41results about How to "Improve the evaluation index" patented technology

The invention relates to a visual attention detection method for a three-dimensional video. The method is characterized in that firstly a characteristic contrast is calculated by utilizing a low-level visual characteristic so as to obtain spatial saliency of a three-dimensional video frame; and in addition, time saliency is obtained by utilizing motion information, and motion saliency is calculated by using a motion in plane and a motion in depth for the motion information in the three-dimensional video. Finally a saliency map is obtained by the spatial saliency and the time saliency, and a common fate law and a tightness law in Getalf psychology are applied in a combination process. An experimental result shows that a good effect is achieved in saliency prediction of the three-dimensional video.

The invention discloses an answer ordering method for a community question-answer platform. According to the method, rich metadata, such as themes, timestamps and text content of questions and answers, of the questions and answers can be sufficiently used for solving the answer ordering problem; meanwhile, an enhancement-type attention mechanism recurrent neural network model (EARNN) is used for solving the answer ordering problem, and compared with a tradition model, more information is used. A predicted result is improved to a certain degree on multiple evaluation indexes.

The invention relates to a two-dimensional video saliency detection method based on long-term and short-term memory. The method is characterized by comprising the following steps: firstly, extractingshort-term time sequence features by utilizing a 3D convolutional network (3D-ConvNet); secondly, extracting long-term time sequence features by adopting a bidirectional long-term and short-term memory network (B-ConvLSTM); fusing the extracted short-term time sequence features and long-term time sequence features; and finally, obtaining a saliency map through fusion result deconvolution. The model is combined with long-term and short-term time sequence characteristics, the operation information of the salient target in the video can be effectively reserved, and the saliency prediction experiment result of the two-dimensional video proves that the proposed model can obtain a good detection effect.

The invention provides a machine reading understanding model training method and device based on a joint loss function. Specifically, when machine reading understanding model training is carried out,a loss function composed of a maximum likelihood estimation function and a minimum risk training function is used as an evaluation index of a machine reading understanding model so as to guide adjustment of parameters of the machine reading understanding model. The idea of the minimum risk training function is to use the loss function to describe the difference degree between the answer output bythe model and the standard answer. The maximum likelihood estimation function, namely loss, tries to find a group of model parameters, so that the loss value of the machine reading understanding modelon the training set is minimum, and therefore, compared with the mode of singly utilizing the maximum likelihood estimation function, the model trained by the method provided by the invention can extract answers more accurately.

The invention discloses a synthetic aperture radar (SAR) image compression method based on target area extraction and direction wave, and mainly solves the problems of less edge information stream distribution and loss of important information due to the fact that the same transform compression strategy is used in an image target area and a background area in the existing method. The method comprises the implementation steps of extracting a texture map of an SAR image by a variation coefficient; carrying out quadtree partitioning on the SAR image, dividing an image block into the target area and the background area by the texture map; detecting a diverting pair of the target area by the texture map, and carrying out pruning processing on partitioned images; and carrying out directionlets transformation on the target area, carrying out wavelet transformation on the background area, and respectively encoding the coefficients of the target area and the background area by a set partitioning in hierarchical trees (SPIHT) encoding method. The SAR image compression method has the advantage of well protecting the information of the target area by different transform compression strategies in the target area and the background area, and can be applied to real-time transmission and storage of the SAR image.

The invention discloses a method for predicting a next track point of a user. The method comprises the following steps: crawling a certain amount of data: an ID of the user, position information of aseries of short-term and long-term historical track points corresponding to the user, and a timestamp of each track point; constructing a feature interaction self-attention network model based on thecrawled information, and making attention in combination with a result that the position information of the long-term historical track points of each user passes through a self-attention layer; performing optimal training on the parameters by using a cross entropy loss function; for a new user and a series of historical track points thereof, and constructing a series of instances by utilizing theID information, the position information of a series of historical track points corresponding to the user and the timestamp of each track point, and inputting the instances into a trained feature interaction self-attention network model, thereby obtaining a series of sorting scores of predicted positions. According to the method, the problem of predicting the next track point by utilizing the richmetadata of the user and the historical track is solved, and the prediction accuracy is greatly improved.

The invention discloses an image super-resolution reconstruction method. The method includes two parts: 1. Adaptive decomposition of the image. The so-called self-adaptation here refers to the self-adaptive drive decomposition process of the image data. 2. Perform super-resolution amplification on the basis of image adaptive decomposition. On the basis of adaptively decomposing the image into low-frequency information and high-frequency information, a Gaussian mixture model is used as a priori model in the high-frequency part for super-resolution amplification, and spline interpolation is used in the low-frequency part, and finally superimposed. The invention can realize high-quality enlargement of images.

The invention discloses a sequence position recommendation method based on geographic perception. The problem of position recommendation can be solved by fully utilizing rich metadata (user ID, timestamp, position and place longitude and latitude information) of users and historical behavior tracks. Meanwhile, the geographic perception sequence recommender GeoSAN based on the self-attention network is used for recommending the user position, compared with a traditional model, geographic information is more fully utilized, the data sparsity problem can be well solved, and the recommendation result is improved to a certain extent on multiple evaluation indexes.

The invention relates to the technical field of machine learning, in particular to a negative sample extraction method and device, computer equipment and a storage medium, and the method comprises thesteps: obtaining the page burying point information of a display page in an application platform, and determining the label information and popularity information of each burying point object according to the page burying point information; acquiring historical behavior information of a user in the application platform, and determining a label weight of the user in the application platform according to the historical behavior information; according to the label weight and the label information of each buried point object, determining the sampling probability of each buried point object sampled by the user in the application platform; generating a negative sample distribution sequence according to the popularity information and the sampling probability of each buried point object, and extracting a negative sample from the negative sample distribution sequence. According to the scheme, the sampling logic of the negative sample is optimized, the calculation amount in the model training process is reduced, and the model effect and the evaluation index are improved.

The invention relates to a multi-mode video dense event description algorithm of an interactive Transform, and belongs to the technical field of video algorithms. The method comprises the following steps: 1, extracting visual features, audio features and voice features in a video; information in the video is better utilized through multi-modal feature extraction; 2, fusing the visual features with the audio features and the voice features through an interactive attention module in the interactive Transform, and further coding the video features; 3, completing model training in two stages; firstly, a description model is trained based on real video segments, then the encoder weight of the trained description model is frozen, and then a segment proposal model is trained. According to the method, the feature information in the video is fully utilized, the multi-modal features are interactively fused, and a good dense video description effect is shown.

The invention provides a remote sensing image text generation and optimization method based on self-reinforcement learning. The remote sensing image text generation and optimization method based on self-reinforcement learning comprises the following steps: S1, extracting remote sensing image semantic understanding feature; S2, obtaining a training set, pre-training the text generation model, and extracting text generation model parameters; and S3, inputting the extracted feature vectors, a prior text library, pre-trained text generation model parameters and task requirements of a user into a remote sensing image text generation network, and restoring image feature information represented by the extracted feature vectors into text description through a deep learning natural language processing technology. The text is generated by adopting the self-reinforcement learning remote sensing image text generation algorithm based on the strategy gradient algorithm, the training effect of the remote sensing image generation model is improved, parameters are promoted to converge towards expected values, and the accuracy of generation description is improved.

The invention provides a single image rain removal method based on a multi-scale fusion generative adversarial network. The method comprises the following steps of firstly, detecting a rain image through image saliency to obtain a saliency map, fusing the saliency map with a rain image Concat, and accurately recognizing a raindrop area to be restored, then, using multi-scale fusion to generate anadversarial network for rain removal, conducting multi-scale fusion on l12, l14 and l16 layers of networks in a generator network, enhancing the quality of a generated rain removal image, enabling thediscriminator network to adopt combination of global discrimination and local discrimination, and conducting training to obtain a final network model, and inputting the test set into the trained model to obtain a rain removed image, and evaluating the generated image according to the SSIM and PSNR indexes. The image generated by the image rain removal method provided by the invention is better invisual effect, the removed raindrop area is more authentic and coherent, and each evaluation index is improved.

The invention relates to the field of function testing, and in particular relates to a testing method for a noise reduction function of a device and the relevant device. The testing method for the noise reduction function of the device comprises the following steps: acquiring performance parameters of a to-be-tested device, and establishing a testing scene of the to-be-tested device according to the performance parameters; acquiring parameters of the testing scene, according to the parameters of the testing scene, extracting any original sound corresponding to the parameters of the testing scene from a preset sound library, and according to the differences between a sound wave curve of the original sound and a preset sound wave curve, correcting the original sound, so that a sound sample is obtained; and inputting the sound sample into the to-be-tested device, carrying out playing through the to-be-tested device, receiving the played sound sample, thus obtaining a testing sample, comparing the testing sample with the sound sample, and generating a noise reduction testing report according to the comparing result. The evaluation index for testing the noise reduction function is increased, the testing dimensionality is expanded, the testing efficiency is promoted, and the automatic testing can easily access in quality testing.

The invention discloses a motion direction change-based candidate seed point cloud single target tracking method, which comprises a training module and a test module, and is characterized in that the training module firstly preprocesses point cloud to obtain template point cloud, then uses Gaussian sampling to obtain candidate point cloud, and then inputs the template point cloud and the candidate point cloud into an encoder for encoding, a corresponding feature vector is obtained, a distance loss function and a direction loss function are calculated, and the whole model is trained. The test module firstly uses a pre-trained PointRcnn model to carry out target detection, then carries out candidate area sampling, then inputs sampled candidate point cloud and tracking target point cloud of a previous frame into the trained model to carry out coding, and finally carries out target tracking on a coded feature vector by using cosine similarity comparison. According to the method, the single-target tracking precision can be improved, and a wrong tracking phenomenon can be effectively prevented.

The invention provides a retina vessel segmentation method based on residual network feature extraction, which is applied to a neural network model, and comprises the following steps: enabling an original retina vessel image to pass through a pre-trained VGG coding layer to obtain a plurality of images, wherein the number of the images is five, the sizes of the images have a preset proportional relation with the sizes of the original retinal blood vessel images. Therefore, the segmentation precision of the network is obviously improved, and the fitting ability and generalization ability of the model are better optimized. Compared with a Unet, the network also generally has better performance in other data sets.

The invention relates to a risk prediction method based on clinical examination and medication intervention data, and the method comprises the steps: selecting start and end nodes from the clinical examination data in an individual observation period, carrying out the vectorization modeling, and obtaining an input vector x1; constructing an intervention dictionary, calculating the characteristic frequency of medication intervention, and performing vectorization modeling on individual medication intervention data to obtain an input vector x2; combining the input vector x1 and the input vector x2 to obtain an input feature vector X; inputting the input feature vector X into a prediction model, obtaining a real result Y through fitting, optimizing prediction model parameters, and obtaining a final prediction model; inputting the individual data into the final prediction model subjected to parameter adjustment and outputting model prediction results. According to the method, the design is reasonable, the relationship between different medication intervention combinations and the influence of the medication intervention combinations on the individual state can be explored, the prediction is accurate and reliable, and each evaluation index is improved.

The invention provides a student score prediction method based on a two-way attention mechanism. The student score prediction method comprises the following steps of obtaining student attribute characteristics, first-stage historical scores and second-stage historical scores; obtaining a feature vector f1 of each student attribute feature with respect to a first-stage historical score and a feature vector f2 of each student attribute feature with respect to a second-stage historical score; fusing the feature vector f1 and the feature vector f2 to obtain a feature vector f; and calculating a student score prediction score p from the feature vector f through a three-layer full-connection neural network MLP. By introducing a double-path attention mechanism, the defects of a single-path attention mechanism can be overcome, information complementation is well performed, and the prediction performance of the model is further improved; and for a prediction result, a plurality of evaluation indexes are remarkably improved.