Topic
Upsampling
About: Upsampling is a research topic. Over the lifetime, 2426 publications have been published within this topic receiving 57613 citations.
Papers published on a yearly basis
Papers
More filters
••
TL;DR: In this paper, a simple and efficient method for refining maps or correspondences by iterative upsampling in the spectral domain that can be implemented in a few lines of code is presented.
Abstract: We present a simple and efficient method for refining maps or correspondences by iterative upsampling in the spectral domain that can be implemented in a few lines of code. Our main observation is that high quality maps can be obtained even if the input correspondences are noisy or are encoded by a small number of coefficients in a spectral basis. We show how this approach can be used in conjunction with existing initialization techniques across a range of application scenarios, including symmetry detection, map refinement across complete shapes, non-rigid partial shape matching and function transfer. In each application we demonstrate an improvement with respect to both the quality of the results and the computational speed compared to the best competing methods, with up to two orders of magnitude speed-up in some applications. We also demonstrate that our method is both robust to noisy input and is scalable with respect to shape complexity. Finally, we present a theoretical justification for our approach, shedding light on structural properties of functional maps.
98 citations
••
TL;DR: A novel modified generative adversarial network (GAN) is proposed to train a DL-based feature extractor without supervision, and replaces the original Jensen–Shannon divergence with the Wasserstein distance, aiming to mitigate the unstability and difficulty of the training of GAN frameworks.
Abstract: Feature extraction (FE) is a crucial research area in hyperspectral image (HSI) processing. Recently, due to the powerful ability of deep learning (DL) to extract spatial and spectral features, DL-based FE methods have shown great potentials for HSI processing. However, most of the DL-based FE methods are supervised, and the training of them suffers from the absence of labeled samples in HSIs severely. The training issue of supervised DL-based FE methods limits their application on HSI processing. To address this issue, in this paper, a novel modified generative adversarial network (GAN) is proposed to train a DL-based feature extractor without supervision. The designed GAN consists of two components, which are a generator and a discriminator. The generator can focus on the learning of real probability distributions of data sets and the discriminator can extract spatial–spectral features with superior invariance effectively. In order to learn upsampling and downsampling strategies adaptively during FE, the proposed generator and discriminator are designed based on a fully deconvolutional subnetwork and a fully convolutional subnetwork, respectively. Moreover, a novel min–max cost function is designed for training the proposed GAN in an end-to-end fashion without supervision, by utilizing the zero-sum game relationship between the generator and discriminator. Besides, the proposed modified GAN replaces the original Jensen–Shannon divergence with the Wasserstein distance, aiming to mitigate the unstability and difficulty of the training of GAN frameworks. Experimental results on three real data sets validate the effectiveness of the proposed method.
97 citations
•
04 Oct 1999TL;DR: In this paper, the IPP architecture is integrated onto a Digital Signal Processor (DSP) as a coprocessor to assist in the computation of sum of absolute differences, symmetrical row/column Finite Impulse Response (FIR) filtering with a downsampling (or upsampling) option, and generic algebraic functions.
Abstract: The proposed architecture is integrated onto a Digital Signal Processor (DSP) as a coprocessor to assist in the computation of sum of absolute differences, symmetrical row/column Finite Impulse Response (FIR) filtering with a downsampling (or upsampling) option, row/column Discrete Cosine Transform (DCT)/Inverse Discrete Cosine Transform (IDCT), and generic algebraic functions The architecture is called IPP, which stands for image processing peripheral, and consists of 8 multiply-accumulate hardware units connected in parallel and routed and multiplexed together The architecture can be dependent upon a Direct Memory Access (DMA) controller to retrieve and write back data from/to DSP memory without intervention from the DSP core The DSP can set up the DMA transfer and IPP/DMA synchronization in advance, then go on its own processing task Alternatively, the DSP can perform the data transfers and synchronization itself by synchronizing with the IPP architecture on these transfers This architecture implements 2-D filtering, symmetrical filtering, short filters, sum of absolute differences, and mosaic decoding more efficiently than the previously disclosed architectures of the prior art
96 citations
••
TL;DR: In this paper, the authors developed an iterative image-reconstruction algorithm for application to low-intensity computed tomography projection data, which is based on constrained, total-variation (TV) minimization.
Abstract: Purpose: The authors developed an iterative image-reconstruction algorithm for application to low-intensity computed tomography projection data, which is based on constrained, total-variation (TV) minimization. The algorithm design focuses on recovering structure on length scales comparable to a detector bin width. Methods: Recovering the resolution on the scale of a detector bin requires that pixel size be much smaller than the bin width. The resulting image array contains many more pixels than data, and this undersampling is overcome with a combination of Fourier upsampling of each projection and the use of constrained, TV minimization, as suggested by compressive sensing. The presented pseudocode for solving constrained, TV minimization is designed to yield an accurate solution to this optimization problem within 100 iterations. Results: The proposed image-reconstruction algorithm is applied to a low-intensity scan of a rabbit with a thin wire to test the resolution. The proposed algorithm is compared to filtered backprojection (FBP). Conclusions: The algorithm may have some advantage over FBP in that the resulting noise level is lowered at equivalent contrast levels of the wire.
95 citations
••
29 Mar 2010TL;DR: This paper extends and refines the patch-based image model of Freeman et al. and interprets the image as a tiling of distinct textures, each of which is matched to an example in a database of relevant textures, resulting in enhanced appearance textured regions.
Abstract: Image upsampling is a common yet challenging task, since it is severely underconstrained. While considerable progress was made in preserving the sharpness of salient edges, current methods fail to reproduce the fine detail typically present in the textured regions bounded by these edges, resulting in unrealistic appearance. In this paper we address this fundamental shortcoming by integrating higher-level image analysis and custom low-level image synthesis. Our approach extends and refines the patch-based image model of Freeman et al. [10] and interprets the image as a tiling of distinct textures, each of which is matched to an example in a database of relevant textures. The matching is not done at the patch level, but rather collectively, over entire segments. Following this model fitting stage, which requires some user guidance, a higher-resolution image is synthesized using a hybrid approach that incorporates principles from example-based texture synthesis. We show that for images that comply with our model, our method is able to reintroduce consistent fine-scale detail, resulting in enhanced appearance textured regions.
94 citations