改进的神经网络非均匀性校正算法研究  

Research on Improved Neural Network Non-uniformity Correction Algorithm

作　　者：陈芳林[1,2] 张宝辉[2] 汪贵华[1] 杨开峰[2] 于世孔 

机构地区：[1]南京理工大学电子工程与光电技术学院,南京210094 [2]北方夜视科技集团有限公司南京研发中心,南京211106

出　　处：《科学技术与工程》2016年第33期215-220,共6页Science Technology and Engineering

摘　　要：针对传统神经网络非均匀性校正算法所存在的像素点期望值估计不准确、场景长时间静止时校正图像发生衰退以及校正图像出现鬼影的问题,在原有的神经网络算法的基础上做了进一步的改进。主要包括三个部分:首先,计算场景的局部方差,设置合理判断阈值,区别不同的场景区域;再通过比较前后两帧场景的绝对误差值,判别场景是否静止,来控制校正参数是否更新;然后利用图像配准的方法,计算出帧间位移,对校正步长做出修正。最后与传统神经网络算法进行对比实验,结果表明,改进的神经网络算法在鬼影抑制和非均匀性校正方面都具有较好的效果。The traditional Neural network non-uniformity correction algorithms exists following questions： the expectations of pixels are estimated inaccurate,when the scene for a long time still the corrected image will degrade,and the ghosting will appear in corrected image. Neural network algorithm was improued based on the original algorithm made further improvements to solve the above questions. First,the local variance of scene was calculated,and a reasonable judgment thresold was set to distinguish between different areas of the scene. Next,the absolute error of two adjacent scenes was compared,and the scene state was discriminater which is stationary or not to determine whether to update the correction parameters. Then,the image registration was used to calculate the interframe displacement and corrects the correction step. Finally,the experiment is done to compare with the traditional neural network algorithm. Experimental results show that the improved neural network algorithm has better effect correction in suppressing ghosting and non-uniformity.

关 键 词：神经网络 非均匀性 图像配准 红外图像 

分 类 号：TN911.73[电子电信—通信与信息系统]