Research On Infrared Scene Modeling Algorithm Based On Real Geographic Information

Infrared scenes have effective characteristics in low visibility environments such as nights and dense,and can be used for detection,reconnaissance,and other tasks.Under low illumination or harsh weather conditions,infrared scenes can characterize the thermal radiation of objects to produce high-contrast images,thereby improving the efficiency of object detection and recognition.Obtaining high-quality infrared scene images has a certain degree of difficulty,requiring not only expensive equipment but also harsh conditions.Secondly,infrared scene images obtained using conventional algorithms often have semantic differences and uneven gray-scale brightness distribution problems.Therefore,this study takes real geographic data as the research object and studies semi-supervised modeling and multiband modeling techniques.The main research content and contributions are summarized as follows:1.Research on semi supervised infrared scene modeling technology.Conventional technologies use satellite images or aerial images to obtain real geographic information about the target area,and then fed into supervised or unsupervised networks to obtain corresponding infrared scenes.However,there are semantic-level problems such as edge blur and detail loss in this process.To address this problem,this article combines supervised and unsupervised learning methods to obtain a semi supervised learning algorithm,named SISMGAN.SISMGAN is a semi supervised learning strategy for infrared image translation using generation confrontation networks,and an effective generation module,gradient graph based discrimination module,and perceptual loss function are designed to address the semantic differences contained in infrared generated images.SISMGAN uses a large amount of unpaired data and a small amount of paired images for joint training.Compared with the supervised algorithm Pix2 Pix,the peak signal-to-noise ratio and structural similarity index values are increased by 1.012 and0.015,respectively.Compared with the test results of other unsupervised translation networks,the performance is also greatly improved.2.Research on multi-band infrared scene modeling technology.The above output of the infrared scene only contains data from a specific wavelength band.To efficiently obtain data from different wavelength bands with different information,research on multi-domain transfer techniques is conducted.When different infrared scene data in various wavelength bands are fed into a multi-domain translation network,there may be an uneven distribution of grayscale brightness levels in the synthesized target band image.To solve this problem,this study improved the multi-domain image modeling algorithm to obtain a multi-band scene modeling algorithm BEISGAN.BEISGAN introduces the condition Unet network architecture into the generator module,and adds a spatial attention mechanism to make the learning map closer to the original data distribution.Through subjective and objective comparative tests,it is demonstrated that BEISGAN outperforms current multi domain algorithms in terms of both the effect and quality of infrared band images obtained by migration under the condition of unsupervised image training.3.Model network inference optimization.In the process of submitting GPU tasks to devices in a deep learning framework,task scheduling can consume a large amount of time,resulting in reduced efficiency in reasoning and computing.From the perspective of practical application,this thesis uses efficient advance scheduling technology to optimize the GAN inference and improve the calculation efficiency of the model.