基于神经网络的非线性滤波残差估计非均匀校正算法  

Nonuniform Correction Algorithm for Residual Estimation of Nonlinear Filter Based on Neural Network

作　　者：张科航 宋鸿飞[1] 谭文 曹文晓[1] 郭飞 ZHANG Kehang;SONG Hongfei;TAN Wen;CAO Wenxiao;GUO Fei(School of Opto-Electronic Engineering,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学光电工程学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2023年第5期66-74,共9页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技厅发展计划项目(20200401066GX)。

摘　　要：目前由于非均匀性技术的限制,红外焦平面阵列具有条纹非均匀性,这种非均匀性一般是一种固定的模式噪声,它会在原有非均匀性的基础上随着时间和温度的变化而变化。传统的残差计算忽略了非均匀性的特点,场景细节的变化会被误认为非均匀性而被过滤掉。提出了一种基于场景的条纹非均匀性校正算法。该方法采用非线性滤波的方法滤除单柱图像的非均匀性,并计算与原始图像的实际残差,然后利用前一帧的预测残差和实际残差得到当前残差。最后,根据场景鬼影抑制参数自适应校正计算增益系数和偏置系数。实验结果表明,提出的算法在200帧测试图像中滤除了61.3%的非均匀性,与其他自适应校正算法相比,非均匀性校正效果更明显,并且在一定程度上有效地保护了图像边缘,收敛速度快,质量高,能有效去除列条纹和非均匀随机噪声。Due to the limitation of the non-uniformity technology,the infrared focal plane array has stripes non-uniformity,the non-uniformity is generally a fixed pattern noise,which will change with time and temperature based on the original non-uniformity.The traditional residual calculation ignores the characteristics of non-uniformity,and the changes in scene details will be mistaken for non-uniformity and filtered out.In this paper,we proposed a scene-based stripes non-uniformity correction algorithm.The proposed method uses nonlinear filtering to filter out the non-uniformity of a single column and calculate the actual residual with the original image.Then use the predicted residue of the previous frame and the actual residue to obtain the current residue.Finally,we calculate the gain coefficient and bias coefficient according to the adaptive correction of scene ghost suppression parameters.The experimental results show that the proposed algorithm can eliminate 61.3%of the non-uniformity in 200 frames of test images.Compared with other adaptive correction algorithms,the non-uniformity correction effect is more obvious,and the image edge is effectively protected to a certain extent.The convergence speed is fast,the quality is high,and it can effectively remove column stripes and non-uniform random noise.

关 键 词：非均匀性 条纹噪声 非线性滤波 残差神经网络 自适应校正 

分 类 号：TP33[自动化与计算机技术—计算机系统结构] TN495[自动化与计算机技术—计算机科学与技术]