对比约束下的非局部关联单图像去反光级联算法  

A single image reflection removal cascaded algorithm using non-local correlation and contrast constraint

作　　者：罗超 缪君[1] 郑义林 华锋 储珺[2] LUO Chao;MIAO Jun;ZHENG Yi-lin;HUA Feng;Chu Jun(College of Aeronautical Manufacturing Engineering,Nanchang Hangkong University,Nanchang 330063;Institute of Computer Vision,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学航空制造工程学院,江西南昌330063 [2]南昌航空大学计算机视觉研究所,江西南昌330063

出　　处：《计算机工程与科学》2024年第5期861-871,共11页Computer Engineering & Science

基　　金：国家自然科学基金(62162045,62366032)。

摘　　要：图像中存在的反光不但显著降低了图像质量,而且严重影响了后续的计算机视觉任务。因此,提出了一种对比约束下的非局部关联单图像去反光级联算法NCRR,该算法通过LSTM传递跨级联信息的双支路方式,利用反光特征和背景特征相互补充信息并迭代细化预测精度,使2条支路的预测效果相互促进。为了便于多个级联步骤的训练,提出了一种正负对比的正则化损失,将背景图像和原图像的特征分别作为正、负样本,确保目标图像在表示空间中拉近背景图像,推远原图像,缩小预测范围,较好地缓解不适定性问题。此外,提出了一种高效、计算量少的非局部关联预测模块,它能获取十字交叉路径上所有像素的上下文信息。通过进一步级联操作,使每个像素捕获整幅图像长距离的依赖关系,能利用周围像素点信息来预测被强反光遮挡的背景信息。实验结果表明,本文算法能够有效去除玻璃的强反光,并且玻璃去反光的评估结果都超过了其他对比算法,具有较好的鲁棒性。Reflection in the image not only significantly reduces the image quality,but also seriously affects the subsequent computer vision tasks.So proposed a single image reflection removal cascaded algorithm using non local correlation and contrast constraint.This algorithm utilizes a dual-branch approach for LSTM-based information propagation across cascades.It employs reflection and background features to complement each other and iteratively refine prediction accuracy,ensuring mutual enhancement of the two branches'prediction results.To facilitate training for multiple cascade steps,a positive-negative contrastive regularization loss is introduced.This loss treats background images and original images features as positive and negative samples,respectively.This ensures that the target image is brought closer to the background image while moving away from the original image in the representation space,narrowing the prediction range and effectively alleviating the ill-posed problem.Additionally,an efficient,low-computational-cost non-local correlation prediction module is proposed,capable of capturing contextual information for all pixels along cross paths.Through further cascade operations,each pixel captures long-distance dependencies across the entire image,enabling the use of surrounding point information to predict background information obscured by strong reflections.Experimental results demonstrate that,compared to current algorithms,the proposed algorithm achieves superior results and exhibits robust performance.

关 键 词：图像反光 级联 正则化 上下文信息 远程依赖 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]