Retinal vessel segmentation is a challenging medical task owing to small size of dataset,micro blood vessels and low image contrast.To address these issues,we introduce a novel convolutional neural network in this pap...Retinal vessel segmentation is a challenging medical task owing to small size of dataset,micro blood vessels and low image contrast.To address these issues,we introduce a novel convolutional neural network in this paper,which takes the advantage of both adversarial learning and recurrent neural network.An iterative design of network with recurrent unit is performed to refine the segmentation results from input retinal image gradually.Recurrent unit preserves high-level semantic information for feature reuse,so as to output a sufficiently refined segmentation map instead of a coarse mask.Moreover,an adversarial loss is imposing the integrity and connectivity constraints on the segmented vessel regions,thus greatly reducing topology errors of segmentation.The experimental results on the DRIVE dataset show that our method achieves area under curve and sensitivity of 98.17%and 80.64%,respectively.Our method achieves superior performance in retinal vessel segmentation compared with other existing state-of-the-art methods.展开更多
Semantic segmentation for mixed scenes of aerial remote sensing and road traffic is one of the key technologies for visual perception of flying cars.The State-of-the-Art(SOTA)semantic segmentation methods have made re...Semantic segmentation for mixed scenes of aerial remote sensing and road traffic is one of the key technologies for visual perception of flying cars.The State-of-the-Art(SOTA)semantic segmentation methods have made remarkable achievements in both fine-grained segmentation and real-time performance.However,when faced with the huge differences in scale and semantic categories brought about by the mixed scenes of aerial remote sensing and road traffic,they still face great challenges and there is little related research.Addressing the above issue,this paper proposes a semantic segmentation model specifically for mixed datasets of aerial remote sensing and road traffic scenes.First,a novel decoding-recoding multi-scale feature iterative refinement structure is proposed,which utilizes the re-integration and continuous enhancement of multi-scale information to effectively deal with the huge scale differences between cross-domain scenes,while using a fully convolutional structure to ensure the lightweight and real-time requirements.Second,a welldesigned cross-window attention mechanism combined with a global information integration decoding block forms an enhanced global context perception,which can effectively capture the long-range dependencies and multi-scale global context information of different scenes,thereby achieving fine-grained semantic segmentation.The proposed method is tested on a large-scale mixed dataset of aerial remote sensing and road traffic scenes.The results confirm that it can effectively deal with the problem of large-scale differences in cross-domain scenes.Its segmentation accuracy surpasses that of the SOTA methods,which meets the real-time requirements.展开更多
Background Recent advances in serial femtosecond crystallography(SFX)using X-ray free electron lasers(XFELs)have facilitated accurate structure determination for biological macromolecules.However,given the many fluctu...Background Recent advances in serial femtosecond crystallography(SFX)using X-ray free electron lasers(XFELs)have facilitated accurate structure determination for biological macromolecules.However,given the many fluctuations inherent in SFX,the acquisition of SFX data of sufficiently high quality still remains challenging.Method Aimed at enhancing the accuracy of SFX data,this study proposes an iterative refinement method to optimally match pairs of the observed and predicted reflections on the detector plane.This method features a combination of detector geometry optimization and diffraction model refinement in an alternate manner,concomitant with a cycle-by-cycle peak selection procedure.Result To demonstrate whether this iterative method is convergent and feasible,both numerical simulations and experimental tests have been performed.The results reveal that this method can gradually improve overall quality of the integrated SFX data and therefore accelerate the convergence of Monte Carlo integration,while simultaneously suppressing correlations inherent in certain parameters and precluding outliers to some extent during the refinement.Conclusion We have demonstrated that our iterative refinement method is applicable to both simulated and experimental SFX data.It is expected that this method could provide meaningful insights into the refinement of SFX data and take the step forward toward more accurate Monte Carlo integration.展开更多
Massive multiple-input multiple-output provides improved energy efficiency and spectral efficiency in 5 G. However it requires large-scale matrix computation with tremendous complexity, especially for data detection a...Massive multiple-input multiple-output provides improved energy efficiency and spectral efficiency in 5 G. However it requires large-scale matrix computation with tremendous complexity, especially for data detection and precoding. Recently, many detection and precoding methods were proposed using approximate iteration methods, which meet the demand of precision with low complexity. In this paper, we compare these approximate iteration methods in precision and complexity, and then improve these methods with iteration refinement at the cost of little complexity and no extra hardware resource. By derivation, our proposal is a combination of three approximate iteration methods in essence and provides remarkable precision improvement on desired vectors. The results show that our proposal provides 27%-83% normalized mean-squared error improvement of the detection symbol vector and precoding symbol vector. Moreover, we find the bit-error rate is mainly controlled by soft-input soft-output Viterbi decoding when using approximate iteration methods. Further, only considering the effect on soft-input soft-output Viterbi decoding, the simulation results show that using a rough estimation for the filter matrix of minimum mean square error detection to calculating log-likelihood ratio could provideenough good bit-error rate performance, especially when the ratio of base station antennas number and the users number is not too large.展开更多
Knowledge distillation(KD) enhances student network generalization by transferring dark knowledge from a complex teacher network. To optimize computational expenditure and memory utilization, self-knowledge distillati...Knowledge distillation(KD) enhances student network generalization by transferring dark knowledge from a complex teacher network. To optimize computational expenditure and memory utilization, self-knowledge distillation(SKD) extracts dark knowledge from the model itself rather than an external teacher network. However, previous SKD methods performed distillation indiscriminately on full datasets, overlooking the analysis of representative samples. In this work, we present a novel two-stage approach to providing targeted knowledge on specific samples, named two-stage approach self-knowledge distillation(TOAST). We first soften the hard targets using class medoids generated based on logit vectors per class. Then, we iteratively distill the under-trained data with past predictions of half the batch size. The two-stage knowledge is linearly combined, efficiently enhancing model performance. Extensive experiments conducted on five backbone architectures show our method is model-agnostic and achieves the best generalization performance.Besides, TOAST is strongly compatible with existing augmentation-based regularization methods. Our method also obtains a speedup of up to 2.95x compared with a recent state-of-the-art method.展开更多
This study introduces a novel sequential data assimilation method that uses conditional denoising score matching(CDSM).The CDSM leverages iterative refinement of noisy samples guided by conditional score functions to ...This study introduces a novel sequential data assimilation method that uses conditional denoising score matching(CDSM).The CDSM leverages iterative refinement of noisy samples guided by conditional score functions to achieve real-time state estimation by incorporating observational constraints at each time step.Unlike traditional methods,such as variational assimilation and Kalman filtering,which rely on Gaussian assumptions and can be computationally expensive because of iterations or ensembles,CDSM is based on stochastic differential equations(SDEs).展开更多
文摘Retinal vessel segmentation is a challenging medical task owing to small size of dataset,micro blood vessels and low image contrast.To address these issues,we introduce a novel convolutional neural network in this paper,which takes the advantage of both adversarial learning and recurrent neural network.An iterative design of network with recurrent unit is performed to refine the segmentation results from input retinal image gradually.Recurrent unit preserves high-level semantic information for feature reuse,so as to output a sufficiently refined segmentation map instead of a coarse mask.Moreover,an adversarial loss is imposing the integrity and connectivity constraints on the segmented vessel regions,thus greatly reducing topology errors of segmentation.The experimental results on the DRIVE dataset show that our method achieves area under curve and sensitivity of 98.17%and 80.64%,respectively.Our method achieves superior performance in retinal vessel segmentation compared with other existing state-of-the-art methods.
基金supported by the National Key Research and Development of China(No.2022YFB2503400).
文摘Semantic segmentation for mixed scenes of aerial remote sensing and road traffic is one of the key technologies for visual perception of flying cars.The State-of-the-Art(SOTA)semantic segmentation methods have made remarkable achievements in both fine-grained segmentation and real-time performance.However,when faced with the huge differences in scale and semantic categories brought about by the mixed scenes of aerial remote sensing and road traffic,they still face great challenges and there is little related research.Addressing the above issue,this paper proposes a semantic segmentation model specifically for mixed datasets of aerial remote sensing and road traffic scenes.First,a novel decoding-recoding multi-scale feature iterative refinement structure is proposed,which utilizes the re-integration and continuous enhancement of multi-scale information to effectively deal with the huge scale differences between cross-domain scenes,while using a fully convolutional structure to ensure the lightweight and real-time requirements.Second,a welldesigned cross-window attention mechanism combined with a global information integration decoding block forms an enhanced global context perception,which can effectively capture the long-range dependencies and multi-scale global context information of different scenes,thereby achieving fine-grained semantic segmentation.The proposed method is tested on a large-scale mixed dataset of aerial remote sensing and road traffic scenes.The results confirm that it can effectively deal with the problem of large-scale differences in cross-domain scenes.Its segmentation accuracy surpasses that of the SOTA methods,which meets the real-time requirements.
基金This work was financially supported by the grants from the Strategic Priority Research Program of CAS(XDB08030103)the National Natural Science Foundation of China(31570744,31670059).
文摘Background Recent advances in serial femtosecond crystallography(SFX)using X-ray free electron lasers(XFELs)have facilitated accurate structure determination for biological macromolecules.However,given the many fluctuations inherent in SFX,the acquisition of SFX data of sufficiently high quality still remains challenging.Method Aimed at enhancing the accuracy of SFX data,this study proposes an iterative refinement method to optimally match pairs of the observed and predicted reflections on the detector plane.This method features a combination of detector geometry optimization and diffraction model refinement in an alternate manner,concomitant with a cycle-by-cycle peak selection procedure.Result To demonstrate whether this iterative method is convergent and feasible,both numerical simulations and experimental tests have been performed.The results reveal that this method can gradually improve overall quality of the integrated SFX data and therefore accelerate the convergence of Monte Carlo integration,while simultaneously suppressing correlations inherent in certain parameters and precluding outliers to some extent during the refinement.Conclusion We have demonstrated that our iterative refinement method is applicable to both simulated and experimental SFX data.It is expected that this method could provide meaningful insights into the refinement of SFX data and take the step forward toward more accurate Monte Carlo integration.
文摘Massive multiple-input multiple-output provides improved energy efficiency and spectral efficiency in 5 G. However it requires large-scale matrix computation with tremendous complexity, especially for data detection and precoding. Recently, many detection and precoding methods were proposed using approximate iteration methods, which meet the demand of precision with low complexity. In this paper, we compare these approximate iteration methods in precision and complexity, and then improve these methods with iteration refinement at the cost of little complexity and no extra hardware resource. By derivation, our proposal is a combination of three approximate iteration methods in essence and provides remarkable precision improvement on desired vectors. The results show that our proposal provides 27%-83% normalized mean-squared error improvement of the detection symbol vector and precoding symbol vector. Moreover, we find the bit-error rate is mainly controlled by soft-input soft-output Viterbi decoding when using approximate iteration methods. Further, only considering the effect on soft-input soft-output Viterbi decoding, the simulation results show that using a rough estimation for the filter matrix of minimum mean square error detection to calculating log-likelihood ratio could provideenough good bit-error rate performance, especially when the ratio of base station antennas number and the users number is not too large.
基金supported by the National Natural Science Foundation of China (62176061)。
文摘Knowledge distillation(KD) enhances student network generalization by transferring dark knowledge from a complex teacher network. To optimize computational expenditure and memory utilization, self-knowledge distillation(SKD) extracts dark knowledge from the model itself rather than an external teacher network. However, previous SKD methods performed distillation indiscriminately on full datasets, overlooking the analysis of representative samples. In this work, we present a novel two-stage approach to providing targeted knowledge on specific samples, named two-stage approach self-knowledge distillation(TOAST). We first soften the hard targets using class medoids generated based on logit vectors per class. Then, we iteratively distill the under-trained data with past predictions of half the batch size. The two-stage knowledge is linearly combined, efficiently enhancing model performance. Extensive experiments conducted on five backbone architectures show our method is model-agnostic and achieves the best generalization performance.Besides, TOAST is strongly compatible with existing augmentation-based regularization methods. Our method also obtains a speedup of up to 2.95x compared with a recent state-of-the-art method.
基金supported by the National Natural Science Foundation of China(grant number 42450178).
文摘This study introduces a novel sequential data assimilation method that uses conditional denoising score matching(CDSM).The CDSM leverages iterative refinement of noisy samples guided by conditional score functions to achieve real-time state estimation by incorporating observational constraints at each time step.Unlike traditional methods,such as variational assimilation and Kalman filtering,which rely on Gaussian assumptions and can be computationally expensive because of iterations or ensembles,CDSM is based on stochastic differential equations(SDEs).
