Research On Color Consistency Of Remote Sensing Image Based On Deep Learning

In the process of satellite remote sensing image acquisition,due to the different sensors,shooting time,shooting lighting conditions and atmospheric conditions of different satellites in the original data imaging,there are problems such as uneven brightness and color distortion in the image interior and between image scenes,which makes it difficult to process and apply the image subsequent splicing,and affects the overall color consistency of the image after orthophoto correction and fusion.To solve this problem,traditional satellite remote sensing image data splicing and fusion uses the combination of base map template and manual color blending to process color,which consumes a lot of human resources;If only the base map template is used for color blending,due to the complexity of the ground feature types and seasonal effects,it is impossible to achieve full automatic color consistency processing.Based on the above problems,through the research and analysis of the image color model,combined with the basic theory of convolution network and the idea of generative adversarial networks,this thesis proposes a color consistency method that introduces the attention mechanism to improve the model of cycle generative adversarial networks,which can realize the automatic color consistency processing of complex objects,and provide a new idea for the color consistency of satellite remote sensing images.The main achievements of this thesis are as follows:(1)Make a sample dataset of image color consistency template.The single scene image entity data and sub meter satellite image service data are selected respectively,and the typical geographic region data of China are selected respectively.The color consistency template of China’s image is made by preprocessing,grouping and cutting the data.(2)Improve the Cycle GAN model and conduct experiments.By changing the residual block network structure,adding jump connections,changing the normalization method and other operations,the improved Cycle GAN model that is more suitable for color correction is obtained.Experiments on single scene image entity data and sub meter satellite image service data prove that the improved Cycle GAN model proposed in this thesis has a good effect on image color consistency.(3)Through the method of thinning data sets and introducing attention,the color consistency model with attention mechanism is obtained for different surface feature composition data.The model with attention mechanism is compared with the model before attention mechanism.It is proved that the network effect with attention mechanism is better than that without attention mechanism,and the image data types suitable for each attention mechanism are obtained.