Search

In the learning of existing radial basis function neural networks-based methods, it is difficult to propagate errors back. This leads to an inconsistency between the learning and recognition task. This article proposes a geodesic basis function neural network with subclass extension learning (GBFNN-ScE). The geodesic basis function (GBF), which is defined here for the first time, uses the geodetic distance in the manifold as a measure to obtain the response of the sample with respect to the local center. To learn network parameters by back-propagating errors for the purpose of correct classification, a specific GBF based on a pruned gamma encoding cosine function is constructed. This function has a concise and explicit expression on the hyperspherical manifold, which is conducive to the realization of error back propagation. In the preprocessing layer, a sample unitization method with no loss of information, nonnegative unit hyperspherical crown (NUHC) mapping, is proposed. The sample can be mapped to the support set of the GBF. To alleviate the problem that one-hot encoding is not effective enough in the differential expression of data labels within a class, a subclass extension (ScE) learning strategy is proposed. The ScE learning strategy can help the learned network be more robust. For the working of GBFNN-ScE, the original sample is projected onto the support set of GBF through the NUHC mapping. Then the mapped samples are sent to the nonlinear computation units composed of GBFs in the hidden layer. Finally, the response obtained in the hidden layer is weighted by the learned weight to obtain the network output. This article theoretically proves that the separability of the data with ScE learning is stronger. The experimental results show that the proposed GBFNN-ScE has a better performance in recognition tasks than existing methods. The ablation experiments show that the ideas of the GBFNN-ScE contribute to the algorithm performance.

Harmonic mode locking (HML) was proved to be an effectively technique for generating high repetition rate ultrafast pulses. However, none of the various research methods currently used could improve the defects of HML's low signal to noise ratio (SNR), mostly less than 40 dB, limiting its further application. Using the self-made single walled carbon nanotubes (SWCNTs) as saturable absorber (SA) in an erbium-doped fiber laser to obtain HML pulses with an ultra-high SNR (>62 dB) was reported here. A series of harmonics from fundamental frequency to sixth orders could be observed by simply increasing the pump power, and the all-polarization-maintaining (PM) fiber path ensured time domain frequency stability. Moreover, by adjusting the polarization direction of the directional SA, a stable traditional soliton pulse was obtained. In this mode, the repetition frequency no longer split with the increase of pump power, and the SNR was as high as 83 dB. Such an environmentally stable, all-PM fiber, versatile operations laser could not only expand the industrial application prospects of high repetition frequency HML pulses, but also be an excellent seed source for ultra-high-power lasers research.

A passively mode-locked thulium-doped fiber laser using a tungsten ditelluride saturable absorber (${{\rm WTe}_2}\mbox{-}{\rm SA}$WTe_2-SA) is demonstrated. High-power mode-locked pulses with an average output power of 108.1 mW were achieved by incorporating the ${{\rm WTe}_2}\mbox{-}{\rm SA}$WTe_2-SA into a thulium-doped fiber oscillator. To the best of our knowledge, this is the highest average power obtained from a ${{\rm WTe}_2}\mbox{-}{\rm SA}$WTe_2-SA-based fiber laser. We further amplified the output power to 5.60 W with an all-fiber thulium-doped double-cladding fiber amplifier. Our result indicates that ${{\rm WTe}_2}\mbox{-}{\rm SA}$WTe_2-SA could be an excellent candidate for a high-power fiber laser system.

Remote sensing (RS) multi-spectral images are usually suffered from cloud and fog cover, which canlead to analysis troubles and application limitations. An ovel patch-based dark channel prior dehaze method is proposed for solving this problem. An Atmospheric light (AL) curved surface hypothesis, instead of globally invariable plane, is applied to describe AL distribution, and apatch-based approach is given to estimate curved surface. By using AL curved surface estimation, a new recovering model for RS multi-spectral images is given to obtain dehaze-free images. Comparative experiments are conducted, those results illustrate that the proposed method can produces visually impressive restored images, and the proposed method is superior to other relative methods interms of image quality evaluations.

Abstract: A terahertz time-domain spectroscopy (THz-TDS) imaging system can obtain high-dimensional signals with substance fingerprint information. By introducing geometric algebra, a novel signal analysis approach to THz-TDS signals is developed based on an optical physical mechanism. Using this approach, signals are represented with vectors in the high-dimensional real vector space. Geometric distribution properties and algebraic relationships of THz-TDS signals are deduced. It is proved that every complex refractive index of substances relates to a unique 2-blade, the vectors corresponding to the samples of the same substance are collinear and belong to the intrinsic 2-blade of the substance. When decomposed through the conformal split with respect to a 2-blade, THz-TDS signals of high dimensionality can be related to vectors in a 2-dimensional subspace. Based on the conformal split properties we deduced, two criteria for substance identification on the basis of THz-TDS signals are proposed. Accordingly, a novel substance identification method via the conformal split is presented. In the method, the 2-blade related to each “known” substance is calculated with two vectors corresponding to THz-TDS signals measured from samples of the substance but with different thicknesses. Using the conformal split with respect to those 2-blades, an identified vector corresponding to a THz-TDS signal is linearly related to the vector in a 2-dimensional subspace. The substance of a sample can be identified using criteria on the projected vectors in the subspaces. This method can contribute to accurate classification and identification. Finally, two experiments are presented that show the feasibility and accuracy of this method.

A stable noise-like (NL) mode-locked Tm-doped fiber laser (TDFL) relying on a nonlinear optical loop mirror (NOLM) was experimentally presented. Different from the previous NL mode-locked TDFL with NOLM, the entire polarization-maintaining (PM) fiber construction was utilized in our laser cavity, which makes the oscillator have a better resistance to environmental perturbations. The robust TDFL can deliver stable bound-state NL pulses with a pulse envelope tunable from ∼14.1 ns to ∼23.6 ns and maximum pulse energy of ∼40.3 nJ at a repetition rate of ∼980.6 kHz. Meanwhile, the all-PM fiber laser shows good power stability (less than ∼0.7%) and repeatability.

In landmark-based image registration, estimating the landmark correspondence plays an important role. In this letter, a novel landmark correspondence estimation technique using mean shift algorithm is proposed. Image corner points are detected as landmarks and mean shift iterations are adopted to find the most probable corresponding point positions in two images. Mutual information between intensity of two local regions is computed to eliminate mis-matching points. Multi-level estimation (MLE) technique is proposed to improve the stability of corresponding estimation. Experiments show that the precision in location of correspondence landmarks is exact. The proposed technique is shown to be feasible and rapid in the experiments of various mono-modal medical images.

Multisensor image fusion could improve system performances such as detection, tracking, and identification greatly. In this paper, a long distance target detection approach is presented based on multisensor image features fusion. This method extracts two different features from visual and infrared (IR) image sequences respectively to detect regions of motion information content. Temporal change feature is extracted from the visual image sequence using temporal decomposition based on wavelet, which reflects the dynamical content variation at a pixel at any time. And correlation features between local regions are extracted from IR image sequence to distinguish regions with potential moving targets. All these features are merged into a multi-dimensional space and the support vector machine is trained to select regions that have the potential target at each pixel location. The method is robust and feasible to detect long distance targets in clutter background scene.

We report on mode-locked thulium-doped fiber lasers with high-energy nanosecond pulses, relying on the transmission in a semiconductor saturable absorber (SESA) and a carbon nanotube (CNTs-PVA) film separately. A section of an SMF–MMF–SMF structure multimode interferometer with a transmission peak wavelength of ∼2003 nm was used as a wavelength selector to fix the laser wavelength. When the SESA acted as a saturable absorber (SA), the mode-locked fiber laser had a maximum output power of ∼461 mW with a pulse energy of ∼0.14 μJ and a pulse duration of ∼9.14 ns. In a CNT-film-based mode-locked fiber laser, stable mode-locked pulses with the maximum output power of ∼46 mW, pulse energy of ∼26.8 nJ and pulse duration of ∼9.3 ns were obtained. To the best of our knowledge, our experiments demonstrated the first 2 μm region ‘real’ SA-based dissipative soliton resonance with the highest mode-locked pulse energy from a ‘real’ SA-based all-fiberized resonator.

Mutual information (MI) based image registration has been found to be quite effective in many medical image applications. However, standard MI hampers the convergence of registration transformation parameters since it contains local maxima. In this paper, a novel registration method is proposed. At first, MI based on edge width matching is computed to avoid great change of joint probability distribution and get less local maxima. Particle swarm optimization (PSO), which combines local search methods with global ones balancing exploration and exploitation, is done to search the optimal registration parameter. PSO has less computational complexity as its complex behavior follows only a few simple rules. It could avoid local maxima and reach global optimal results. This method is applicable to a variety of multimodal images, and suitable to different interpolation methods. Theoretical analysis and experiments show that this method is effective and accurate to register multimodal medical images.

In the linear discriminative analysis, especially in the high dimension case, it is insufficient to project the data onto a one-dimensional subspace for the two-category classification problem. Therefor a weak component approach (WCA) was proposed to project patterns to a low dimensional subspace with rich number of classification features. The role of the weak component in pattern classification was discussed. And the abundance of discriminative information contained in weak components was explored. Firstly, a definition of the weak component was given. Secondly, an improved regularization method was proposed. The regularization is a biased estimate of the variance in the corresponding dimension of the training data and the population data. Then a construction method of the feature subspace based on weak component was given, which extracts the eigenvector of the scatter matrixes according to their discriminative information. Finally, the proposed approach was validated in the experiments by comparing it with LDA. A better classification accuracy of the presented method was achieved. As WCA extracts the dims on which the data distributes closer, it is applicable to the high-dimensional data which distributes elliptically.