Research On Outdoor Static Scene Image Prediction Model Based On Environmental Multi-factor Fusion

Outdoor static scene image prediction is a branch of image prediction,which has important application value in scene matching,aircraft simulation,virtual reality and so on.Due to the long prediction time and the complex outdoor environment,outdoor static scene image prediction technology has become a new challenge for image prediction technology.In order to solve the problem of long prediction time and complex relationship between outdoor environment,this thesis proposes an outdoor static scene image prediction model based on multi-factor fusion,which is suitable for real outdoor scene prediction under normal weather.Secondly,according to the related imaging theory of haze days,the model is modified and applied to outdoor static scenes of haze days,which expands the application scope of the prediction model.Main work of this thesis:1.Based on the causal relationship between the outdoor environment and the image,a physical prediction model based on environmental multi-factor fusion is proposed,which can be applied to the outdoor static scene images.The model integrates various environmental parameters and image capture time to realize image prediction of static scenes in adjacent days.An experimental data acquisition device is built,the images are then corrected and restored,in order to establish an outdoor scene database.2.Based on the theory of light decomposition and the theory of environmental effects,an image prediction algorithm for outdoor static scenes under normal weather is proposed.The composition of the skylight basis image and sunlight basis image is analyzed,and the modulation transfer function is utilized to characterize the atmospheric effect.The relationship between the aerosol and turbulence modulation transfer function and the weather parameters is then analyzed.The prediction results of scene images which are captured on two random days are as follows: for the first day,the cosine similarity between the predicted image and the real reference image is 0.9633,the peak signal-to-noise ratio is 25.4198 d B and the structural similarity is 0.9277;for another day,the cosine similarity of the images is 0.9140,the peak signal-to-noise ratio is 23.7065 d B and the structural similarity is 0.9334,which verify the effectiveness of the proposed algorithm.3.The atmospheric modulation transfer function of haze weather is introduced to modify the prediction model,then an image prediction algorithm which is suitable for haze static scenes is proposed.Two neural network models of Cycle GAN and pix2 pix are introduced to predict haze scenes.By comparing the three above models,the results are as follows: the poor prediction stability of pix2 pix model makes it not suitable for static image prediction.The peak signal-to-noise ratio of the physical modified model is better than the neural network model,while the structure similarity of Cycle GAN model is higher.