detecting a face forgery image by using deep learning. A deep learning network model is integrally built, and the network model is divided into three parts: a double-flow network, a feature extraction network and a classifier. Because all face counterfeit data sets disclosed at present are videos, the videos need to be clipped into frame pictures by using OpenCV. In addition, the frame picture contains a large amount of background information, so that a human face area needs to be cut out by using a human face positioning algorithm. And inputting the obtained face region image into a double-flow network and a feature extraction network to extract and learn features. And finally, inputting the learned features into a classifier, and identifying whether the face image is true or false. The method is used for solving the problem of partial limitation, namely weak generalization ability, of the existing face counterfeiting detection scheme, and meanwhile, the compression resistance of the model is improved through the double-flow framework, so that the method is more in line with the common face video quality in daily life.

The invention belongs to the field of medical equipment, and particularly relates to a handheld three-dimensional scanner-based spatial registration method for surgical navigation. The method adopts facial recognition feature point matching registration to perform preliminary registration on space point clouds of a to-be-tested person obtained by the handheld three-dimensional scanner and image space point clouds obtained by a medical image; a preliminary registration result uses a point cloud iterative registration method to complete the registration of the space of the to-be-tested person ofsurgical navigation and the image space; implementation results show that the method has fast registration speed and stable registration precision, and can meet the requirements of a surgical navigation system and improve the registration accuracy of the surgical navigation system.

The invention discloses a radar radiation source identification method based on a multistage jumper residual network, and belongs to the field of image processing. The radar radiation source identification method comprises the following steps: performing time-frequency transformation on a radar radiation source signal to generate a time-frequency image of the radar radiation source signal; carrying out the preprocessing of the time frequency image through an ImageDataGenerator, so as to obtain a one-dimensional time frequency image; and inputting the one-dimensional time-frequency image into a deep residual network for feature extraction and signal classification so as to identify and obtain the radar radiation source. According to the radar radiation source identification method based on the multistage jumper residual network provided by the invention, eight sequentially connected residual blocks are arranged in a deep residual network and are connected by using jumpers to form four residual units, so that the constructed deep residual network with the total convolution layer number of 18 can extract deep information of a signal time-frequency image, and meanwhile, the problems of gradient disappearance, gradient explosion and the like of the network are also avoided.

The invention discloses an artificial intelligence-based pest control system. The system comprises a wireless visual front end, a communication module and a cloud server, wherein the wireless visual front end comprises a solar photovoltaic panel, a solar charging control circuit, a lithium battery, an MCU, a GPS positioning module and an image acquisition module. The the image acquisition module shoots a high-definition image of a pest sticking plate placed in a tea garden at regular time and sends the high-definition image to the MCU; the MCU sends the high-definition image to the cloud server through the communication module; the cloud server receives high-definition image data, analyzes the types and the number of pests in the tea garden based on artificial intelligence, and provides prediction and forecast of the pest condition; and in the process, the MCU controls the solar charging control circuit to realize charging and discharging of the lithium battery by the solar photovoltaic panel. The system has the advantages of high automation degree, strong real-time performance and accurate prevention and control.

The invention relates to a low-illumination image target detection method based on image fusion and a target detection network, and the method combines a non-subsampled shearlet transform image fusionalgorithm and an improved target detection network YOLOV4, and comprises the steps: 1) employing an infrared and low-illumination visible light image fusion technology, so that the information complementation can be performed on infrared and low-illumination visible light images, and texture information is highlighted while the contour definition is improved; 2) sending the fused image to an improved YOLOV4 target detection network for detection, and outputting target information under a low-illumination condition; and 3) in order to improve the feature extraction capability of the YOLOV4 network, replacing the residual blocks in the YOLOV4 backbone network with dense link blocks, wherein compared with the residual blocks, the dense link blocks enable the network to improve the feature expression capability and improve the feature extraction capability of the network; finally, experiments prove that the method can improve the target detection capability under the condition of low illumination.

The invention discloses a method and device for determining an optimal verbal skill sequence and a storage medium. The method comprises the following steps: determining user feature information and dialing information of a target user; by utilizing a pre-trained prediction model, according to the user feature information and the dialing information, respectively determining a predicted hang-up rate and a predicted hang-up conversion rate of each verbal skill in the plurality of verbal skill sequences and interaction process information of interaction with the target user, wherein the interaction process information comprises a current verbal skill sequence interacting with the target user, a verbal skill subsequence interacting with the target user and a subsequent verbal skill sequence interacting with the target user; forming a plurality of verbal skill sequences by verbal skill corresponding to each verbal skill node; and determining an optimal verbal skill sequence interacting with the target user according to the predicted hanging-up rate and the predicted hanging-up conversion rate of each verbal skill in the plurality of verbal skill sequences.

The invention discloses an ultrasonic damage judging method and system for an embedded part of a train axle. The ultrasonic damage judging method comprises the steps: acquiring a single-angle ultrasonic B-scan image of the embedded part of a train wheel; decomposing the ultrasonic B-scan image into a plurality of first images with different scales, calculating the significance value of each pixel block of the plurality of first images, and generating a target area image; and repeating the step S1 and step S2 until a plurality of target area images at all angles are generated, and extracting an overlapping area image of the plurality of target area images as a defect area. According to the method, multiple multi-angle ultrasonic B-scan images of the train wheel embedding part are collected, feature information of wheel set noise and defects is increased, Gaussian pyramid image decomposition is carried out on the ultrasonic B-scan images to carry out significance detection, target defect features are further enhanced, background noise is further weakened, and the detection accuracy is improved. Moreover, an accurate defect area is extracted, a detection result is generated, and the problem that the reliability of the detection result is low in an existing ultrasonic damage judgment method for the axle embedding part is solved.

The invention relates to the field of neural networks, in particular to a vector fusion calculation method and device applied to neural network data processing, and the method and device adopt an instruction set control mode to encode two different Route operators and Shartcut operators into different parameters in an instruction set, and control the operation of a system through a state machine mode. The SRAM of the instruction set stores the function information of each layer of the network, then decodes the function information in the decoding state of the state machine, and when the operator type of the instruction set is decoded to be Route or Shartcut, corresponding operator processing is carried out. As the Route operator and the Shartcut operator both carry out data processing on the two intermediate layers in the network layer, no matter splicing or corresponding position data addition is carried out, the feature information of the output feature map is increased, and the subsequent processing precision of the image can be improved.

The invention belongs to the technical field of computer vision. The invention discloses a facial expression recognition method and system, a storage medium, a computer program and a terminal, and themethod comprises the steps: pre-training an image generation model according to the combination of a given depth image and an RGB image, and enabling the trained image generation model to convert aninput depth image into an RGB image according to an RGB image style used for training; generating eyebrows, eyes and mouths of expressions in the RGB images, training a convolutional neural network considering the eyebrows, the eyes and the mouths, and achieving expression recognition through the convolutional neural network. According to the invention, the feature information of eyes, eyebrows and mouths is enhanced, and the recognition accuracy is higher; the effect of the image generation model is good, important information about expressions is reserved through the image generation model,and the forms of RGB images used for expression recognition are unified; the accuracy of expression recognition is higher; when only one channel of the depth map is used for identification, the effectobtained by the method is better.