- 标题
- 摘要
- 关键词
- 实验方案
- 产品
-
[IEEE 2018 25th IEEE International Conference on Image Processing (ICIP) - Athens, Greece (2018.10.7-2018.10.10)] 2018 25th IEEE International Conference on Image Processing (ICIP) - Stereo Generation from a Single Image Using Deep Residual Network
摘要： In this paper, we propose a framework to generate stereoscopic content from a single image using the relative depth label predicted from deep residual network. Specifically, our framework first obtains a coarse relative depth label from the network and refines it to painting depth by sampling and interpolation, then an unsupervised clustering algorithm is employed to separate pixels of different depths into different layers to generate stereoscopic images. Experimental results with good visual effects demonstrate that the proposed method can be generally applied in both outdoor and indoor scenes. Meanwhile the quantitative results on relative depth estimation from a single image are comparable to state-of-the-art. Further experiments show the application possibility of our method in VR and panorama.
关键词： layered images,residual networks,relative depth,Stereo generation
更新于2025-09-23 15:22:29
-
[IEEE 2018 24th International Conference on Pattern Recognition (ICPR) - Beijing (2018.8.20-2018.8.24)] 2018 24th International Conference on Pattern Recognition (ICPR) - An Extensive Study of Cycle-Consistent Generative Networks for Image-to-Image Translation
摘要： Image-to-image translation between different domains has been an important research direction, with the aim of arbitrarily manipulating the source image content to become similar to a target image. Recently, cycle-consistent generative network (CycleGAN) has become a fundamental approach for general-purpose image-to-image translation, while almost no work has examined what factors may influence its performance. To provide more insights, we propose two new models roughly based on CycleGAN, namely LongCycleGAN and NestCycleGAN. First, LongCycleGAN cascades several generators to perform the domain translation in a long cycle. It shows the benefit of stacking more generators on the generation quality. In addition to the long cycle, NestCycleGAN develops new inner cycles to bridge intermediate generators directly, which can help constrain the unsupervised mappings. In the experiments, we conduct qualitative and quantitative comparisons for tasks including photo?label, photo?sketch, and photo colorization. The quantitative and qualitative results demonstrate the effectiveness of our two proposed models.
关键词： CycleGAN,NestCycleGAN,unsupervised learning,image-to-image translation,cycle-consistent generative networks,LongCycleGAN
更新于2025-09-23 15:22:29
-
Adaptive SVD Domain-Based White Gaussian Noise Level Estimation in Images
摘要： Noise level estimation is a challenging area of digital image processing with a variety of applications, including image enhancement, image segmentation, and feature extraction. In this paper, an adaptive estimation of additive white Gaussian noise level based on the singular value decomposition (SVD) of images is proposed. The proposed algorithm aims to improve the performance of noise level estimation in the SVD domain at low noise levels. An initial noise level estimate is used to adjust the parameters of the algorithm in order to increase the accuracy of noise level estimation. The proposed algorithm exhibits the ability to adapt the number of considered singular values and to accordingly adjust the slope of a linear function that describes how the average value of the singular value tail varies with noise levels. Although, for each image, the proposed algorithm performs the noise level estimation twice in two distinct stages, the singular value decompositions are only performed in the first stage of the algorithm. The experimental results demonstrate that the proposed algorithm improves the noise level estimation at low noise levels without a significant increase in computational complexity. At noise level σ = 15, the improvements in the mean square level are about 39% at the expense of slightly higher additional computational time.
关键词： artificial neural networks,singular value decomposition,image analysis,noise level estimation,Digital images,AWGN,least square methods
更新于2025-09-23 15:22:29
-
[IEEE 2018 IEEE 6th Workshop on Advances in Information, Electronic and Electrical Engineering (AIEEE) - Vilnius, Lithuania (2018.11.8-2018.11.10)] 2018 IEEE 6th Workshop on Advances in Information, Electronic and Electrical Engineering (AIEEE) - Deep Neural Network-based Feature Descriptor for Retinal Image Registration
摘要： Feature description is an important step in image registration workflow. Discriminative power of feature descriptors affects feature matching performance and overall results of image registration. Deep Neural Network-based (DNN) feature descriptors are emerging trend in image registration tasks, often performing equally or better than hand-crafted ones. However, there are no learned local feature descriptors, specifically trained for human retinal image registration. In this paper we propose DNN-based feature descriptor that was trained on retinal image patches and compare it to well-known hand-crafted feature descriptors. Training dataset of image patches was compiled from nine online datasets of eye fundus images. Learned feature descriptor was compared to other descriptors using Fundus Image Registration dataset (FIRE), measuring amount of correctly matched ground truth points (Rank-1 metric) after feature description. We compare the performance of various feature descriptors applied for retinal image feature matching.
关键词： artificial neural networks,biomedical imaging,machine learning,image registration,retinal images,feature descriptors
更新于2025-09-23 15:22:29
-
[IEEE 2018 25th IEEE International Conference on Image Processing (ICIP) - Athens, Greece (2018.10.7-2018.10.10)] 2018 25th IEEE International Conference on Image Processing (ICIP) - Perception Preserving Decolorization
摘要： Decolorization is a basic tool to transform a color image into a grayscale image, which is used in digital printing, stylized black-and-white photography, and in many single-channel image processing applications. While recent researches focus on retaining as much as possible meaningful visual features and color contrast. In this paper, we explore how to use deep neural networks for decolorization, and propose an optimization approach aiming at perception preserving. The system uses deep representations to extract content information based on human visual perception, and automatically selects suitable grayscale for decolorization. The evaluation experiments show the effectiveness of the proposed method.
关键词： Color-to-gray conversion,perception preserving,deep neural networks
更新于2025-09-23 15:22:29
-
[IEEE 2018 25th International Conference on Mechatronics and Machine Vision in Practice (M2VIP) - Stuttgart, Germany (2018.11.20-2018.11.22)] 2018 25th International Conference on Mechatronics and Machine Vision in Practice (M2VIP) - Generative models for direct generation of CNC toolpaths
摘要： Today, numerical controls (CNC) are the standard for the control of machine tools and industrial robots in production and enable highly flexible and efficient production, especially for frequently changing production tasks. A numerical control has discrete inputs and outputs. Within the NC channel, however, it is necessary to analytically describe curves for the calculation of the position setpoints and the jerk limitation. The resulting change between discrete and continuous description forms and the considerable restrictions in the parallelisation of the interpolation of continuous curves within the NC channel lead to a performance overhead that limits the performance of the NC channel with regard to the calculation of new position setpoints. This can lead to a drop in production speed and thus to longer production times. To solve this problem, we propose a new approach in this paper. This is based on the use of deep generative models and allows the direct generation of interpolated toolpaths without calculation of continuous curves and subsequent discretization. The generative models are being trained to create curves of certain types such as linear and parabolic curves or splines directly as discrete point sequences. This approach is very well feasible with regard to its parallelization and reduces the computing effort within the NC channel. First results with straight lines and parabolic curves show the feasibility of this new approach for the generation of CNC toolpaths.
关键词： machine learning,computerized numerical control,interpolation,CNC,generative adversarial networks
更新于2025-09-23 15:22:29
-
Gated CNN for visual quality assessment based on color perception
摘要： Visual quality assessment aims to build a computational model which can evaluate the quality of image with respect to human perception. As one of the important aspects to represent images, color provides rich information of images and deeply influences the visual perception. Thus, color harmony, which is defined as 'two or more colors are sensed together as a single, pleasing, collective impression' [1], is one of the important features to determine the visual/aesthetic quality of images. Despite the recent progress of color harmony models in respect to aesthetic quality assessment, most conventional approaches represent the color harmony only considering the distribution of co-occurrence colors but ignoring the spatial relationships between those neighbored colors. To overcome this limitation, we propose to take advantage of the intrinsic structural properties from conditional random field (CRF) to model the color harmony of images. In the CRF framework, we present a novel method that uses gated convolutional neural networks (CNNs) to calculate the probabilities of being high aesthetic quality for small patches and compute the harmony compatibilities between them, which can be considered as the associated and interactive potentials of CRF. Semantic tag of each image is also employed in our work to improve the proposed harmony model's discriminant capability, which shows promising improvements for aesthetic quality assessment compared with existing color harmony models.
关键词： Aesthetic quality,Deep neural networks,Conditional random fields
更新于2025-09-23 15:22:29
-
Performance-Aware Energy Saving for Data Center Networks
摘要： Today’s data center networks (DCNs) tend to have tens to hundreds of thousands of servers to provide massive and sophisticated services. The architectural design of DCNs usually over-provisioned for peaks workloads and fault-tolerance. Statistically, DCNs remain highly under-utilized with typical utilization of around 30%. Network over-provisioning and under-utilization can be exploited for energy-saving. Most research efforts on data center network energy saving focus on how to save maximum energy with little or no consideration to the performance of the residual network. Thus, the DCN performance degraded and the network left vulnerable to sudden traffic surges. In this paper, we have studied energy-saving problem in DCNs while preserving network performance. The problem was formulated as MILP that is solvable by CPLEX to minimize the energy consumed by DCN, meanwhile, safety threshold constraints for links utilization are met. To overcome CPLEX high computational time, a heuristic algorithm to provide practical and efficient solution for the MILP is introduced. The heuristic algorithm uses switches grouping and links consolidation to switch the traffic to a small number of network devices and turn-off unused switches and links. Valiant load-balancing is used to distribute the loads over active links. Simulation experiments using synthetic and real packet traces were conducted to validate the heuristic in terms of energy consumption and network performance. The results show that the heuristic can save up to 45% of the network energy and improves the average imbalance-scores for links and switches by more than 50% with minimal effect on network performance.
关键词： Load Balancing,Energy Saving,Data Center Networks
更新于2025-09-23 15:22:29
-
Ship detection based on squeeze excitation skip-connection path networks for optical remote sensing images
摘要： Ship detection plays a crucial role in remote sensing image processing, which has drawn great attention in recent years. A novel neural network architecture named squeeze excitation skip-connection path networks (SESPNets) is proposed. A bottom-up path is added to feature pyramid network to improve feature extraction capability, and path-level skip-connection structure is firstly proposed to enhance information flow and reduce parameter redundancy. Also, squeeze excitation module is adopted, which can adaptively recalibrate channel-wise feature responses by adding an extra branch after each shortcut path connection block. The multi-scale fused region of interest (ROI) align is then proposed to obtain more accurate and multi-scale proposals. Finally, soft-non-maximum suppression is utilized to overcome the problem of non-maximum suppression (NMS) in ship detection. As demonstrated in the experiments, it can be seen that the SESPNets model has achieved the state-of-the-art performance, which shows the effectiveness of proposed method.
关键词： Skip-connection path networks,Squeeze excitation,Ship detection,Optical remote sensing images,Deep learning
更新于2025-09-23 15:22:29
-
[IEEE 2018 25th IEEE International Conference on Image Processing (ICIP) - Athens, Greece (2018.10.7-2018.10.10)] 2018 25th IEEE International Conference on Image Processing (ICIP) - Integrating Multi-Level Convolutional Features for Correlation Filter Tracking
摘要： Discriminative correlation filters (DCFs) have drawn increasing interest in visual tracking. In particular, a few recent works treat DCFs as a special layer and adding it into a Siamese network for visual tracking. However, they adopt shallow networks to learn target representations, which lack robust semantic information in deeper layers and make these works fail to handle significant appearance changes. In this paper, we design a novel network to fuse multi-level convolutional features, each level of which characterize target from different perspectives. Then we integrate our network with the DCF layer to construct an end-to-end deep architecture for visual tracking. The overall architecture is trained end-to-end offline to adaptively learn target representations, which are not only enabled to encode high-level semantic features and low-level spatial detail features, but also closely related to correlation filters. Experiments show that our proposed tracker achieves superior performance against state-of the-art trackers.
关键词： correlation filters,visual tracking,convolutional neural networks
更新于2025-09-23 15:22:29