RESUMEN
Recent advances in deep learning-based methods have led to significant progress in the hyperspectral super-resolution (SR). However, the scarcity and the high dimension of data have hindered further development since deep models require sufficient data to learn stable patterns. Moreover, the huge domain differences between hyperspectral image (HSI) datasets pose a significant challenge in generalizability. To address these problems, we present a general hyperspectral SR framework via meta-transfer learning (MTL). We randomly sample various spectral ranges for SR tasks during MTL, allowing the model to accumulate diverse task experiences. Additionally, we implement a task schedule to gradually expand the number of bands, bridging the significant domain differences between datasets. By leveraging multiple datasets, we are able to achieve better performance and greater generalizability, making it applicable under various circumstances. Meanwhile, as a general framework, our scheme can be applied to existing methods to obtain performance improvements. In addition, we design an advanced network architecture based on the multifusion features to further improve the performance. Experiments demonstrate that our method not only achieves superior performance in both qualitative and quantitative terms but also can adapt robustly to a new and difficult sample, where few epochs can yield quite considerable results.
RESUMEN
Semi-supervised deep learning algorithm is an effective means of medical image segmentation. Among these methods, multi-task learning with consistency regularization has achieved outstanding results. However, most of the existing methods usually simply embed the Signed Distance Map (SDM) task into the network, which underestimates the potential ability of SDM in edge awareness and leads to excessive dependence between tasks. In this work, we propose a novel triple-task mutual consistency (TTMC) framework to enhance shape and edge awareness capabilities, and overcome the task dependence problem underestimated in previous work. Specifically, we innovatively construct the Signed Attention Map (SAM), a novel fusion image with attention mechanism, and use it as an auxiliary task for segmentation to enhance the edge awareness ability. Then we implement a triple-task deep network, which jointly predicts the voxel-wise classification map, the Signed Distance Map and the Signed Attention Map. In our proposed framework, an optimized differentiable transformation layer associates SDM with voxel-wise classification map and SAM prediction, while task-level consistency regularization utilizes unlabeled data in an unsupervised manner. Evaluated on the public Left Atrium dataset and NIH Pancreas dataset, our proposed framework achieves significant performance gains by effectively utilizing unlabeled data, outperforming recent state-of-the-art semi-supervised segmentation methods. Code is available at https://github.com/Saocent/TTMC.
Asunto(s)
Imagenología Tridimensional , Humanos , Imagenología Tridimensional/métodos , Aprendizaje Profundo , AlgoritmosRESUMEN
INTRODUCTION: This study proposed an automatic diagnosis method based on deep learning for adenoid hypertrophy detection on cone-beam computed tomography. METHODS: The hierarchical masks self-attention U-net (HMSAU-Net) for segmentation of the upper airway and the 3-dimensional (3D)-ResNet for diagnosing adenoid hypertrophy were constructed on the basis of 87 cone-beam computed tomography samples. A self-attention encoder module was added to the SAU-Net to optimize upper airway segmentation precision. The hierarchical masks were introduced to ensure that the HMSAU-Net captured sufficient local semantic information. RESULTS: We used Dice to evaluate the performance of HMSAU-Net and used diagnostic method indicators to test the performance of 3D-ResNet. The average Dice value of our proposed model was 0.960, which was superior to the 3DU-Net and SAU-Net models. In the diagnostic models, 3D-ResNet10 had an excellent ability to diagnose adenoid hypertrophy automatically with a mean accuracy of 0.912, a mean sensitivity of 0.976, a mean specificity of 0.867, a mean positive predictive value of 0.837, a mean negative predictive value of 0.981, and a F1 score of 0.901. CONCLUSIONS: The value of this diagnostic system lies in that it provides a new method for the rapid and accurate early clinical diagnosis of adenoid hypertrophy in children, allows us to look at the upper airway obstruction in three-dimensional space and relieves the work pressure of imaging doctors.
Asunto(s)
Tonsila Faríngea , Aprendizaje Profundo , Niño , Humanos , Tonsila Faríngea/diagnóstico por imagen , Tomografía Computarizada de Haz Cónico/métodos , Nariz , Hipertrofia/diagnóstico por imagen , Procesamiento de Imagen Asistido por Computador/métodosRESUMEN
Spectral unmixing (SU), which refers to extracting basic features (i.e., endmembers) at the subpixel level and calculating the corresponding proportion (i.e., abundances), has become a major preprocessing technique for the hyperspectral image analysis. Since the unmixing procedure can be explained as finding a set of low-dimensional representations that reconstruct the data with their corresponding bases, autoencoders (AEs) have been effectively designed to address unsupervised SU problems. However, their ability to exploit the prior properties remains limited, and noise and initialization conditions will greatly affect the performance of unmixing. In this article, we propose a novel technique network for unsupervised unmixing which is based on the adversarial AE, termed as adversarial autoencoder network (AAENet), to address the above problems. First, the image to be unmixed is assumed to be partitioned into homogeneous regions. Then, considering the spatial correlation between local pixels, the pixels in the same region are assumed to share the same statistical properties (means and covariances) and abundance can be modeled to follow an appropriate prior distribution. Then the adversarial training procedure is adapted to transfer the spatial information into the network. By matching the aggregated posterior of the abundance with a certain prior distribution to correct the weight of unmixing, the proposed AAENet exhibits a more accurate and interpretable unmixing performance. Compared with the traditional AE method, our approach can greatly enhance the performance and robustness of the model by using the adversarial procedure and adding the abundance prior to the framework. The experiments on both the simulated and real hyperspectral data demonstrate that the proposed algorithm can outperform the other state-of-the-art methods.
RESUMEN
For robust feature matching, a popular and particularly effective method is to recover smooth functions from the data to differentiate the true correspondences (inliers) from false correspondences (outliers). In the existing works, the well-established regularization theory has been extensively studied and exploited to estimate the functions while controlling its complexity to enforce the smoothness constraint, which has shown prominent advantages in this task. However, despite the theoretical optimality properties, the high complexities in both time and space are induced and become the main obstacle of their application. In this article, we propose a novel method for multivariate regression and point matching, which exploits the sparsity structure of smooth functions. Specifically, we use compact Fourier bases for constructing the function, which inherently allows a coarse-to-fine representation. The smoothness constraint can be explicitly imposed by adopting a few low-frequency bases for representation, resulting in reduced computational complexities of the induced multivariate regression algorithm. To cope with potential gross outliers, we formulate the learning problem into a Bayesian framework with latent variables indicating the inliers and outliers and a mixture model accounting for the distribution of data, where a fast expectation-maximization solution can be derived. Extensive experiments are conducted on synthetic data and real-world image matching, and point set registration datasets, which demonstrates the advantages of our method against the current state-of-the-art methods in terms of both scalability and robustness.
RESUMEN
The representation-based algorithm has raised a great interest in hyperspectral image (HSI) classification. l1-minimization-based sparse representation (SR) attempts to select a few atoms and cannot fully reflect within-class information, while l2-minimization-based collaborative representation (CR) tries to use all of the atoms leading to mixed-class information. Considering the above problems, we propose the pairwise elastic net representation-based classification (PENRC) method. PENRC combines the l1-norm and l2-norm penalties and introduces a new penalty term, including a similar matrix between dictionary atoms. This similar matrix enables the automatic grouping selection of highly correlated data to estimate more robust weight coefficients for better classification performance. To reduce computation cost and further improve classification accuracy, we use part of the atoms as a local adaptive dictionary rather than the entire training atoms. Furthermore, we consider the neighbor information of each pixel and propose a joint pairwise elastic net representation-based classification (J-PENRC) method. Experimental results on chosen hyperspectral data sets confirm that our proposed algorithms outperform the other state-of-the-art algorithms.
RESUMEN
Recently, convolutional sparse representation (CSR) has improved the preservation of details of source images in the fusion results. This is mainly because the CSR has a global representation character that can improve spatial consistency in image representation. However, during image fusion processing, since the CSR expresses infrared and visible images separately, it ignores connections and differences between them. Further, CSR-based image fusion is not able to retain both strong intensity and clear details in the fusion results. In this paper, a novel fusion approach based on joint CSR is proposed. Specifically, we establish a joint form based on the CSR. The joint form is able to guarantee spatial consistency during image representation while obtaining distinct features, such as visible scene details and infrared target intensity. Experimental results illustrate that our fusion framework outperforms traditional fusion frameworks of sparse representation.
RESUMEN
In this paper, we proposed a new end-to-end model, termed as dual-discriminator conditional generative adversarial network (DDcGAN), for fusing infrared and visible images of different resolutions. Our method establishes an adversarial game between a generator and two discriminators. The generator aims to generate a real-like fused image based on a specifically designed content loss to fool the two discriminators, while the two discriminators aim to distinguish the structure differences between the fused image and two source images, respectively, in addition to the content loss. Consequently, the fused image is forced to simultaneously keep the thermal radiation in the infrared image and the texture details in the visible image. Moreover, to fuse source images of different resolutions, e.g., a low-resolution infrared image and a high-resolution visible image, our DDcGAN constrains the downsampled fused image to have similar property with the infrared image. This can avoid causing thermal radiation information blurring or visible texture detail loss, which typically happens in traditional methods. In addition, we also apply our DDcGAN to fusing multi-modality medical images of different resolutions, e.g., a low-resolution positron emission tomography image and a high-resolution magnetic resonance image. The qualitative and quantitative experiments on publicly available datasets demonstrate the superiority of our DDcGAN over the state-of-the-art, in terms of both visual effect and quantitative metrics.
RESUMEN
This paper proposes a new algorithm for infrared and visible image fusion based on gradient transfer that achieves fusion by preserving the intensity of the infrared image and then transferring gradients in the corresponding visible one to the result. The gradient transfer suffers from the problems of low dynamic range and detail loss because it ignores the intensity from the visible image. The new algorithm solves these problems by providing additive intensity from the visible image to balance the intensity between the infrared image and the visible one. It formulates the fusion task as an l1-l1-TV minimization problem and then employs variable splitting and augmented Lagrangian to convert the unconstrained problem to a constrained one that can be solved in the framework of alternating the multiplier direction method. Experiments demonstrate that the new algorithm achieves better fusion results with a high computation efficiency in both qualitative and quantitative tests than gradient transfer and most state-of-the-art methods.
RESUMEN
Denoising is an important preprocessing step to further analyze the hyperspectral image (HSI), and many denoising methods have been used for the denoising of the HSI data cube. However, the traditional denoising methods are sensitive to outliers and non-Gaussian noise. In this paper, by utilizing the underlying low-rank tensor property of the clean HSI data and the sparsity property of the outliers and non-Gaussian noise, we propose a new model based on the robust low-rank tensor recovery, which can preserve the global structure of HSI and simultaneously remove the outliers and different types of noise: Gaussian noise, impulse noise, dead lines, and so on. The proposed model can be solved by the inexact augmented Lagrangian method, and experiments on simulated and real hyperspectral images demonstrate that the proposed method is efficient for HSI denoising.
RESUMEN
The state-of-the-art ultra-spectral sensor technology brings new hope for high precision applications due to its high spectral resolution. However, it also comes with new challenges, such as the high data dimension and noise problems. In this paper, we propose a real-time method for infrared ultra-spectral signature classification via spatial pyramid matching (SPM), which includes two aspects. First, we introduce an infrared ultra-spectral signature similarity measure method via SPM, which is the foundation of the matching-based classification method. Second, we propose the classification method with reference spectral libraries, which utilizes the SPM-based similarity for the real-time infrared ultra-spectral signature classification with robustness performance. Specifically, instead of matching with each spectrum in the spectral library, our method is based on feature matching, which includes a feature library-generating phase. We calculate the SPM-based similarity between the feature of the spectrum and that of each spectrum of the reference feature library, then take the class index of the corresponding spectrum having the maximum similarity as the final result. Experimental comparisons on two publicly-available datasets demonstrate that the proposed method effectively improves the real-time classification performance and robustness to noise.
RESUMEN
Deposition of nanostructured metals on substrates is important for the fundamental study and practical application, such as in optics and catalysis. In this paper, we report the deposition of gold (Au) nanoplates and porous platinum (Pt) structures on substrates through solvent-free chemical reductions of chloroauric acid (HAuCl(4)) and chloroplatinic acid (H(2)PtCl(6)) with ethylene glycol (EG) vapor at temperatures below 200 °C. The process includes two steps. The first step is the formation of a thin layer of a metal precursor on substrates by coating solution of a metal precursor. The thin metal precursor layer is subsequently dried by annealing. The second step is the chemical reduction of the metal precursor with EG vapor at 160 or 180 °C in the absence of solvent. Both the Au and Pt nanostructures deposited by this method have good adhesion to substrates, but they have different morphologies. The Au nanostructures appear as separate two-dimensional islands on the substrates, and up to 70% of them can be triangular nanoplates with the (111) crystal plane as the basal plane. In contrast, the reduction of H(2)PtCl(6) gives rise to a 3-dimensional porous Pt structure on substrates. The different morphologies of nanostructured Au and Pt are tentatively related to the different surface energies of Au and Pt.
Asunto(s)
Cloruros/química , Glicol de Etileno/química , Compuestos de Oro/química , Oro/química , Nanopartículas del Metal/química , Compuestos de Platino/química , Platino (Metal)/química , Oxidación-Reducción , Tamaño de la Partícula , Porosidad , Propiedades de Superficie , VolatilizaciónRESUMEN
Electronically and ionically conductive gels were fabricated by mixing and mechanically grinding neutral tetrathiafulvalene (TTF) and tetracyanoquinodimethane (TCNQ) in ionic liquids (ILs) like 3-ethyl-1-methylimidazolium dicyanoamide (EMIDCA), 1-ethyl-3-methylimidazolium thiocyanate (EMISCN), 1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide (EMITf(2)N), trihexyltetradecylphosphonium bis(trifluoromethylsulfonyl)imide (P(14,6,6,6)Tf(2)N), and methyl-trioctylammonium bis(trifluoromethylsulfonyl)imide (MOATf(2)N). Charge-transfer TTF-TCNQ crystallites were generated during the mechanical grinding as indicated by the UV-visibile-near-infrared (UV-vis-NIR) absorption spectroscopy, Fourier transform infrared (FTIR) spectroscopy, and X-ray diffraction. The charge-transfer TTF-TCNQ crystallites have a needle-like shape. They form solid networks to gelate the ILs. The gel behavior is confirmed by the dynamic mechanical measurements. It depends on both the anions and cations of the ILs. In addition, when 1-methyl-3-butylimidazolium tetrafluoroborate (BMIBF(4)) and 1-methyl-3-propylimidazolium iodide (PMII) were used, the TTF-TCNQ/IL mixtures did not behave as gels. The TTF-TCNQ/IL gels are both electronically and ionically conductive, because the solid phase formed by the charge-transfer TTF-TCNQ crystallites is electronically conductive, while the ILs are ionically conductive. The gel formation is related to needle-like charge-transfer TTF-TCNQ cyrstallites and the π-π and Coulombic interactions between TTF-TCNQ and ILs.
Asunto(s)
Geles/química , Compuestos Heterocíclicos/química , Líquidos Iónicos/química , Nitrilos/química , Conductividad Eléctrica , Iones/química , Estructura Molecular , Tamaño de la Partícula , Propiedades de SuperficieRESUMEN
High-performance dye-sensitized solar cells (DSCs) with binder-free films of carbon nanotubes (CNTs), including single-walled CNTs (SWCNTs) and multi-walled CNTs (MWCNTs), as the counter electrode are reported. The CNT films were fabricated by coating gels, which were prepared by dispersing CNTs in low-molecular-weight poly(ethylene glycol) (PEG) through mechanical grinding and subsequent ultrasonication, on fluorine tin oxide (FTO) glass. PEG was removed from the CNT films through heating. These binder-free CNT films were rough and exhibited good adhesion to substrates. They were used as the counter electrode of DSCs. The DSCs with SWCNT or MWCNT counter electrodes exhibited a light-to-electricity conversion efficiency comparable with that with the conventional platinum (Pt) counter electrode, when the devices were tested immediately after device fabrication. The DSCs with an SWCNT counter electrode exhibited good stability in photovoltaic performance. The efficiency did not decrease after four weeks. On the other hand, DSCs with the MWCNT or Pt counter electrode exhibited a remarkable decrease in the photovoltaic efficiency after four weeks. The high photovoltaic performance of these DSCs is related to the excellent electrochemical catalysis of CNTs on the redox of the iodide/triiodide pair, as revealed by the cyclic voltammetry and ac impedance spectroscopy.