56results about How to "Fast training convergence" patented technology

The invention discloses an image tampering detection method based on a full-resolution mixed attention mechanism, and the method comprises the steps: obtaining a to-be-detected image, inputting the to-be-detected image into a pre-trained image tampering detection model based on the full-resolution mixed attention mechanism, and outputting a tampering detection result. The method has the advantages that the problem that the image resolution is reduced due to potential tampering of region texture features and loss of edge high-frequency features in an existing method is effectively solved, the direction of a network output vector is optimized, and the network training convergence speed is increased; a training method based on feedback ablation points is provided, and modular end-to-end training in a limited memory is achieved; the feedback of each module is combined to optimize and adjust parameters, and the reasoning test speed of the model is accelerated. The method has better accuracy, stronger robustness, faster training speed and smaller memory occupation in the aspect of detection tasks with multiple tampering types simultaneously, and has a very wide application prospect.

The invention discloses a multi-dimensional isomorphic heterogeneous BP neural network optical power ultrashort-term prediction method. The method specifically comprises the steps of: SS1, utilizing grid-connected active power measured data, meteorological substation measured data, grid-connected active power historical data, meteorological substation historical data and weather forecast data comprehensively, analyzing data segmented by the day, calculating an index of similarity under the conditions of approximate meteorological conditions and similar active power, and classifying the data according to the index to form historical data samples; SS2, correcting numerical weather information according to the weather forecast data and the meteorological substation measured data; SS3, matching the samples under the conditions of approximate meteorological conditions according to the corrected numerical weather information and classified historical data samples, and selecting the approximate samples as input training samples of an artificial neural network; SS4, carrying out input data normalization, training sample selection and predictive output for a BP neural network; and SS5, repeating the process from step SS1 to step SS4 when the prediction in the next time period starts.

The invention relates to a target detection and positioning method and system for unmanned vehicles. The method includes: obtaining a projection matrix; determining the pixel coordinates and type information of the 2D bounding box of the detection target in the image; combining the laser radar data with the previous step Time synchronization and filtering of the information; the filtered lidar data is projected into the detection image through the projection matrix, and the lidar data in the 2D bounding box of the detection target in the corresponding image is obtained and clustered; the lidar data is in x Calculate the mean value in the y direction to obtain the coordinates of the detection target relative to the lidar coordinate system; calculate the relative coordinates of the detection target relative to the origin of the vehicle center coordinate system; obtain the UTM coordinates of the vehicle, and calculate the coordinates of the detection target in the UTM projection coordinate system , and converted to the WGS1984 geographic coordinate system. The above method in the present invention can detect the target in real time and obtain the longitude and latitude coordinates of the target in the WGS1984 coordinate system.