The embodiment of the invention discloses a high-resolution remote sensing image weak and small target detection method and device based on deep learning. The method comprises the steps of obtaining ato-be-processed remote sensing image; inputting the remote sensing image to be processed into a pre-trained convolutional neural network, carrying out 4-time downsampling, 8-time downsampling and 16-time downsampling respectively on the remote sensing image to be processed through the convolutional neural network; obtaining the priori boxes of different sizes corresponding to the to-be-processedremote sensing image, identifying the target priori boxes of which the target category confidence is greater than a preset threshold, and determining the coordinate information of a target included inthe to-be-processed remote sensing image through a preset clustering algorithm according to the coordinate information of each target priori box, wherein the first layer of the convolutional neural network comprises a residual component, the second layer, the third layer and the fourth layer of the convolutional neural network each comprise four residual components, and each residual component comprises two convolutional layers and a fast link. By applying the scheme provided by the embodiment of the invention, the weak and small target detection precision can be improved.
The invention provides a modeling method based on wind power fluctuation multi-scale decomposition, comprising: according to the first wind power historical data collected in advance, analyzing the time characteristic and statistical characteristic of wind power fluctuation, so as to determine the wind power fluctuationdecomposition component and the time scale corresponding to the wind power fluctuationdecomposition component; According to the wind power fluctuation decomposition component and its corresponding time scale, a two-stage WMMF filter is used to decompose the wind power multi-scale fluctuation of the first wind power historical data to obtain a low-frequency trend component, an intermediate-frequency fluctuation component and a high-frequency fluctuation component. Accordingto the low frequency trend component, the intermediate frequency fluctuation component and the high frequency fluctuation component, a multi-dimensional probability model is established. Through theabove method, the multi-dimensional probability model can be established according to the fluctuation characteristics and correlation of wind power fluctuation of the original wind power time series,thus retaining the characteristics of the complete wind power fluctuation process and simulating the wind power output characteristics to the maximum extent.
The invention belongs to the technical field of artificial intelligencegraph classification, and provides a graph classification method and system fusing high-order structure embedding and composite pooling, and the method comprises the steps: obtaining a to-be-classified graph; inputting a to-be-classified graph into the graph neural network to obtain a category to which the graph belongs; wherein for each sub-graph set of the graph, each convolutional layer calculates the feature of each sub-graph based on the sub-graph set output by the previous neural network layer, each composite pooling layer updates the sub-graph set based on the feature of each sub-graph output by the convolutional layer, and meanwhile, for each sub-graph in the updated sub-graph set, the feature of each composite pooling layer is calculated based on the feature of each sub-graph output by the convolutional layer. The features of the sub-graphs in the local neighborhood are fused through an attention mechanism, and the features of the sub-graphs are updated; and obtaining a graph representation vector by the reading layer, and inputting the graph representation vector into the classifier to obtain a category to which the graph belongs. A high-order structure is utilized, messages are directly transmitted among the sub-graphs, structural information invisible in node level is captured, and the classification precision of the graphs is improved.
The invention discloses a three-dimensional model retrieval method based on LSTM network multi-modalinformation fusion, and the method comprises the steps: for a given three-dimensional model, extracting a plurality of views of the three-dimensional model arranged according to a rotation angle sequence; extracting skeleton characteristics of a plurality of views in a multi-task and multi-angle manner, and obtaining structured information of the three-dimensional model according to the skeleton characteristics; extracting view feature vectors of a plurality of views, and inputting the view feature vectors into a layer of LSTM network structure; checking whether other feature vectors need to be extracted continuously or not; connecting the skeleton feature vector with the view feature vector subjected to one layer of LSTM to form a new feature vector, and inputting the new feature vector into a second layer of LSTM network structure for fusion; checking whether other to-be-fused featurevectors exist or not, if yes, forming a new feature vector again and inputting the new feature vector into the next layer of LSTM network structure for fusion; and taking the output of the last fusion as the final feature vector Q of the three-dimensional model, and finishing the final detection process of the three-dimensional model in combination with a similarity measurement method.
The invention discloses a pneumonia image processing method and system and a storage medium; the method comprises the steps: carrying out the filtering reconstruction and feature enhancement of a to-be-processed pneumonia image, and carrying out the fusion of the to-be-processed pneumonia image with an original pneumonia image, and maintaining the features of the original pneumonia image and the feature-enhanced image, thereby facilitating the subsequent neural network learning. Through verification of an InceptionV3 network, the pneumonia image is processed by adopting the method disclosed by the invention, and compared with an unprocessed pneumonia image and a pneumonia image which is processed only by using a Retinex algorithm, the obtained pneumonia image is improved in both accuracy and specificity.