YZYunxin Zhao

Yunxin Zhao

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Yunxin Zhao has a total of 39 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • PointGrad: A Unified Spatially Varying Tensor Gradient Operator for Point Cloud Segmentation

    by Rina Bao, Yunxin Zhao, Kannappan Palaniappan

    Accurate representation of geometric shapes and structures plays a pivotal role in facilitating the recognition and segmentation of point clouds, particularly in remote sensing applications involving irregular, sparse and large-scale data. Graph Neural Networks (GNNs), a widely used method for encoding geometric relationships among points through edge operators within a latent space projection. However, prevailing graph operators of point cloud within GNNs have not been explicitly tailored to incorporate directional gradient operations, which are crucial for capturing essential structural information, therefore, failed in challenging cases. To address this, we introduce a novel graph convolution operator, termed the Point Gradient (PointGrad) operator. PointGrad is meticulously crafted to effectively capture directional tensor gradients of individual points, subsequently mapping them into a high-dimensional, multi-scale tensor space. The PointGrad operator adeptly encodes diverse orders of partial derivatives, thereby comprehensively encapsulating multi-scale geometric shape details and curvature characteristics. Integrating the PointGrad operator into graph convolutions over distributed, irregular point sets helps capture pivotal structural particulars within the point cloud across varying spatial and feature scales to empower efficient learning. Furthermore, through the strategic stacking of PointGrad operators, we establish a family of deep networks known as PointGradNet. The comprehensive experimental analysis demonstrates the effectiveness and generalization of the proposed PointGrad operator across a range of benchmark datasets from varying in spatial scale, data sources, and task granularities, including urban scene airborne lidar point cloud segmentation, indoor scene semantic segmentation, and object part segmentation. Experiments on remote sensing benchmarks, including the IEEE DFT4 and COU2024 LiDAR datasets, demonstrate that PointGrad significantly improves segmentation accuracy and robustness in complex geospatial environments, providing a powerful, interpretable tool for scene understanding in earth observation and geoscience analytics. Codes is available at https://github.com/baorina/PointGradNet.

  • Open Access

    Lightweight SDENet Fusing Model-Based and Learned Features for Computational Histopathology

    by Rina Bao, Yunxin Zhao, Akhil Srivastava, Shellaine Frazier, Kannappan Palaniappan

    Model-based deep learning has the potential to significantly reduce the size of deep architectures while matching the competitive performance of much deeper and wider architectures. We demonstrate the advantage of combining model-based handcrafted features with learned features for AI-enabled computational pathology. Digital histopathology with whole slide image analysis using gigapixel-sized images and deep neural networks are being actively investigated for diagnosis and treatment, but require tens to hundreds of millions of learnable parameters (network weights). Additionally, using deep architectures effectively in medical applications, including pathology, has a number of challenges, including limited supervisory manual expert labels, a paucity of training data to cover clinical heterogeneity, many rare disease classes, complex anatomical structures and very large deep network architectures that have difficulty with domain adaptation along with generalization to new patient cohorts. We propose a lightweight squeeze, delineate, and excitation network (SDENet) deep learning architecture for pathology image cell and nuclei segmentation. SDENet is based on a novel hybrid network modular design, composed of a combination of model-based engineered or predefined filters for extracting salient information based on expert medical knowledge, with a learnable stack of convolution filters to capture structural relationships using complementary bottom-up data driven features. The proposed SDENet model-based approach learns rich feature representations of histopathology images, achieving a highly competitive performance on the MoNuSeg dataset for cell and nuclei segmentation with almost 90% fewer learnable parameters, and generalizes better to unseen image datasets, achieving about 20% higher accuracy on TNBC, compared to the widely used UNet architecture.


  • Open Access

    GLSNet++: Global and Local-Stream Feature Fusion for LiDAR Point Cloud Semantic Segmentation Using GNN Demixing Block

    by Rina Bao, Kannappan Palaniappan, Yunxin Zhao, Guna Seetharaman

    Semantic point cloud segmentation is a critical task in 3D computer vision, offering valuable contextual information for navigation, cartography, landmarks, object recognition, and building modeling. We developed GLSNet++, an innovative deep learning architecture for robust context-dependent 3D point cloud segmentation. GLSNet++ uniquely combines dual streams of global and local feature manifolds to capture multiscale contextual and structural information, addressing challenges due to highly varying object sizes in urban scenes. To effectively and efficiently refine mixed class labels from cross-scale global and local streams, GLSNet++ incorporates a novel graph neural network (GNN) based demixing block (GDB) for accurately resolving class membership near voxel boundaries with spatial context-dependent feature fusion. We validate GLSNet++ on the IEEE DFT4 LiDAR dataset, achieving competitive city-scale semantic segmentation that can be extended to more classes, higher-resolution point clouds, and larger geographic regions. GLSNet++ exhibits strong generalization when tested on an independent LiDAR dataset from Columbia, Missouri evaluated using OpenStreetMap.


  • Modeling Speech Structure to Improve T-F Masks for Speech Enhancement and Recognition

    by Suliang Bu, Yunxin Zhao, Tuo Zhao, Shaojun Wang, Mei Han

    Time-frequency (TF) masks are widely used in speech enhancement (SE). However, accurately estimating TF masks from noisy speech remains a challenge to both statistical or neural network (NN) approaches. Statistical model based mask estimation usually depends on a good parameter initialization, while NN-based method relies on setting proper and stable learning targets. To address these issues, we propose to extract TF speech structure from clean speech and partition noisy speech spectrogram into mutually exclusive regions. We investigate modeling clean speech by utterance-specific narrowband complex Gaussian mixture models to derive the regions, and using the region targets to supervise the training of UNet++, a high-performance NN, for predicting regions from noisy speech. For multichannel SE, we consider two scenarios of using speech regions: 1) integrating the regions with TF masks by constraining the mask values or the model parameter updates, and 2) using the predicted regions in place of TF masks. For single-channel SE, we consider using the region targets to improve TF mask targets. Furthermore, we propose to use UNet++ for TF mask estimation. Our experiment results on speech recognition (CHiME-3) and SE (CHiME-3 and LibriSpeech) have demonstrated the effectiveness of our proposed approach of modeling speech region structure to improve TF masks for speech recognition and enhancement.


  • Voice Conversion for Persons with Amyotrophic Lateral Sclerosis

    by Yunxin Zhao, Mili Kuruvilla-Dugdale, Minguang Song

    Amyotrophic lateral sclerosis (ALS) results in progressive paralysis of voluntary muscles throughout the body. As speech deteriorates, individuals rely on pre-programmed messages available on commercial speech generating devices to communicate using one of the generic electronic voices on the device. To replace these generic voices and restore vocal identity, our aim is to develop personalized voices for people with ALS via the approach of voice conversion. The task is challenging because very few people have large quantities of their premorbid healthy speech recorded. Therefore, we have to rely on small quantities of dysarthric speech concomitant with an individual's disease stage. Further, progressive fatigue prohibits acquisition of large speech datasets and individuals display a range of dysarthria severities resulting from breathing, voice, articulation, resonance, and prosody disturbances. As the first step to address these problems, we use healthy source speakers and propose the approach of combining a structured sparse spectral transform with multiple linear regression-based frequency warping prediction for spectral conversion, and interpolating the transformed spectral frames for speech rate modification. Our experimental data included four healthy source speakers from the ARCTIC dataset, and four target ALS speakers with mild to severe dysarthria, forming 16 speaker pairs. Subjective listening evaluations showed that on average, (i) the proposed approach improved speech intelligibility by about 80% over the target speakers' speech, (ii) the converted voice was 3 times more similar to the target speakers' speech than to the source speakers' speech, and (iii) the converted speech quality was close to the MOS scale “good” relative to the source speakers' speech being “excellent.”


  • Structured Sparse Spectral Transforms and Structural Measures for Voice Conversion

    by Yunxin Zhao, Mili Kuruvilla-Dugdale, Minguang Song

    We investigate a structured sparse spectral transform method for voice conversion (VC) to perform frequency warping and spectral shaping simultaneously on high-dimensional (D) STRAIGHT spectra. Learning a large transform matrix for high-D data often results in an overfit matrix with low sparsity, which leads to muffled speech in VC. We address this problem by using the frequency-warping characteristic of a source-target speaker pair to define a region of support (ROS) in a transform matrix, and further optimize it by nonnegative matrix factorization (NMF) to obtain structured sparse transform. We also investigate structural measures of spectral and temporal covariance and variance at different scales for assessing VC speech quality. Our experiments on ARCTIC dataset of 12 speaker pairs show that embedding the ROS in spectral transforms offers flexibility in tradeoffs between spectral distortion and structure preservation, and the structural measures provide quantitatively reasonable results on converted speech. Our subjective listening tests show that the proposed VC method achieves a mean opinion score of “very good” relative to natural speech, and in comparison with three other VC methods, it is the most preferred one in naturalness and in voice similarity to target speakers.


  • Open Access

    Slim Embedding Layers for Recurrent Neural Language Models

    by Zhongliang Li, Raymond Kulhanek, Shaojun Wang, Yunxin Zhao, Shuang Wu

    Recurrent neural language models are the state-of-the-art models for language modeling. When the vocabulary size is large, the space taken to store the model parameters becomes the bottleneck for the use of recurrent neural language models. In this paper, we introduce a simple space compression method that randomly shares the structured parameters at both the input and output embedding layers of the recurrent neural language models to significantly reduce the size of model parameters, but still compactly represent the original input and output embedding layers. The method is easy to implement and tune. Experiments on several data sets show that the new method can get similar perplexity and BLEU score results while only using a very tiny fraction of parameters.