Research On Algorithms Of Nonuniformity Correction On Infrared Focal Plane Array Based On Neural Networks And FPGA Implementation

The outbreak of COVID-19 is threatening health of human all over the world.In order to detect the body temperature of pedestrians and control the spread of the epidemic,infrared thermal imaging systems are widely used.So the quality of infrared image need to be much better.Infrared focal plane array is one of the most important core devices in the infrared detector,However,due to the limitations of processing technology,materials and design methods,the response of each unit on the infrared focal plane to the uniform radiation of the standard blackbody is inconsistent.And it will cause the fixed pattern noise on the image,which greatly affects the quality of IR images.Therefore,it's very important to implement nonuniformity correction before getting the infrared images.The primary method of non-uniformity correction should be started from the material of the detectors and the process,and it involves several frontier sciences.So,the most convenient and effective way is correcting through digital signal processing,which can be summarized into calibration-based non-uniformity correction and scene-based non-uniformity correction algorithms.Firstly,this paper describes the main scene based non-uniformity correction methods.Comparing them with hardware implementation,we eventually focus on the BP artificial neural networks correction algorithm.We optimize the traditional BP neural networks correction algorithm to obtain a better correction effect.Finally,real-time correction is realized.The specific work of this dissertation is as follows.(1)The traditional BP neural networks algorithm regards the four-neighborhood mean image as the expecting image and processes the original images by multiplying and adding some certain coefficients.Using the error and minimum gradient method to make the output correction image continuously approximate to the expecting image,and finally outputs the correction image.However,the speed of convergence becomes slower and it also has ghosting problem.In order to solve these problems,this dissertation improves the traditional neural networks method of finding the expected image by linking the expected image with the correction result to form a closed loop,which speeds up the convergence speed and solves the "ghosting" problem by using the peripheral of the nine-axis sensor.(2)Since the IR images are output line by line and the scene based correction needs to be done in real time,which is more computationally intensive.So,in this paper,an FPGA-based hardware acceleration method is being proposed.In the hardware implementation,the detector driver module,filtering module,parameter correction module and coefficient storage module are designed in hardware.Then the pipeline architecture is designed based on the processing flow of these modules.In order to accelerate the fast access to DDR,a DMA structure is designed for data buffering between DDR and FPGA in this dissertation.(3)Firstly,we use Python for software simulation to verify the correctness of the algorithm.And then realizing the real-time calibration on FPGA.Finally,we analyze the result by observation and some parameters.Compared with the traditional neural networks nonuniformity calibration algorithm,the improved correction algorithm in this paper improves the correction speed by 23%and the value of PSNR by about 3dB.