重庆理工大学学报2025，Vol.39Issue(7)：139-147,9.DOI:10.3969/j.issn.1674-8425(z).2025.04.018
可学习RPCA深度网络的视频显著性检测方法
Video saliency detection method for learnable RPCA deep network
摘要
Abstract
Video saliency detection has garnered keen academic interest in computer vision with the primary goal of extracting the most significant regions or objects from video sequences.Currently,it is widely used in video surveillance,medical imaging,and behavioral analysis.In video surveillance,for example,real-time abnormal behavior detection is achieved by extracting moving objects(such as people or vehicles),thereby improving public safety.In medical imaging,separating the lesion area from healthy tissue enables doctors to make more accurate diagnoses,thereby improving the accuracy and effectiveness of treatment.In addition,video saliency detection has great potentials in autonomous driving and virtual reality. Robust Principal Component Analysis(RPCA),as a classic video saliency detection method,has been widely used thanks to its clear mathematical theoretical foundation,complete interpretability,and good foreground and background separation effect.RPCA effectively handles noise and dynamic background interference by decomposing video frames into two parts:low rank background and sparse foreground.However,traditional RPCA algorithms involve high computational complexity and achieve low efficiency when processing high-dimensional data and real-time videos.In addition,traditional RPCA requires manual selection of parameters in the algorithm,which needs readjustment for different videos.This not only adds labor costs but also limits its practical application. In recent years,deep learning has made great achievements in computer vision.Its powerful feature learning ability and end-to-end optimization mechanism provide new ways for video saliency detection.Inspired by the success of deep learning,this paper combines the theoretical rigor of traditional RPCA optimization methods with the powerful learning capabilities of deep learning,and proposes a video foreground background separation method based on learnable RPCA deep networks.The method not only retains the mathematical theoretical foundation of RPCA,but also achieves adaptive learning of parameters through deep learning,thereby markedly improving the performance and efficiency of the algorithm.First,by analyzing the relationship between sparse regularization functions,low rank regularization functions,and threshold functions,a learnable piecewise linear threshold function is designed to replace the traditional fixed threshold function.This design serves as the activation function of the network,where the parameters are adaptively learned through data-driven methods,thus avoiding the limitations of manually adjusting parameters in traditional RPCA.Then,one iteration of the optimization process of the PCP(Principal Component Pursuit)algorithm is employed as a layer of the network,and a deep network framework is built layer by layer.By combining the backpropagation algorithm,the minimization of the loss function is achieved,and an end-to-end parameter learning mechanism is designed.This mechanism not only improves the automation level of the algorithm,but also significantly enhances computational efficiency.Finally,the effectiveness of the method is verified through extensive experiments.Results show the method outperforms the other five algorithms in terms of visual effects and F-measure values(average of 0.789 5).Compared with that of the best performing Truncated Kernel Norm(TNN)optimization algorithm among these five methods,the F-measure value is 9.89％higher.Moreover,the method outperforms the other ones in computation time,demonstrating its higher efficiency and superiority.关键词
鲁棒主成分分析/深度学习/自适应正则化/视频显著性检测Key words
robust principal component analysis/deep learning/adaptive regularization/video foreground and background separation分类
信息技术与安全科学引用本文复制引用
袁薛程,肖锋,张文娟,沈超,药嘉怡..可学习RPCA深度网络的视频显著性检测方法[J].重庆理工大学学报,2025,39(7):139-147,9.基金项目
国家自然科学基金项目(62171361) （62171361）
国家自然科学基金青年项目(52302505) （52302505）
陕西省科技厅重点研发计划一般项目(工业领域)(2023-YBGY-027) （工业领域）
陕西省教育厅一般专项科研计划项目(22JK0412) （22JK0412）
