18results about How to "Reduce model parameters" patented technology

The invention relates to an electromagnetic inverse scattering imaging method based on an SERU-Net model. The method comprises the following steps: acquiring scattered field information based on antennas arranged around an imaging area to obtain an actual target initial image; then, based on the MNIST handwritten digital picture set, generating an imaging area virtual target to obtain a virtual target initial image; then, taking the virtual target as an SERU-Net deep learning network label value, taking the virtual target initial image as an SERU-Net input value, and constructing an electromagnetic imaging prediction model based on an SERU-Net model; and finally, inputting the actual target initial image into the electromagnetic imaging prediction model to obtain an actual target prediction imaging result. According to the method, the imaging problem under the condition of low signal-to-noise ratio is efficiently processed in a short time, the problems that a classic network has a poor imaging effect on a weak scatterer and ghosting may occur in imaging of a strong scatterer are obviously solved, and the imaging precision is greatly improved.

The invention discloses a SobelConv and CoordGate-based flaw feature extraction method, and the method comprises the following steps: obtaining an original feature map, and constructing a flaw feature extraction method of double-branch parallel and left and right branch information interaction; in a left branch of the defect feature extraction method, an improved deformable convolution DConv and a designed calculation space convolution CoordGate block are used for carrying out adaptive reinforcement learning on defect features of SobeConv; in a right branch of the flaw feature extraction method, a designed calculation space variation convolution CoordGate block stack is used for learning flaw local space features; fusing defect features through multiple times of double-branch information interaction and feature splicing, and then constructing a residual connection layer to fuse with originally input defect features; jointly constructing a new flaw detection model by using the flaw feature extraction method and the RT-DETR; and constructing a flaw image data set, training and testing the flaw image data set by using the new flaw detection model, and observing a flaw detection result. The objective of the invention is to reduce the model parameter quantity, improve the calculation efficiency, reduce the calculation resource consumption, and improve the recognition rate of the defect detection model under the complex background.

The invention discloses an information generation model training method and an information generation method. The training method of the information generation model comprises the steps of obtaining sample data for a basic generation model, constructing a training instruction based on the sample data, inputting the sample data and the training instruction into the basic generation model to obtain a test result, adjusting the training instruction according to the test result to obtain a target training instruction, and training the target training instruction according to the target training instruction. Performing compression processing on model parameters of the basic generation model to obtain a to-be-trained information generation model; inputting the target training instruction and the sample data into an information generation model, and determining a prediction result; and training the information generation model according to the prediction result. According to the scheme, the information generation model with high performance can be trained on the basis of a small number of training samples, the requirement of model performance for the number of samples is remarkably reduced, and the accuracy of the model is effectively improved.

The embodiment of the application relates to the technical field of image processing, and discloses an image comparison method, system, electronic equipment and computer readable storage medium, the method comprising: respectively pre-processing and adaptively binarizing a test image and a template image corresponding to the test image to obtain a first image and a second image; performing target detection on the first image, respectively removing a detected target region in the first image and a region in the second image with the same position as the target region to obtain a third image and a fourth image; extracting feature points in the third image and the fourth image and performing feature point matching to determine matched feature points; performing global binarization on the pre-processed test image and the template image and performing difference processing to obtain a difference image; and obtaining an image comparison result based on the matched feature points and the difference image, so that different elements between images are efficiently identified, elements that do not need to be focused on are excluded, and the workload of a tester during rechecking is reduced.

The embodiment of the present application provides a training method and system of a self-supervised training predictor for speech separation. The method comprises: extracting self-supervised training features of each single person voice source speech by using a pre-training model; extracting shallow features for speech representation and deep features for context information in the self-supervised training features, and determining the shallow features and the deep features of each single person voice source speech as training labels of the self-supervised training predictor; inputting training mixed speech generated by each single person voice source speech into the self-supervised training predictor to obtain estimated features of each single person voice source speech; and training the self-supervised training predictor based on a loss function determined based on the estimated features and the training labels corresponding to each single person voice source speech. The self-supervised training predictor is trained and applied in a speech separation model in the embodiment of the present application, so that the accuracy of the self-supervised training features is improved, the performance of the speech separation system is improved, and the model parameters and the calculation complexity are reduced.

The application relates to a video instance segmentation method based on a dynamic convolution decomposition of a lightweight attention mechanism; the method first inputs a video frame, and a backbone network independently extracts a feature map of each frame in the video; then the feature map extracted by the backbone network enters an HQT encoding and decoding module, the position change of an instance in each frame is accurately positioned by combining an output head, and three prediction branches are used to supervise model training; in the application, a convolution layer of an encoding network adopts dynamic convolution decomposition, a more compact model is obtained, model training is easier, the required parameter quantity is greatly reduced, the training speed is improved, and the training time is shortened; a lightweight HQT encoding and decoding module is provided, which helps to reduce model parameters and improve efficiency; the loss function is improved, model training is more stable, the convergence speed and convergence precision are improved, and the problems of imbalance between foreground and background in samples and imbalance between foreground categories under a long tail condition are relieved.

The invention provides a mask feature progressive knowledge distillation completion method for a user portrait of a nasopharyngeal carcinoma patient, and relates to the technical field of image data processing, and the method comprises the following steps: a pre-distillation stage: carrying out the pre-distillation of a pre-training language model, so as to obtain a high-quality teacher model, generating feature distillation through a mask, and carrying out the feature distillation; applying a mask operation in the input data, and setting a mask proportion according to the number of student model parameters; and S2, mask generation feature distillation: the teacher model and the student model perform mask operation on input data including the triad and text description thereof in a data input stage. In the masking generation characteristic distillation, masking operation is applied to input marks, and a proper masking rate is set in a pre-distillation stage according to student model parameter counting, so that the performance of an initial model is improved. In the progressive distillation stage, a multi-grade student model is designed, and the parameter count and the mask rate are gradually reduced. Each student model is trained under the three kinds of supervision described above.

This invention discloses a fast activity recognition method based on a multi-path parallel MLP mixer architecture. First, multi-dimensional sequence data is used as input. An embedding module encodes the multi-dimensional data to obtain a feature sequence. This feature sequence is then fed into multiple MLP mixer branches, where features are mixed along the time, channel, and frequency domain dimensions, respectively. In each branch, after passing through multiple MLP mixers, global pooling is used to aggregate the feature sequence along the time dimension into a feature vector. Finally, the feature vectors obtained from each branch are concatenated and sent to a classification head for activity recognition. Compared with existing deep learning-based activity recognition methods, this invention achieves more accurate activity recognition with fewer model parameters, less floating-point computation, and faster inference speed, meeting the requirements of high precision and high efficiency in practical applications.

The application discloses a kind of knowledge distillation methods, devices and equipment of mask autoencoder and storage medium, the method is by respectively establishing the teacher model and student model of mask autoencoder, wherein the teacher model and the student model are all visual transformation model, and the size of the teacher model is greater than the student model;The teacher model is pre-trained;The student model is pre-trained based on the pre-trained teacher model, so that the student model learns data generalization ability from the pre-trained teacher model, and obtains better image features representing ability;The student model is fine-tuned based on the pre-trained downstream task training, and the student model can be deployed in the power edge side lacking of computing resources, while ensuring that the model accuracy does not decrease, the model parameter is reduced, and the real-time inference speed is accelerated.

The invention provides a plant station illegal behavior intelligent monitoring method and system based on an improved YOLO algorithm, and belongs to the technical field of industrial plant station safety management. The monitoring method comprises the steps of video stream acquisition and preprocessing, improved YOLO feature extraction and detection, specific illegal behavior identification, illegal behavior decision, alarm linkage, platform interaction and the like. The core improvement lies in that a lightweight backbone network is constructed by adopting GhostNet, and model parameters and calculation amount are reduced; a small target detection effect is optimized by newly adding a high-resolution detection head, introducing a BiFPN fusion mechanism and adopting a Focus loss and attention mechanism; color band channel enhancement, time sequence difference features and optical flow detection are fused, and flame / smoke recognition robustness is improved; and multiple types of illegal behavior recognition are integrated in the same model system, so that unified monitoring is realized. According to the method, the problems of insufficient model lightweight, unstable small target detection, unreliable fire identification, complex deployment and the like in the prior art are solved.

A hyperspectral image pixel-level classification method based on a MADWA network belongs to the field of remote sensing image classification. The present invention solves the problem of poor classification accuracy of existing deep neural networks under the background of noisy labels. The present invention uses a multi-scale structure to fully learn the real label and noise label data information, so that the network's ability to learn information is improved; and a spectral spatial attention structure is designed to more effectively focus on the unique features to be learned in each dimension, so as to make more full use of the local features of hyperspectral data and improve classification accuracy. And by adopting an asymmetric convolution kernel, the model parameters can be greatly reduced, the running speed can be improved, and the computational complexity can be reduced. The method of the present invention can be applied to remote sensing image classification.

The application discloses a voltage sag analysis method based on a 1D V-net deep learning model, which can directly learn voltage sag characteristic information caused by motor starting, transformer switching, single-phase short circuit, two-phase short circuit and three-phase short circuit from original monitoring data, and avoids a tedious manual feature extraction process. Compared with a one-way cyclic network structure formed by RNN, LSTM, GRU and the like, the application can improve voltage sag disturbance type classification accuracy and voltage sag start-stop time positioning accuracy. Compared with a two-way cyclic network structure formed by RNN, LSTM, GRU and the like, the application can guarantee voltage sag disturbance type classification accuracy and voltage sag start-stop time positioning accuracy on the basis of reducing model parameters.

The application discloses a lightweight dual-prediction branch semantic segmentation water body extraction deep learning method and system, first, the coding part of the lightweight dual-prediction branch semantic segmentation model is constructed, three convolution blocks in the coding structure are designed based on the visual perception, and high, medium and low resolution feature maps are extracted respectively; secondly, the decoding part of the lightweight dual-prediction branch semantic segmentation model is constructed, the dual-prediction branch is designed, the medium and low resolution feature maps are fused and used as one of the discriminant features, input into one of the prediction branches, the high resolution feature map is input into the other prediction branch, and the maximum confidence of the two prediction branches is used as the final prediction probability; finally, the model is trained by using a satellite remote sensing image labeled data set and the adaptive learning rate. The deep learning model provided in the application can effectively solve the water body extraction in a large range of satellite remote sensing images, and provide an intelligent method for dynamic change monitoring of water resources.

The invention discloses an auditory space attention decoding method based on a low-density electroencephalogram cross-hemisphere attention mechanism, electronic equipment and a medium, and belongs to the field of brain-computer interface and neural signal processing. The method comprises the following steps: preprocessing low-density electroencephalogram signals, and respectively mapping the low-density electroencephalogram signals into three-dimensional tensors with a time dimension and a two-dimensional space dimension according to a left hemisphere and a right hemisphere; a space-channel attention module is utilized to extract key spatio-temporal characteristics of each hemisphere; dynamic interaction between left and right hemisphere features is modeled through a cross-hemisphere attention module, and information fusion is realized; and finally, classifying the fused features, and outputting auditory attention directions. According to the method, an electroencephalogram space-time structure is reserved through three-dimensional tensor mapping, feature discrimination is enhanced in combination with a double attention mechanism, high-precision decoding can still be achieved under the low-density condition that only 4-8 channels are used, meanwhile, the method has the advantages of being low in calculation complexity and high in physiological interpretability, and the method is suitable for wearable brain-computer interfaces and intelligent hearing-aid equipment.

The invention discloses an urban traffic infrastructure operation state detection and evaluation method, and belongs to the technical field of traffic infrastructure monitoring. In order to achieve rapid and intelligent identification and evaluation of urban traffic infrastructure operation state indexes, the method comprises the steps that the traffic infrastructure operation state indexes are set, and then a data set used for urban traffic infrastructure operation state detection and evaluation is constructed; constructing an intelligent identification model of the traffic infrastructure operation state indexes, wherein the intelligent identification model comprises a basic model, a first auxiliary model, a second auxiliary model and a strengthening model; using the obtained data set for urban traffic infrastructure operation state detection and evaluation to train the constructed basic model, the first auxiliary model, the second auxiliary model and the strengthening model, and using the obtained trained basic model as a final traffic infrastructure operation state index intelligent identification model to construct a digital twin model; and a traffic infrastructure operation state evaluation method is constructed, and the traffic infrastructure operation state is evaluated.

The invention discloses a method and a device for restoring a compressed image, electronic equipment and a computer program product. Relates to the audio and video field, and the method comprises the following steps: obtaining a target compressed image, decoding the target compressed image according to a preset decoding mode to obtain an initial restored image, the preset decoding mode being a decoding mode corresponding to a compression mode of the target compressed image; the initial restored image is input into a target model, a target restored image is obtained, and the target model is composed of N shallow feature extraction modules, M deep feature extraction modules and an up-sampling module. Through the method and the device, the problem of high computing resource consumption in a compressed image restoration process in the related art is solved.

The invention provides a smoked sheet rubber impurity removal path planning method based on visual guidance. According to the method, work is executed based on an impurity identification model and a path planning model. Wherein the core of the impurity identification model is a neural network which is established on the basis of an optimized YOLO algorithm and is used for smoked sheet rubber image target identification, and in addition, an SAHI model is used for carrying out auxiliary prediction of impurities; the core of the path planning model is to carry out global path planning based on an ant colony algorithm. The specific execution process for completing smoked sheet rubber impurity identification and elimination path planning work by using the method comprises the steps of constructing a smoked sheet rubber data set, completing training of an impurity analysis model, performing impurity prediction based on the impurity analysis model, performing coordinate conversion on a smoked sheet rubber impurity image set, and obtaining the smoked sheet rubber impurity identification and elimination path planning work. And calculating an elimination path corresponding to the impurities based on the path planning model.

A link prediction method and apparatus based on symbolic graph neural networks includes acquiring a symbolic graph comprising multiple drug nodes and multiple links. The multidimensional attribute features of the multiple drug nodes in the symbolic graph constitute an attribute feature matrix. This attribute feature matrix is ​​input into a preset link prediction model. After processing by a symbolic graph neural network based on a low-pass filter, a fully connected layer, and a regression classifier within the link prediction model, the predicted link between any two drug nodes in the symbolic graph is iteratively output. The error between the predicted link and the actual link in the symbolic graph is analyzed, and the model parameters of the link prediction model are adjusted based on the error until the training termination condition is met. The trained link prediction model is then used to predict the link between the drug node pairs to be predicted. This invention designs a symbolic graph neural network based on a low-pass filter. The end-to-end prediction model can simultaneously train a feature extractor and a link type discriminator, resulting in high accuracy in link prediction.