基于阴影解耦和重参的轻量级阴影检测算法  

Fast and Efficient Shadow Detection Algorithm via Shadow Decoupling and Reparameterization

作　　者：陈珏宇 杨雨泓 邢冠宇 刘艳丽[2] CHEN Jueyu;YANG Yuhong;XING Guanyu;LIU Yanli(College of Computer Sci.,Sichuan Univ.,Chengdu 610065,China;National Key Lab.of Fundamental Sci.on Synthetic Vision,Sichuan Univ.,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,四川成都610065 [2]四川大学视觉合成图形图像技术国防重点学科实验室,四川成都610065

出　　处：《工程科学与技术》2024年第5期297-306,共10页Advanced Engineering Sciences

基　　金：国家自然科学基金项目(61972271);四川省科技计划项目(2023YFS0454)。

摘　　要：由于阴影边界上的像素数量往往远小于阴影内部区域的像素数量,准确检测阴影边界区域相比于检测阴影内部区域像素更困难。为了提高在阴影边界上的检测准确率,本文提出了一种新颖且高效的轻量级边界感知阴影检测算法RBNet。首先,将图像分为阴影区域和非阴影区域,并分别采用距离变换将图像的阴影区域、非阴影区域的边界部分和主体解耦。接着,提出RBNet的解码器通过学习阴影区域的边界特征,在监督学习中平衡阴影边界与区域内部像素的对阴影检测性能的影响。然后,在RBNet中设计了一种可多分支融合的结构化重参模块RepConv,通过重参进行多分支融合来减少模型参数量、降低模型计算量并提高模型推理速度。将提出的RB-Net和其他常见的阴影检测算法进行了一系列阴影检测对比实验和算法模型消融对比实验。实验结果表明,本文提出的阴影检测算法RBNet不仅模型最小,且具有最快的推理速度,同时,在性能上优于现有的阴影检测算法。RBNet可在小型移动端设备上进行广泛应用,结合阴影去除算法,显著提高目标检测或分割任务的准确性。Since the number of pixels along shadow boundaries is often significantly smaller than that of pixles within shadow regions,accurately detecting shadow boundary areas is more challenging than detecting pixels within the shadow interior.To improve detection accuracy at shadow boundaries,a novel and efficient lightweight boundary-aware shadow detection algorithm called RBNet was proposed in this paper.During the supervised training phase,the input image was divided into shadow and non-shadow regions,and the distance transforms were applied to decouple the shadow and non-shadow regions into a boundary part and a body part respectively.Secondly,the boundary features of shadow regions were learned and the impact of shadow boundaries and interior pixels were balanced in the loss function of RBNet.Additionally,a multi-branch fusion structured re-parameterization module named RepConv was designed in RBNet.Through re-parameterization,the model parameters and computational cost were reduced,and the inference speed was improved.A series of shadow detection comparison experiments and algorithm model comparison experiments between the proposed RBNet and other common shadow detection algorithms is conducted in the paper.Experimental results demonstrate that the proposed shadow detection algorithm,RBNet,not only has the smallest model size but also achieves the fastest inference speed,while outperforming existing shadow detection algorithms in terms of BER performance.RBNet is highly applicable to mobile devices.When combined with shadow removal algorithms,it can significantly enhance the accuracy of object detection or segmentation tasks.

关 键 词：阴影检测 深度学习 图像解耦 卷积网络 重参 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]