Research On Nonuniformity Correction And Parallel Acceleration Technology In Infrared Focal Plane

Staring imaging technology based on infrared focal plane array has become the main direction of the development of infrared imaging technology.Compared with the single element detector system,infrared focal plane array pixel has high sensitivity and is able to acquire more scene detail information and higher variable frame rate.But the nonuniform response of array pixels severely limits the temperature resolution and imaging signal-tonoise ratio of infrared imaging system.Therefore,the nonuniformity correction of infrared focal plane has become one of the most important research contents in the field of infrared imaging.The nonuniformity correction method has two main categories: nonuniformity correction algorithms based on infrared reference radiation sources and scene-based adaptive correction methods.The reference-based nonuniformity correction method has high computational efficiency and is suitable for real-time applications,but its parameter calibration requires shutdown and it is difficult to solve the problem of parameter drift caused by scene change.The scene-based adaptive correction method can adjust the correction parameters in real time as the scene changes,and its adaptability is stronger.But its large amount of calculation is a major obstacle to its application.To solve the problems mentioned above,this thesis mainly studies the technical methods to improve the accuracy and computational efficiency of scene-based adaptive correction methods.This thesis first introduces the total variation model that achieves significant results in the field of image denoising,and analyzes the bilateral total variation model and the nonlocal total variation model.The denoising effect of the former and the processing speed of the latter still needs to be improved.In this thesis,the conception of patch in the nonlocal total variation model is introduced in the bilateral total variation model,and a patch-based bilateral total variation model is proposed.With PSNR improved by 1d B,experimental results show that the new model has better denoising results,and the processing speed is 5 times faster than that of the nonlocal total variation modelSubsequently,this thesis introduces a classic neural network correction method.This method does not need to manually calibrate the correction parameters.But the detail information of the infrared image after correction is seriously lost.The patch-based bilateral total variation regularization term is able to preserve the detail information of the image while smoothing the nonuniform noise of the infrared image.In this thesis,the patch-based bilateral total variation regularization term is introduced into the neural network correction method.The experimental results show that the new algorithm has higher correction accuracy of infrared image and better ability to keep the details with PSNR improved by 2d B.In addition,this thesis proposes a new deghosting adaptive strategy,which can effectively speed up the slow convergence speed resulted from the threshold strategy which the current deghosting methods use.In order to solve the problem that the scene-based adaptive correction algorithm has a large amount of calculation and cannot realize real-time processing,this thesis firstly decompose the neural network correction algorithm based on the total variation which uses patch-based bilateral conception according to the implementation process.Through Open CL programming standard,the parallelization of each step of algorithm is implemented on the GPU device.After the steps above,parallel algorithm optimization is done by caching multiplexed data into local memory,caching fixed template data into constant memory and merging kernel functions.The experimental results show that the processing speed of the neural network correction algorithm based on the patch-based bilateral total variation after CPU + GPU heterogeneous acceleration is increased by more than 20 times compared with the CPU serial mode.The parallel algorithm can correct 20-30 frames of infrared images per second,basically meeting the needs of real-time correction of infrared images.