RGB深度图像显著性目标检测方法设计  

Design of salience target detection method for RGB depth images

作　　者：王雨夕 徐杨[1] 袁旭祥 WANG Yuxi;XU Yang;YUAN Xuxiang(School of Computer and Software Engineering,University of Science and Technology Liaoning,Anshan 114051,China)

机构地区：[1]辽宁科技大学计算机与软件工程学院,辽宁鞍山114051

出　　处：《液晶与显示》2025年第4期607-616,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.61775169);辽宁省教育厅科研项目(No.LJKZ0310)。

摘　　要：为了高效利用深度特征信息辅助显著性检测,实现对不同尺度特征信息的融合,本文提出了一种基于CDINet算法改进的RGB-D图像显著性目标检测算法。首先,添加了多尺度特征融合模块用来加强编码器和解码器之间特征信息的传输,有效减少浅层特征丢失,通过辅助解码器的跳跃连接获得更多的显著物体的特征信息。接着,在CDINet的网络结构尾部连接了一个循环注意力模块,通过使用面向记忆的场景理解功能,逐渐优化局部细节。最后,对损失函数进行调整,使用一致性增强损失(CEL)处理因为不同尺度特征融合产生的空间一致性等问题,并在不增加参数的情况下均匀突出显著区域。实验结果表明,改进后的模型与原CDINet算法模型相比,在LFSD数据集上的F-measure提高了0.6%,MAE下降了0.4%;在STERE数据集上的F-measure提高了0.4%,S-measure提升了0.5%。相对于其他算法模型,本模型基本满足检测性能更好、适应性更高等要求。In order to efficiently use depth feature information to assist salient object detection,the fusion of different scale feature information is realized.In this paper,an improved salient object detection algorithm for RGB-D image saliency based on CDINet algorithm is proposed.Firstly,a multi-scale feature fusion module is added to enhance the transmission of feature information between encoder and decoder,so as to effectively reduce shallow feature loss,and obtain more feature information of salient objects through the jump connection of auxiliary decoder.Next,a circular attention module is connected at the tail of the CDINet’s network structure,which gradually optimizes local details by using memory-oriented scene understanding.Finally,the loss function is adjusted,and the consistency enhanced loss(CEL)is used to deal with the spatial consistency caused by the fusion of different scale features,and the salient areas are uniformly highlighted without increasing parameters.The experimental results show that compared with the original CDINet algorithm model,the improved model has an F-measure increase of 0.6%and a MAE decrease of 0.4%on LFSD data set,and an F-measure increase of 0.4%and a S-measure decrease of 0.5%on STERE data set.Compared with other algorithm models,this model basically meets the requirements of better detection performance and higher adaptability.

关 键 词：显著性目标检测 计算机视觉 边缘检测 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]