METHOD OF DETECTING ANOMALY OF HYPERSPECTRAL IMAGE BASED ON "TRAINING-TRAINED" MODEL, COMPUTER DATA MEDIUM AND DEVICE Russian patent published in 2024 - IPC G06F17/10 

FIELD: computer engineering.

SUBSTANCE: invention relates to computer engineering for detecting hyperspectral image anomaly. Result is achieved due to the fact that the data to be detected contain n types of background data, and training data of the model of the trained network contain normal data from n types of background data; inputting the training data of the model of the trained network into the model of the training network, trained at the stage S2, and obtaining the embedding derived by the training network model; simultaneously entering training data of the model of the trained network into the model of the trained network, obtaining the embedding derived by the model of the trained network, performing training on the model of the trained network multiple times; entering the data to be detected into the training network model trained at step S2 and into the trained network model trained at step S4 to detect an anomaly; calculating to obtain an anomaly estimate and determining that the corresponding pixel is an abnormal pixel, in accordance with a predetermined threshold value T of the anomaly estimate, if the anomaly estimate is greater than or equal to the threshold value T; if the anomaly estimate is less than the threshold value T, determining that the corresponding pixel is a normal pixel; abnormal pixels are screened out from experimental data and subsequent detection of objects is performed.

EFFECT: high accuracy of detecting an existing anomaly of a hyperspectral image.

6 cl, 13 dwg, 2 tbl