Research On Simulation Image Evaluation Based On Deep Learning

With the continuous development of computer processing capabilities and software,the research of infrared scene simulation technology has progressed rapidly,and it has been widely used in both military and civilian applications,showing good prospects.High fidelity is the prerequisite for infrared scene simulation technology to be put into application.As the performance of infrared scene simulation systems continues to improve,the fidelity of infrared scene simulation images is getting higher and higher.How to evaluate the simulation results has become a key issue that affects the development and application of infrared scene simulation technology.At present,there are few studies on evaluation schemes for infrared scene simulation,and an effective evaluation method and a mature evaluation system have not yet been formed.This paper combines the research team's many years of research foundation on infrared simulation technology with their learning and understanding in the field of deep learning,and Combining with actual scientific research projects,an evaluation scheme for the similarity and fidelity of infrared scene simulation images based on deep learning is proposed.The main research content has the following three aspects:(1)The principle of the infrared scene simulation process is analyzed,and infrared scene simulation is carried out for two practical problems,and some simulation images are obtained.The simulation evaluation problem is discussed.For this paper,the evaluation direction of the simulation image is refined into two aspects:image similarity and image fidelity evaluation.Finally,the typical traditional simulation image similarity evaluation method is analyzed and implemented,and a typical evaluation scheme for the simulation system is selected for analysis.(2)Combining deep learning with traditional similarity evaluation methods,a simulation image similarity evaluation method based on twin neural networks is proposed,and the training process of the model is improved according to the problems encountered in the experiment.The ground background image generated by an infrared scene simulation is evaluated,and the similarity evaluation method in this paper is verified by traditional methods.(3)Introduce the idea of generating and discriminating against each other in the generation of adversarial networks.The simulation image fidelity evaluation method based on the generation of adversarial networks is proposed.The real scene images are used as data sets to train the generation of adversarial networks.The image was tested for fidelity evaluation.In response to the problems that occurred during the experiment,the network model and evaluation method were improved.Finally,the evaluation results are analyzed to prove the reliability of the fidelity evaluation method proposed in this paper.