Research On Key Technologies Of Multi-source Image Fusion In Complex Environment

As an objective reflection of the natural landscape,images are an important medium for mankind to understand the world.Due to the limitations of front-end imaging equipment and the complexity of the environment,the single-source image information accepted by a single sensor is limited and cannot meet specific needs.Different source images acquired by different sensors are complementary and can reflect different characteristics,but the storage and transmission costs are relatively high.Combining different source images and retaining effective information from each source image can not only reduce storage and transmission costs,but also facilitate subsequent decision-making.This article carries out related research,mainly in three aspects.1.Polarization and visible light image fusion.Aiming at the problem of the complexity and low robustness of the traditional fusion method,a fusion method based on the two-way dense connection generation confrontation network is proposed.The two-way dense connection network is used as the generator,and the training method based on the generative adversarial network is used to effectively improve The robustness of the fusion method and the quality of the fusion image.2.Infrared and visible light image fusion.In view of the complex design of traditional methods and the incomplete retention of gradient information in most deep learning methods,a dual-channel Unet autoencoder is proposed as a fusion network,which eliminates the training method of GAN network,adopts an end-to-end training method and improves the gradient loss function,Which can effectively reduce training costs and improve the quality of fusion images.3.Multi-focus image fusion.in view of the cumbersome process of mask-based generation methods in processing the edge of the mask,and the poorly processed masks have obvious segmentation traces on the resulting image,the end-to-end generation model is used and the method of supervised learning is adopted.The continuity of edge processing is better than that based on segmentation.