Method and device for creating remote sensing image style migration model

Abstract

The embodiment of the invention discloses a method and a device for creating a remote sensing image style migration model, and the method comprises the steps: employing an iteration mode to execute the following steps until the obtained overall feature loss reaches a preset threshold value: inputting an original image into an initial learning algorithm model, and obtaining a generated image; inputting the original image, the generated image and the style image into a pre-training model to obtain overall feature loss; adjusting parameters of the initial learning algorithm model, weights of a plurality of different pieces of feature information in the pre-training model and a mixing ratio between the weight of the content feature information and the weight of the style feature information according to the overall feature loss; and taking the initial learning algorithm model adjusted according to the overall feature loss conforming to a preset threshold as a remote sensing image style migration model. Through the scheme of the embodiment, the style imbalance of the strong style image is avoided, the texture feature and the color feature of the original image are maintained, and the obvious real style effect is achieved.

Description

technical field [0001] The embodiment of the present application relates to remote sensing image processing technology based on deep learning, especially a method and device for creating a remote sensing image style migration model. Background technique [0002] The current art style migration algorithm based on deep learning: [0003] The stylized output of the content image obtained by the generative model of artistic style and the model based on optimization has large local changes, distortion of lines and texture, and cannot reflect the true color of the image, making the current artistic style migration algorithm based on deep learning ineffective. The style transfer suitable for real remote sensing images can hardly be applied to the style transfer of remote sensing images. At the same time, the style transfer model based on optimization needs a long iterative optimization time, and the generated image after stylization has large color blocks and artifacts. [0004] T...

AI Technical Summary

Problems solved by technology

The image instance segmentation step not only increases the time cost, but also makes it difficult to ensure the accuracy of the segmentation, which in turn affects the final effect of style transfer

Using the pre-trained model as the training of the feature loss control model, although the inference speed is improved, the final generated image is difficult to bring satisfactory results on the remote sensing image due to the influence of the global style feature of the style image.

At the same time, the style characteristics of global strong style images can only be applied to images with simple texture lines and image content

Images