Model Reconstruction Technique Research Articles

Additive manufactured high-performance topology-optimized lattice structure: Compressive behavior and flow heat transfer characteristics

In this study, topological optimization techniques were employed to construct innovative TopS lattice structures, to achieve a lightweight design while simultaneously optimizing their mechanical properties and enhancing flow and heat transfer performance. Through model reconstruction and proportional scaling techniques, lightweight configurations with increased surface areas were achieved, leading to the successful fabrication of Ti-6Al-4V lattice structures via Laser Powder Bed Fusion (LPBF) technology. A comprehensive investigation combining numerical simulations and experimental testing revealed that the TopS-L16 lattice structure outperforms TopS-L1 and TopS-L4 in terms of compressive strength, demonstrating values of 386.00 MPa for compressive strength, 352.00 MPa for upper compressive yield strength, and 6938.00 MPa for elastic modulus. Furthermore, multilayer-filled lattice structures were found to effectively improve streamline distribution and vorticity, reducing boundary layers, which indicates a significant enhancement in fluid flow and heat transfer performance, with TopS-L16 exhibiting the best performance. The novelty of this work lies in its exploration of the interplay between lattice structures and their interactions with compressive deformation, energy absorption capabilities, and flow heat transfer.

Delay-dependent stability analysis for load frequency control systems with time-varying delays

This paper introduces a delay-dependent criterion that addresses both computational accuracy and efficiency in dealing with delay-dependent stability challenges within multi-area load frequency control (LFC) systems. Firstly, the traditional power system LFC model is decomposed into delay-related and delay-independent parts using a model reconstruction technique. Subsequently, an augmented Lyapunov–Krasovskii functional (LKF) is formulated, focusing on the delay-related part of the model. The double integral component within the derivative of the LKF is constrained using the parameter-dependent slack matrices approach, resulting in a less conservative stability criterion. Finally, the proposed result is validated through case studies. Simulation results demonstrate that the proposed stability criterion significantly improves computational accuracy while maintaining better validity compared to existing approaches. Moreover, compared to the traditional LFC model, the reconstructed model offers substantial computational efficiency improvements, albeit with a slight decrease in computational precision.

Reverse Engineering Based on Digital Data Capture In Situ as a Methodology for the Study of Space Labor Risk in Construction Works and Its Applicability in BIM

In the intelligent manufacturing process and digitization, new research spaces have been created, such as reverse engineering and its use for the detection of risks in construction. For this reason, in this research, an analysis methodology is proposed that, through terrestrial laser scanner records in different phases of the work, allows for an analysis and evaluation of the identification of security risks. In addition, by capturing different processes, it can be determined not only interferences from different operators that can generate risk but also establish quality standards in construction, such as determining the thicknesses of the coating of the concrete manufactured “in situ”. Therefore, this research responds to quality analysis in the material execution of the activity and construction of formwork systems and their safety, as well as defining the location of the static and dynamic elements of the model. This value of the morphometric parameter of the building and the scene to be represented characterizes the morphometrics of the work execution activities and finally evaluates the semantic information model of the BIM construction, in which the deformations between the ideal model and what was executed. With the results of this study, it can be concluded that the latest 3D model reconstruction techniques (BIM), together with comparison algorithms, can determine millimeter precision, and with that, it can establish a better method for inspection, evaluation, and control through non-destructive technology.

Dataset and method for deep learning-based reconstruction of 3D CAD models containing machining features for mechanical parts

ABSTRACT Three-dimensional (3D) computer-aided design (CAD) model reconstruction techniques are used for numerous purposes across various industries, including free-viewpoint video reconstruction, robotic mapping, tomographic reconstruction, 3D object recognition, and reverse engineering. With the development of deep learning techniques, researchers are investigating the reconstruction of 3D CAD models using learning-based methods. Therefore, we proposed a method to effectively reconstruct 3D CAD models containing machining features into 3D voxels through a 3D encoder–decoder network. 3D CAD model datasets were built to train the 3D CAD model reconstruction network. For this purpose, large-scale 3D CAD models containing machining features were generated through parametric modeling and then converted into a 3D voxel format to build the training datasets. The encoder–decoder network was then trained using these training datasets. Finally, the performance of the trained network was evaluated through 3D reconstruction experiments on numerous test parts, which demonstrated a high reconstruction performance with an error rate of approximately 1%.

Technical aspects to maximize the hyperaccuracy three-dimensional (HA3D™) computed tomography reconstruction for kidney stones surgery: a pilot study.

The aim of the current prospective pilot study was to describe a hyperaccuracy three-dimensional (HA3D™) model reconstruction technique, specifically developed to maximize the visualization of the renal collecting system's anatomy, and its relationship with the stones, vessels and renal parenchyma, and to compare the HA3D™ virtual models with the intraoperative findings. The image acquisition was performed using a CT scanner (Toshiba, Aquilion Prime) and included the unenhanced, arterial, venous and excretory phases. The DICOM format CT images were processed by MEDICS Srl ( www.medics3d.com , Turin, Italy). In total, study included three patients with renal stone scheduled for non-papillary prone percutaneous nephrolithotomy (PCNL). The median age and BMI were 51 (range 49-54) and 25.5 (range 25.0-32.7), respectively. The median stone size was 1170 mm2 (range 830-1520) and median stone density was 1130 HU (range 600-1340). In all cases, the quality of the CT images acquired with our protocol was adequate to perform the HA3D™ reconstruction. Median operative and puncture time were 39.4 (range 35.2-44.0) and 1.9 (range 1.8-2.1) mins, respectively. The success rate for the first attempt of the percutaneous puncture was 100%, and only one PCNL tract was sufficient to complete the surgery. All three patients were stone-free on the third postoperative day. A dedicated imaging acquisition protocol and a tailored 3D model reconstruction process specifically developed for kidney stones treatment allow obtaining HA3D™ highly relevant models to greatly match intraoperative findings during PCNL with the potential of minimizing bleeding and organ injury complications.

Exotic image formation in strong gravitational lensing by clusters of galaxies – I. Cross-section

ABSTRACT In a recent paper, we have discussed the higher order singularities in gravitational lensing. We have shown that a singularity map, comprising A3- lines and unstable (point) singularities (A4 and D4), is a compact representation of high magnification regions corresponding to a given lens model for all possible source redshifts. It marks all the optimal locations for deep surveys in the lens plane. Here, we present singularity maps for 10 different clusters lenses selected from the Hubble Frontier fields (HFF) and the Reionization Lensing Cluster Survey (RELICS) surveys. We have identified regions in the lens plane with a high magnification for sources up to redshift 10. To determine the dependence of unstable (point) singularities on lens mass model reconstruction techniques, we compared singularity maps corresponding to the different mass models (provided by various groups in the HFF survey) for each cluster lens. We find that the non-parametric (free-form) method of lens mass reconstruction yields the least number of point singularities. In contrast, mass models reconstructed by various groups using a parametric approach have a significantly larger number of point singularities. We also estimate the number of galaxies lying near these unstable (point) singularities, which can be observed with the James Webb Space Telescope (JWST). We find that we expect to get at least one hyperbolic umbilic and one swallowtail image formation for a source at z > 1 for every five clusters with JWST. These numbers are much higher than earlier estimates.

Generative adversarial network as a stochastic subsurface model reconstruction

In geosciences, generative adversarial networks have been successfully applied to generate multiple realizations of rock properties from geological priors described by training images, within probabilistic seismic inversion and history matching methods. Here, the use of generative adversarial networks is proposed not as a model generator but as a model reconstruction technique for subsurface models where we do have access to sparse measurements of the subsurface properties of interest. We use sets of geostatistical realizations as training datasets combined with observed experimental data. These networks are applied to reconstruct nonstationary sedimentary channels and continuous elastic properties, such as P-wave propagation velocity, in the presence and absence of conditioning data. The reconstruction examples shown herein can be considered a post-processing step applied after seismic inversion and performed at those locations where the convergence of the inversion is low, and therefore, the inverted models are associated with high uncertainty. The application examples show the suitability of generative adversarial networks in learning the spatial structure of the data from sets of geostatistical realizations. The generated models reproduce the first- and second-order statistical moments and the spatial covariance matrix of the training dataset.

Load Frequency Stability Analysis of Time-Delayed Multi-Area Power Systems With EV Aggregators Based on Bessel-Legendre Inequality and Model Reconstruction Technique

This paper investigates the stability of time-delayed load frequency control (LFC) systems with electric vehicles (EV) aggregators based on Lyapunov theory and linear matrix inequality (LMI) technology. A novel Lyapunov-Krasovskii functional (LKF) is constructed. Then, based on Bessel-Legendre (B-L) inequality and model reconstruction technique, two stability criteria are derived, respectively. Some simulations are carried out to verify the effectiveness of the proposed methods. And the interaction of time-delays between areas and the effect of gains of EVs on delay margins are discussed.

Delay-Dependent Stability Analysis of Multi-Area Load Frequency Control With Enhanced Accuracy and Computation Efficiency

This paper aims at improving calculation accuracy and reducing calculation burden of delay-dependent stability analysis for large-scale multi-area load frequency control (LFC) schemes in traditional and deregulated environments. The special features of the LFC models have been exploited via model reconstruction technique to equivalently transform the original LFC model into a delay-free part and a delay-related part by a transition matrix. The improvement of the calculation efficiency is obtained by decreasing both the number of decision variables and the maximal order of the linear matrix inequalities. Based on the reconstructed model, a novel augmented Lyapunov functional is constructed mainly based on the delay-related part with much lower order and its derivative is bounded with the Wirtinger inequality to establish a stability condition with less conservatism. Case studies are based on two-area LFC systems to verify the effectiveness of proposed stability criteria based on the time-domain indirect method, with a great improvement of the calculation accuracy that can achieve almost accurate delay margins, such as the frequency domain method. Moreover, comparing with the criterion based on the original system model without reconstruction, the calculation efficiency has been greatly improved with the cost of small deduction of accuracy.

Optimized categorization algorithm of coronary artery calcification score on non-gated chest low-dose CT screening using iterative model reconstruction technique

ObjectivesTo investigate the optimized categorization algorithm of coronary artery calcification score (CACS) for more accurate risk assessment on non-gated chest low-dose CT (LDCT) screening using iterative model reconstruction (IMR) technique. MethodsWe enrolled 102 patients who required coronary artery CTA examination and had coronary artery calcification (CAC) in this study. The CACS on non-gated LDCT and ECG-gated CT images were both measured by Agatston analysis software on Philips workstation. ResultsAccording to the original algorithm (1–100, 100–400 and >400), the CACS measured by non-gated LDCT scan showed a good agreement with ECG-gated CT scan (weighted kappa value of 0.602, P < 0.05). Two ROC curves were drawn to evaluate the accuracy of CACS categorization. The cutoff values were 87.44 and 255.26 respectively. Based on the best diagnostic cutoff value, the CACS measured by the non-gated LDCT scan showed an excellent agreement with ECG-gated CT scan (weighted kappa value of 0.781, P < 0.05). ConclusionsThe CACS on non-gated LDCT may have been underestimated. We therefore developed an optimized categorization algorithm of non-gated CACS in this study.

Application of low-tube current with iterative model reconstruction on Philips Brilliance iCT Elite FHD in the accuracy of spinal QCT using a European spine phantom.

To investigate the repeatability and accuracy of quantitative CT (QCT) measurement of bone mineral density (BMD) by low-mAs using iterative model reconstruction (IMR) technique based on phantom model. European spine phantom (ESP) was selected and measured on the Philips Brilliance iCT Elite FHD machine for 10 times. Data were transmitted to the QCT PRO workstation to measure BMD (mg/cm3) of the ESP (L1, L2, L3). Scanning method: the voltage of X-ray tube is 120 kV, the electric current of X-ray tube output in five respective groups A-E were: 20, 30, 40, 50 and 60 mAs. Reconstruction: all data were reconstructed using filtered back projection (FBP), IR levels of hybrid iterative reconstruction (iDose4, levels 1, 2, 3, 4, 5, 6 were used) and IMR (levels 1, 2, 3 were used). ROIs were placed in the middle of L1, L2 and L3 spine phantom in each group. CT values, noise and contrast-to-noise ratio (CNR) were measured and calculated. One-way analysis of variance (ANOVA) was used to compare BMD values of different mAs and different IMR. Radiation dose [volume CT dose index (CTDIvol) and dose length product (DLP)] was positively correlated with tube current. In L1 with low BMD, different mAs in FBP showed P<0.05, indicating statistically significant BMD in ESP. In other iterative algorithms, different mAs under same iterative algorithms showed P>0.05, indicating no difference in BMD. And P>0.05 was observed among BMD of spine phantom in L1, L2 and L3 under same mAs joined with varied iterative reconstruction. The BMD in L1 varied greatly during FBP reconstruction, and less variation was observed in reconstruction of IMR [1] and IMR [2]. The BMD of L2 changed more during FBP reconstruction, where less was observed in IMR [2]. The BMD of L3 varied greatly during FBP reconstruction, and was less varied in all levels of iDose4 and reconstruction of IMR [2]. In addition, along with continuous mAs incensement, the CNRs in various algorithms continued to increase. Among them, CNR with the FBP algorithm is the lowest, and CNR of the IMR [3] algorithm is the highest. Repeated measurements of BMD with QCT in the ESP multicenter showed that BMD changes in L1-L3 are the least varied at IMR [2] algorithm. It is recommended to scan at 120 kV with 20 mAs combined with IMR [2] algorithm. In this way, the BMD of spine by QCT could be accurately measured, while radiation dosage significantly reduced and imaging quality improved at the same time.

2 Milisievert Low Dose Dedicated Breast CT with Iterative Model Reconstruction Technique in Evaluation of Breast Mass: Feasibility Study in Comparison with MRI

Objective To investigate the feasibility of low dose dedicated computed tomography (CT) with IMR technique compared with magnetic resonance imaging (MRI) in preoperative evaluation of breast cancer. Methods Dedicated CT and breast MRI were performed in 21 patients with diagnosed breast mass for preoperative evaluation. A dedicated protocol combined with chest diagnostic CT and dynamic scans of breast was developed with the use of IMR technique as well as optimization of scan parameters to ensure acceptable radiation dose. Image quality evaluations of CT images were performed using a five-point scale. The number, site and size (maximum diameter) of breast lesion were recorded respectively in CT and MRI images. The enhancement patterns of breast lesion were classified to 3 types (washout, plateau, and persistence) according to time intensity curve derived from CT and time signal curve derived from MRI, respectively. Pearson’s correlation, Bland-Altman analysis, and Cohen’s kappa test were used for statistical assessment. Results The mean effective radiation dose of dedicated breast CT, with the image quality diagnostic acceptable for evaluation of breast lesion and pulmonary structures, was (2.15±0.39) mSv, which was no more than a routine chest diagnostic CT. Results of breast lesion number and lesion site obtained by CT and MRI was consistent, with a total of 24 lesions detected in 21 patients. The mean lesion size was (26.5±12.1) mm in CT and (26.1±12.9) mm in MR, respectively. CT showed an excellent correlation with MRI for lesion size (r=0.99, n=24, p<0.0001). Limits of agreement determined by Bland-Altman analysis for lesion extent was (-3.0mm, 3.8mm). Good agreement was observed between CT and MRI for lesion enhancement patterns (kappa value = 0.936, p=0.000). Conclusion Dedicated breast CT imaging with acceptable radiation dose enabled diagnostic image quality and showed good agreement with breast MR imaging in preoperative evaluation of breast cancer.

CT Pulmonary Angiography at Reduced Radiation Exposure and Contrast Material Volume Using Iterative Model Reconstruction and iDose4 Technique in Comparison to FBP.

PurposeTo assess image quality of CT pulmonary angiography (CTPA) at reduced radiation exposure (RD-CTPA) and contrast medium (CM) volume using two different iterative reconstruction (IR) algorithms (iDose4 and iterative model reconstruction (IMR)) in comparison to filtered back projection (FBP).Materials and Methods52 patients (body weight < 100 kg, mean BMI: 23.9) with suspected pulmonary embolism (PE) underwent RD-CTPA (tube voltage: 80 kV; mean CTDIvol: 1.9 mGy) using 40 ml CM. Data were reconstructed using FBP and two different IR algorithms (iDose4 and IMR). Subjective and objective image quality and conspicuity of PE were assessed in central, segmental, and subsegmental arteries.ResultsNoise reduction of 55% was achieved with iDose4 and of 85% with IMR compared to FBP. Contrast-to-noise ratio significantly increased with iDose4 and IMR compared to FBP (p<0.05). Subjective image quality was rated significantly higher at IMR reconstructions in comparison to iDose4 and FBP. Conspicuity of central and segmental PE significantly improved with the use of IMR. In subsegmental arteries, iDose4 was superior to IMR.ConclusionsCTPA at reduced radiation exposure and contrast medium volume is feasible with the use of IMR, which provides improved image quality and conspicuity of pulmonary embolism in central and segmental arteries.

Knowledge-based iterative model reconstruction technique in computed tomography of lumbar spine lowers radiation dose and improves tissue differentiation for patients with lower back pain

PurposeTo evaluate the image quality and diagnostic confidence of reduced-dose computed tomography (CT) of the lumbar spine (L-spine) reconstructed with knowledge-based iterative model reconstruction (IMR). Materials and methodsProspectively, group A consisted of 55 patients imaged with standard acquisition reconstructed with filtered back-projection. Group B consisted of 58 patients imaged with half tube current, reconstructed with hybrid iterative reconstruction (iDose4) in Group B1 and knowledge-based IMR in Group B2. Signal-to-noise ratio (SNR) of different regions, the contrast-to-noise ratio between the intervetebral disc (IVD) and dural sac (D-D CNR), and subjective image quality of different regions were compared. Higher strength IMR was also compared in spinal stenosis cases. ResultsThe SNR of the psoas muscle and D-D CNR were significantly higher in the IMR group. Except for the facet joint, subjective image quality of other regions including IVD, intervertebral foramen (IVF), dural sac, peridural fat, ligmentum flavum, and overall diagnostic acceptability were best for the IMR group. Diagnostic confidence of narrowing IVF and IVD was good (kappa=0.58–0.85). Higher strength IMR delineated IVD better in spinal stenosis cases. ConclusionLower dose CT of L-spine reconstructed with IMR demonstrates better tissue differentiation than iDose4 and standard dose CT with FBP.

Sub-milliSievert ultralow-dose CT colonography with iterative model reconstruction technique

Purpose. The purpose of this study was to evaluate the technical and diagnostic performance of sub-milliSievert ultralow-dose (ULD) CT colonograpy (CTC) in the detection of colonic and extracolonic lesions.Materials and Methods. CTC with standard dose (SD) and ULD acquisitions of 64 matched patients, half of them with colonic findings, were reconstructed with filtered back projection (FBP), hybrid (HIR) and iterative model reconstruction techniques (IMR). Image noise in six colonic segments, in the left psoas muscle and aorta were measured. Image quality of the left adrenal gland and of the colon in the endoscopic and 2D view was rated on a five point Likert scale by two observers, who also completed the reading of CTC for colonic and extracolonic findings.Results. The mean radiation dose estimate was 4.1 ± 1.4 mSv for SD and 0.86 ± 0.17 mSv for ULD for both positions (p < 0.0001). In ULD-IMR, SD-IMR and SD-HIR, the endoluminal noise was decreased in all colonic segments compared to SD-FBP (p < 0.001). There were 27 small (6–9 mm) and 17 large (≥10 mm) colonic lesions that were classified as sessile polyps (n = 38), flat lesions (n = 3), or as a mass (n = 3). Per patient sensitivity and specificity were 0.82 and 0.93 for ULD-FBP, 0.97 and 0.97 for ULD-HIR, 0.97 and 1.0 for ULD-IMR. Per polyp sensitivity was 0.84 for ULD-FBP, 0.98 for ULD-HIR, 0.98 for ULD-IMR. Significantly less extracolonic findings were detected in ULD-FBP and ULD-HIR, but in the E4 category by C-RADS (potentially important findings), the detection was similar.Conclusion. Both HIR and IMR are suitable for sub-milliSievert ULD CTC without sacrificing diagnostic performance of the study.

High-Quality Three-Dimensional Computed Tomography Angiography of Abdominal Viscera with Small Focal Spot, Low Tube Voltage, and Iterative Model Reconstruction Technique

Purpose: To evaluate the quality of three-dimensional (3D) CT angiography images of the abdominal viscera with small focal spot, low tube voltage, and iterative model reconstruction technique (IMR). Materials and Methods: Seven patients with suspected disease of the pancreatobiliary system had undergone CT with high-quality CTA protocol in the present study. There were 5 men and 2 women, ranging in age from 52 to 80 years (mean: 64 years). Results: Depiction of abdominal small artery, small portal vein was possible in all cases. In two cases that we were able to compare, it was superior to standard CTA in small vascular depiction in CTA made clearly in high quality protocol. Conclusions: Although the use of small focal spot, low tube voltage, and IMR can produce higher-quality images of abdominal vessels than standard CTA, this improvement is not significant at elevated radiation doses.

Ultra-low dose abdominal MDCT: Using a knowledge-based Iterative Model Reconstruction technique for substantial dose reduction in a prospective clinical study

PurposeTo assess lesion detection and image quality parameters of a knowledge-based Iterative Model Reconstruction (IMR) in reduced dose (RD) abdominal CT examinations. Materials and methodsThis IRB-approved prospective study included 82 abdominal CT examinations performed for 41 consecutive patients (mean age, 62±12 years; F:M 28:13) who underwent a RD CT (SSDE, 1.5mGy±0.4 [∼0.9mSv] at 120kV with 17–20mAs/slice) immediately after their standard dose (SD) CT exam (10mGy±3 [∼6mSv] at 120kV with automatic exposure control) on 256 MDCT (iCT, Philips Healthcare). SD data were reconstructed using filtered back projection (FBP). RD data were reconstructed with FBP and IMR. Four radiologists used a five-point scale (1=image quality better than SD CT to 5=image quality unacceptable) to assess both subjective image quality and artifacts. Lesions were first detected on RD FBP images. RD IMR and RD FBP images were then compared side-by-side to SD-FBP images in an independent, randomized and blinded fashion. Friedman's test and intraclass correlation coefficient were used for data analysis. Objective measurements included image noise and attenuation as well as noise spectral density (NSD) curves to assess the noise in frequency domain were obtained. In addition, a low-contrast phantom study was performed. ResultsAll true lesions (ranging from 32 to 55) on SD FBP images were detected on RD IMR images across all patients. RD FBP images were unacceptable for subjective image quality. Subjective ratings showed acceptable image quality for IMR for organ margins, soft-tissue structures, and retroperitoneal lymphadenopathy, compared to RD FBP in patients with a BMI ≤25kg/m2 (median-range, 2–3). Irrespective of patient BMI, subjective ratings for hepatic/renal cysts, stones and colonic diverticula were significantly better with RD IMR images (P<0.01). Objective image noise for RD FBP was 57–66% higher, and for RD IMR was 8–56% lower than that for SD-FBP (P<0.01). NSD showed significantly lower noise in the frequency domain with IMR in all patients compared to FBP. ConclusionIMR considerably improved both objective and subjective image quality parameters of RD abdominal CT images compared to FBP in patients with BMI less than or equal to 25kg/m2.

Overview of Research Methods of Microscopic Pore Structure with Material Properties in Reservoir Rocks

Pore reflects fluid storage capacity of the rock, while the shape and size of the throat, together with the pore throat ratio, control the permeability and fluid reservoir capacity. There are mainly two aspects in microscopic pore structure characteristics evaluation at present: experiment and theoretical modeling. In this paper, several experimental methods and modeling methods that currently exist were compared and analyzed. The analysis shows that in the aspect of experiments, the domestic basic research in the field of three-dimensional pore structure model reconstruction technique is still relatively weak under limited conditions of equipment and technology, while in the aspect of models, the current domestic research is still relatively rare. A conclusion is drawn that the main problems in the study of microscopic pore structure at home are still the degree of improvement in the accuracy of the experiment and the model upgrade.

Digital terrain model reconstruction and preliminary scientific exploration planning of the Chang'E 3

Abstract. At 13:11 (GMT) December 14, 2013 Chang’e 3 (CE-3) successfully landed at 19.51° W, 44.12° N northwestern Mare Imbrium on the Moon, making it China's first planetary mission to land on a celestial body other than Earth. CE-3 explore comprises a lander and a rover. It carries eight scientific instruments onboard, including the descent camera on the lander, and the panoramic camera on the rover. These cameras imaged the topographic features around the landing site. This paper mainly presents the digital terrain model reconstruction techniques for the panoramic camera. Image pairs obtained during the first lunar day are used to reconstructed 3D Digital Terrain Models of 0.02 m resolution near observation points E and S3. The maps have been extensively used to support Yutu operations and strategic planning of the mission. The preliminary scientific exploration planning of the Yutu rover for the second lunar day has been made.

Reliable Accent-Specific Unit Generation With Discriminative Dynamic Gaussian Mixture Selection for Multi-Accent Chinese Speech Recognition

In this paper, we propose a discriminative dynamic Gaussian mixture selection (DGMS) strategy to generate reliable accent-specific units (ASUs) for multi-accent speech recognition. Time-aligned phone recognition is used to generate the ASUs that model accent variations explicitly and accurately. DGMS reconstructs and adjusts a pre-trained set of hidden Markov model (HMM) state densities to build dynamic observation densities for each input speech frame. A discriminative minimum classification error criterion is adopted to optimize the sizes of the HMM state observation densities with a genetic algorithm (GA). To the author's knowledge, the discriminative optimization for DGMS accomplishes discriminative training of discrete variables that is first proposed. We found the proposed framework is able to cover more multi-accent changes, thus reduce some performance loss in pruned beam search, without increasing the model size of the original acoustic model set. Evaluation on three typical Chinese accents, Chuan, Yue and Wu, shows that our approach outperforms traditional acoustic model reconstruction techniques with a syllable error rate reduction of 8.0%, 5.5% and 5.0%, respectively, while maintaining a good performance on standard Putonghua speech.