59 results about "Convolution" patented technology

The invention provides a relevant filtering opposite-thrust target tracking method with an adaptive scale. The method comprises: an initial position and an initial scale of a to-be-tracked target in a video frame are determined, and convolution feature graphs of different layers are extracted respectively by using the initial position as the center and using a deep convolutional neural network; for the extracted convolution feature graph of each layer, tracking is carried out by using a kernel-correlation filtering tracking method to obtain a tracking result; all tracking results are combined by using an adaptive hedging algorithm to obtain a final tracking result as a final position of the to-be-tracked target, so that the to-be-tracked target in the video frame can be localized; after obtaining of the final position of the to-be-tracked target, a final scale of the to-be-tracked target is estimated by using a scale pyramid strategy; and after obtaining of the final position and the final scale of the to-be-tracked target, a to-be-tracked target image block is extracted based on the final scale by using the final position as a center and each kernel-correlation filtering tracking method is trained again to update a coefficient and a template.

The invention discloses a segmentation method of pathological section unconventional cells based on a multi-scale hybrid segmentation model. The method comprises steps that positive and negative samples are respectively scaled to low-resolution, medium-resolution and high-resolution images, the full convolutional network algorithm is utilized for training to acquire a convergent low-resolution segmentation model, a medium-resolution segmentation model and a high-resolution model; a multi-scale hybrid segmentation model is acquired through fusion by a model integration method; after the effective discriminating area of a new pathological section is processed through utilizing the data enhancement method during testing, the processed effective discriminating area is inputted to the multi-scale hybrid segmentation model, the probability of each pixel in the effective segmentation area is outputted, pixels with probability values greater than the threshold t are taken as abnormal cell pixels and are recorded as 1, remaining pixels are taken as normal cell pixels and are recorded as 0, binary images predicted by the multi-scale hybrid segmentation model are acquired, and post-processingon the binary images is carried out to acquire the final segmentation result. The method is advantaged in that high precision is realized, and a Dice value is above 0.869.

The invention discloses a convolutional neural network optimization and rapid target detection method and device. The optimization method comprises the steps of obtaining a first convolutional model trained by adopting floating point type sample image data; wherein a BN layer operation in the first convolution model is located behind convolution layer operation and is used for normalizing a numerical value obtained after the convolution layer operation to a preset data range; adjusting parameters of a convolutional layer according to the parameters of the BN layer; deleting the BN layer to obtain an adjusted second convolution model; adding a quantization layer before the convolution operation of the convolution layer in the second convolution model to obtain a third convolution model; wherein the quantization layer is used for quantizing input data to integer data of a predetermined bit, and enabling a convolution layer to perform convolution operation by using the integer data of thepredetermined bit. According to the method, the problems of large model body quantity and low operation speed are solved by optimizing the convolution operation.

The invention belongs to the field of medical signal processing, particularly provides an eeg high-frequency oscillation signal detecting system based on a convolution neural network, and aims to solve the problem that in a conventional eeg high-frequency oscillation signal detection technique, the false drop rate is high caused by high-frequency noise and peaked wave shape. The eeg high-frequency oscillation signal detecting system comprises a user terminal, a data preprocessing module, a high-frequency oscillation signal predetecting module, the convolution neural network module and a static module, wherein the user terminal is used for acquiring an eeg signal; the data preprocessing module is used for performing data preprocessing; the high-frequency oscillation signal predetecting module is used for performing detection on the eeg signal to obtain suspected high-frequency oscillation fragments; and the convolution neural network module is used for classifying all the suspected high-frequency oscillation fragments. According to the eeg high-frequency oscillation signal detecting system disclosed by the invention, when HFOs are detected, the sensitivity can be effectively improved, the false drop rate is reduced, and the precision rate of epileptogenic focus positioning is increased; and besides, the eeg high-frequency oscillation signal detecting system can use scalp eeg and can also use cortex eeg, so that the egg high-frequency oscillation signal detecting system can enable a doctor to achieve more flexibility.

The invention relates to a deep neural network target detection method integrated with a pixel-level attention mechanism. The deep neural network target detection method comprises the following steps:collecting a training sample image; converting data in the collected image data set into a format required by training the deep convolutional neural network through preprocessing, and performing dataenhancement to improve the training effect of the neural network; designing a deep convolutional neural network structure integrated with a pixel-level attention mechanism, the network structure being used for receiving an input image and outputting bounding box regression information and category information to an object in the image: the network structure comprising two parts, one part being abasic network for preliminarily extracting features; the other part being a plurality of cascaded convolutional layers or downsampling layers added at the tail end of the basic network and being usedfor extracting convolutional feature maps with different scales and fusing the multi-scale feature maps with a pixel-level attention mechanism; and training.

The invention discloses a steel bar detection method based on a deep convolutional neural network. The method comprises the following steps: 1) data acquisition; 2) data processing; 3) constructing and training of a model; 4) model evaluation; 5) model deployment. The target detection algorithm based on the deep convolutional neural network is applied to reinforcing steel bar detection and counting, and a multi-scale and deformation-tolerant reinforcing steel bar detection network framework is provided. Cascade R-CNN with good detection performance, a feature pyramid network capable of effectively solving the multi-scale detection problem, a more stable group normalization module, and a deformable convolution module capable of improving the spatial geometric deformation capacity of network learning and other modules are integrated in the frame. Compared with a traditional steel bar detection method, the network framework is higher in detection precision and better in generalization, and can be deployed in a remote server or a mobile device.

The invention provides a malicious software API (Application Program Interface) call sequence detection method based on graph convolution. The method comprises the following steps: acquiring and recording API call sequence information of processes and sub-processes when a large number of software samples run; performing vectorization processing on the API calling sequence information; extracting aparameter relationship, a dependency relationship and a sequence relationship of the API function; establishing an API call graph; inputting the API call graph into a graph convolutional neural network for training to obtain a malicious software detection network model; collecting API calling sequence information of processes and sub-processes when the executable file to be detected runs; constructing an API call graph of the executable file to be detected, then inputting the API call graph of the executable file to be detected into the malicious software detection network model, If the output result of the malicious software detection network model is 1, indicating that the judgment result is malicious software; If the output result of the malicious software detection network model is 0,indicating that the judgment result is normal software.

The invention discloses a method, system and device for compiling an AI chip and a storage medium, and the method comprises the steps of carrying out the quantification of a network model of the AI chip, and carrying out the precision adjustment of the quantified network model based on the weight of the network model; allocating calculation of the convolutional neural network to different MAC calculation units of the AI chip in the network model after precision adjustment so as to optimize a calculation process of the network model; and acquiring hardware architecture parameters, and generating a continuous calculation flow according to the hardware architecture parameters and the network model. According to the invention, the AI chip is quantified and the precision is adjusted, so that the performance is more perfect; the calculation of the network is optimized to the maximum extent, the calculation steps are reduced, and the utilization rate of the MAC calculation unit is maximized;the utilization rate of the AI equipment is improved by generating the calculation flow capable of enabling the calculation equipment to be always in the busy state in the network reasoning process.

The invention relates to a self-adaptive optical image high resolution restoration method combined with frame selection and blind deconvolution, comprising the following steps: firstly, a short exposure image sequence gn (x, y) is recorded when a self-adaptive optical closed loop is corrected; shannnon entropy of each frame of image in the sequence is calculated; a degraded image gm (x, y) with lower entropy is selected for blind deconvolution image restoration; secondly, an initial value hm (x, y) of a point spread function is generated by utilization of random phase; thirdly, a target f(x, y) is estimated by using the gm(x, y) and the obtained hm (x, y), and an estimated value f(x, y) is obtained after addition of positivity limitation on the target; fourthly, an estimated value hm (x, y) of the point spread function is obtained by using the gm (x, y) and the f(x, y), and an estimated value h(x, y) is obtained after addition of positivity limitation in the same way; fifthly, inspection is made whether an iterated value h(x, y) and an iterated value f(x, y) meet iteration stopping requirements or not; if the iterated values do not meet the iteration stopping requirements, the third step is returned; if the iterated values meet the iteration stopping requirements, circulation is stopped and the f(x, y) and the h(x, y) are outputted. The invention has the advantages of effective improvement of restoration quality, acceleration of convergence rate, capability of well compensating correction capability under hardware limitation of a self-adaptive optical system, and improvement of imaging quality.

The invention discloses a scalp electroencephalogram feature extraction and classification method based on an end-to-end convolutional neural network, and the method comprises the steps: carrying out the data enhancement of training data, and enabling the enhanced training data to train the convolutional neural network; inputting the to-be-detected data into the convolutional neural network for feature extraction and classification. The feature extraction and classification method comprises the following steps: S1, filtering an original scalp electroencephalogram signal by using a band-pass filter to obtain signals xtheta, xmu and xbeta; S2, performing multi-scale time convolution and spatial convolution on the signals xtheta, xmu and xbeta respectively to extract features; s3, performing pooling operation on the feature map output by the convolutional layer; s4, after pooling, carrying out feature fusion, and then sending the feature fusion to a full connection layer to integrate the input abstract features; and S5, sending the output of the full connection layer to a softmax layer for classification. According to the method, a brand new data enhancement technology is applied in a training stage, data is input into a plurality of convolutional neural network branches for multi-scale convolution operation after passing through a filter bank in a test stage, the overfitting phenomenon is reduced, and the classification accuracy is improved.

The invention relates to the technical field of image processing and discloses an image fusion method and a device, and the method comprises the following steps: inputting source images to be fused; respectively carrying out block division treatment on all the source images, and generating block-shaped image matrixes which correspond to the source images; calculating a fusion coefficient matrix of all the block-shaped images in all the source images, carrying out interpolation treatment on the fusion coefficient matrix of all the block-shaped images through the cubic convolution interpolation method, and obtaining the fusion coefficient matrix which corresponds to the source images; and carrying out weighted fusion on the source images by utilizing the fusion coefficient matrix of the source images and generating an image after fusion. The utilization of the invention can lead the image after the fusion to fully reflect the respective main information of all the source images.

One embodiment provides a system for determining a process model for a photolithography process. The photolithography process can use multiple exposure-and-development steps to create features on a wafer. When the photolithography process exposes the wafer to a layout, the wafer can include topography variations which were caused by previous exposure-and-development steps. The process model can be used to predict patterns that are created on the wafer when the wafer is exposed to a second layout, wherein the wafer includes topography variations that were caused by resist features that were created when the wafer was exposed to a first layout. The process model can include a first term and a second term, wherein the first term is convolved with a sum of the first layout and the second layout, and wherein the second term is convolved with the second layout.

The invention discloses a stainless steel weld defect efficient detection method based on multi-domain expression data enhancement and model self-optimization. The method comprises the following steps: deriving a one-dimensional echo time domain signal to spatial domains such as a time-frequency domain, a Gramb angle field domain and a Markov transfer field domain; sequentially inputting the data set constructed by each spatial domain into the MobileNetV3 neural network, and selecting the spatial domain with the most abundant feature expression as a final training data set; constructing a multi-scale depth separable convolution to improve the MobileNetV3 so as to enhance the recognition performance of the network; providing a particle swarm-chaos sparrow search algorithm for automatic optimization of a network structure and parameters; and adopting the CPU + FPGA heterogeneous cooperative calculation to accelerate the reasoning and recognition application speed of the defects. Five types of weld defects such as incomplete fusion, air holes, slag inclusion, incomplete penetration and cracks are taken as objects, the recognition accuracy of the five types of weld defects can reach 98.75%, and the method has practical engineering application value.

The invention discloses a track circuit fault automatic diagnosis method based on a deep convolutional neural network, and the method comprises the steps: firstly obtaining the real-time change data of a track circuit, building a track circuit fault type data set, and dividing the data set into a training sample and a test sample; secondly, establishing a one-dimensional deep convolutional neuralnetwork, inputting the training set sample to train a network model, and evaluating the accuracy of the network model through a test sample to obtain a trained one-dimensional deep convolutional neural network model; and finally, inputting the real-time change data of the track circuit into the trained deep convolutional neural network model, establishing a mapping relationship between input and output through a neural network, outputting a fault type, and positioning a fault position. According to the method, the fault diagnosis efficiency and accuracy of the track circuit can be improved, field maintenance personnel can quickly process faults, and the maintenance and repair efficiency and the intelligent level of the track circuit are improved.

The method of the present invention makes it possible to recognize partial discharges acquired by means of sensors in electrical networks, comprising a series of steps, among which are a post-processing step (13) of the acquired signals and a recognition step (17) of said signals by means of a convolutional neural network (CNN). The method also includes adaptation (15) and training (16) steps of the neural network, as well as a step to build a library (14) of partial discharge signals from known sources that serve as training of the convolutional neural network (CNN).

The invention discloses a clutch main cylinder groove surface defect detection method based on deep learning. Deep learning and defect recognition are combined to achieve high-precision and high-efficiency automatic detection. A multi-scale feature fusion pyramid network is adopted to extract defect features more accurately, and autonomous learning is carried out; a single-target FCOS convolutional neural network is adopted as a detection network model, dependence on an anchor box and a feature candidate region is avoided, the network model is simplified, and the network training speed is increased; experimental results show that the effect of the defect detection algorithm based on deep learning is superior to that of a traditional manual detection method and a machine vision detection method, and the time requirement and precision of defect detection of the clutch main cylinder inner groove on an industrial production line are met; defects generated in the production process of the clutch main cylinder inner groove can be rapidly and accurately detected, the production efficiency of a factory is improved, and the practical significance is achieved.

The invention discloses a method for super-resolution reconstruction of a license plate image. Based on the SRCNN convolution network, the algorithm of the invention adopts a multi-scale filter parallel convolution method for the first layer of the network, thereby adding image features, then the second layer is subjected to nonlinear mapping and the third layer is reconstructed. The experimentalresults show that the method for super-resolution reconstruction of the license plate image has an average improvement of 3dB on the PSNR value compared with the existing method. Therefore, the methodfor super-resolution reconstruction is effective. In the actual criminal investigation image, the acquired license plate image is small in size, and only 3 times magnification is used to reconstructthe image, which cannot meet the actual demand. High-magnification reconstruction of the license plate image is an issue that needs further study in the future.

The invention discloses an AC motor bearing fault diagnosis method adopting a convolutional neural network and a bidirectional long-short term memory network. According to the method, the advantages of parameter sharing, local sensing, downsampling and the like of the convolutional neural network are fully utilized, corresponding spatial features are effectively extracted from original current data, and a complex feature extraction process is avoided. And then the extracted current data spatial features are input into a bidirectional long-short-term memory network, time sequence information of the current data spatial features is captured, and the rolling bearing fault diagnosis accuracy is further improved. According to the invention, the stator current obtained in the motor driving process is used as a fault signal, a non-intrusive fusion system which is easy to form a closed-loop'driving-diagnosis' is provided, and the monitoring cost can be effectively reduced; the bearing fault diagnosis method combines the characteristics of the convolutional neural network and the bidirectional long and short time memory neural network, has good performance in depth and complexity of feature extraction, and can realize accurate and effective extraction of fault features.

The invention provides an image feature extraction system and method based on Talbot effect optical convolution. An optical convolution image processing system is composed of an optical system, a laminated grating, a time-sharing switching system and a detector. The optical system comprises an objective lens, an ocular lens system and a source image and is used for realizing conjugate reduction imaging of an object image. The laminated grating system can generate a Talbot self-imaging effect on a grating Talbot distance, and then optical convolution processing can be carried out on an incidentlight field by utilizing a Moire fringe effect. The time-sharing switching system can realize time-sharing switching of gratings with different periods and different phases, thereby realizing opticalconvolution operation with different convolution parameters and extracting different incident image features. According to the method, image processing is carried out by using optical convolution, sothat the calculation energy consumption can be obviously reduced and the calculation time is saved. According to the scheme, the structure system is compact, and operations such as image filtering can be performed by effectively utilizing optical convolution.