Plug-and-play priors are popular for solving illposed imaging inverse problems. Recent efforts indicate that the convergence guarantee of the imaging algorithms using plug-andplay priors relies on the assumption of bo...Plug-and-play priors are popular for solving illposed imaging inverse problems. Recent efforts indicate that the convergence guarantee of the imaging algorithms using plug-andplay priors relies on the assumption of bounded denoisers. However, the bounded properties of existing plugged Gaussian denoisers have not been proven explicitly. To bridge this gap, we detail a novel provable bounded denoiser termed as BMDual,which combines a trainable denoiser using dual tight frames and the well-known block-matching and 3D filtering(BM3D)denoiser. We incorporate multiple dual frames utilized by BMDual into a novel regularization model induced by a solver. The proposed regularization model is utilized for compressed sensing magnetic resonance imaging(CSMRI). We theoretically show the bound of the BMDual denoiser, the bounded gradient of the CSMRI data-fidelity function, and further demonstrate that the proposed CSMRI algorithm converges. Experimental results also demonstrate that the proposed algorithm has a good convergence behavior, and show the effectiveness of the proposed algorithm.展开更多
Quanta image sensors(QIS) are a new type of singlephoton imaging device that can oversample the light field to generate binary bit-streams. The reconstruction for QIS refers to the recovery of original scenes from the...Quanta image sensors(QIS) are a new type of singlephoton imaging device that can oversample the light field to generate binary bit-streams. The reconstruction for QIS refers to the recovery of original scenes from these binary measurements.Conventional reconstruction algorithms for QIS generally depend solely on one instantiated prior and are certainly insufficient for capturing the statistical properties over high-dimensional space.On the other hand, deep learning-based methods have shown promising performance, due to their excellent ability to learn feature representations from relevant databases. However, most deep models only focus on exploring local features while generally overlooking long-range similarity. In view of this, a dual-prior integrated reconstruction algorithm for QIS(DPI-QIS) is proposed, which combines a deep prior with a non-local self-similarity one using the multi-agent consensus equilibrium(MACE)framework. In comparison to the approaches that utilize a single prior, DPI-QIS fits the reconstruction model sufficiently by leveraging the respective merits of both priors. An effective yet flexible MACE framework is employed to integrate the physical forward model allying with the two prior-based models to achieve an overall better result. Extensive experiments demonstrate that the proposed algorithm achieves state-of-the-art performance in terms of objective and visual perception at multiple oversampling factors, while having stronger robustness to noise.展开更多
The concept of the binary sequence pair is generalized from a single binary sequence. Binary sequence pairs are applied in many fields of radar, sonar or communication systems, in which signals with optimal periodic c...The concept of the binary sequence pair is generalized from a single binary sequence. Binary sequence pairs are applied in many fields of radar, sonar or communication systems, in which signals with optimal periodic correlation are required. Several types of almost perfect binary sequence pairs of length T = 2q are constructed, where q is an odd number. These almost perfect binary sequence pairs are based on binary ideal sequence or binary ideal two-level correlation sequence pairs by using Chinese remainder theorem. For these almost perfect binary sequence pairs with good balanced property, their corresponding divisible difference set pairs(DDSPs) are also derived.展开更多
Weakly supervised object localization mines the pixel-level location information based on image-level annotations.The traditional weakly supervised object localization approaches exploit the last convolutional feature...Weakly supervised object localization mines the pixel-level location information based on image-level annotations.The traditional weakly supervised object localization approaches exploit the last convolutional feature map to locate the discriminative regions with abundant semantics.Although it shows the localization ability of classification network,the process lacks the use of shallow edge and texture features,which cannot meet the requirement of object integrity in the localization task.Thus,we propose a novel shallow feature-driven dual-edges localization(DEL)network,in which dual kinds of shallow edges are utilized to mine entire target object regions.Specifically,we design an edge feature mining(EFM)module to extract the shallow edge details through the similarity measurement between the original class activation map and shallow features.We exploit the EFM module to extract two kinds of edges,named the edge of the shallow feature map and the edge of shallow gradients,for enhancing the edge details of the target object in the last convolutional feature map.The total process is proposed during the inference stage,which does not bring extra training costs.Extensive experiments on both the ILSVRC and CUB-200-2011 datasets show that the DEL method obtains consistency and substantial performance improvements compared with the existing methods.展开更多
基金supported in part by the National Natural Science Foundation of China (62371414,61901406)the Hebei Natural Science Foundation (F2020203025)+2 种基金the Young Talent Program of Universities and Colleges in Hebei Province (BJ2021044)the Hebei Key Laboratory Project (202250701010046)the Central Government Guides Local Science and Technology Development Fund Projects(216Z1602G)。
文摘Plug-and-play priors are popular for solving illposed imaging inverse problems. Recent efforts indicate that the convergence guarantee of the imaging algorithms using plug-andplay priors relies on the assumption of bounded denoisers. However, the bounded properties of existing plugged Gaussian denoisers have not been proven explicitly. To bridge this gap, we detail a novel provable bounded denoiser termed as BMDual,which combines a trainable denoiser using dual tight frames and the well-known block-matching and 3D filtering(BM3D)denoiser. We incorporate multiple dual frames utilized by BMDual into a novel regularization model induced by a solver. The proposed regularization model is utilized for compressed sensing magnetic resonance imaging(CSMRI). We theoretically show the bound of the BMDual denoiser, the bounded gradient of the CSMRI data-fidelity function, and further demonstrate that the proposed CSMRI algorithm converges. Experimental results also demonstrate that the proposed algorithm has a good convergence behavior, and show the effectiveness of the proposed algorithm.
基金supported by Hebei Natural Science Foundation(F2022203030)the National Natural Science Foundation of China(61471313)。
文摘Quanta image sensors(QIS) are a new type of singlephoton imaging device that can oversample the light field to generate binary bit-streams. The reconstruction for QIS refers to the recovery of original scenes from these binary measurements.Conventional reconstruction algorithms for QIS generally depend solely on one instantiated prior and are certainly insufficient for capturing the statistical properties over high-dimensional space.On the other hand, deep learning-based methods have shown promising performance, due to their excellent ability to learn feature representations from relevant databases. However, most deep models only focus on exploring local features while generally overlooking long-range similarity. In view of this, a dual-prior integrated reconstruction algorithm for QIS(DPI-QIS) is proposed, which combines a deep prior with a non-local self-similarity one using the multi-agent consensus equilibrium(MACE)framework. In comparison to the approaches that utilize a single prior, DPI-QIS fits the reconstruction model sufficiently by leveraging the respective merits of both priors. An effective yet flexible MACE framework is employed to integrate the physical forward model allying with the two prior-based models to achieve an overall better result. Extensive experiments demonstrate that the proposed algorithm achieves state-of-the-art performance in terms of objective and visual perception at multiple oversampling factors, while having stronger robustness to noise.
基金supported by the National Natural Science Foundation of China(6160140161501395+6 种基金6160139961671402)Natural Science Foundation of Hebei Province(F2015203150F2016203293F2016203312)Natural Science Research Programs of Hebei Educational Committee(QN2016120)the Independent Research Programs for Young Teachers of Yanshan University(15LGB013)
文摘The concept of the binary sequence pair is generalized from a single binary sequence. Binary sequence pairs are applied in many fields of radar, sonar or communication systems, in which signals with optimal periodic correlation are required. Several types of almost perfect binary sequence pairs of length T = 2q are constructed, where q is an odd number. These almost perfect binary sequence pairs are based on binary ideal sequence or binary ideal two-level correlation sequence pairs by using Chinese remainder theorem. For these almost perfect binary sequence pairs with good balanced property, their corresponding divisible difference set pairs(DDSPs) are also derived.
基金This work was partly supported by National Natural Science Foundation of China(No.62072394)Natural Science Foundation of Hebei Province,China(No.F2021203019)Hebei Key Laboratory Project,China(No.202250701010046).
文摘Weakly supervised object localization mines the pixel-level location information based on image-level annotations.The traditional weakly supervised object localization approaches exploit the last convolutional feature map to locate the discriminative regions with abundant semantics.Although it shows the localization ability of classification network,the process lacks the use of shallow edge and texture features,which cannot meet the requirement of object integrity in the localization task.Thus,we propose a novel shallow feature-driven dual-edges localization(DEL)network,in which dual kinds of shallow edges are utilized to mine entire target object regions.Specifically,we design an edge feature mining(EFM)module to extract the shallow edge details through the similarity measurement between the original class activation map and shallow features.We exploit the EFM module to extract two kinds of edges,named the edge of the shallow feature map and the edge of shallow gradients,for enhancing the edge details of the target object in the last convolutional feature map.The total process is proposed during the inference stage,which does not bring extra training costs.Extensive experiments on both the ILSVRC and CUB-200-2011 datasets show that the DEL method obtains consistency and substantial performance improvements compared with the existing methods.