580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can graph-based methods enhance structural pattern recognition through improved matching and learning?
This research area focuses on the use of graph representations for structural pattern recognition, leveraging graph matching and embedding to capture complex relational information in patterns beyond feature vectors. Graph matching algorithms, including exact, inexact, and graph edit distance methods, play a key role in classification and learning tasks by enabling the comparison of structural objects. Recent advances address computational challenges and integration of graph methods with statistical learning, aiming to balance representational richness with tractability.
Key finding: This paper systematically categorizes and reviews over 180 graph-based pattern recognition methods developed in the past decade. It highlights two main approaches: "graphs from beginning to end" using graph matching... Read more
Key finding: The group integrates expertise in hardware implementation and structural pattern recognition algorithms, focusing on developing active vision systems and learning techniques for industrial applications. Their work... Read more
Key finding: The study employs feature detection and matching methods rooted in structural recognition to align historical images of buildings taken under varying conditions. By focusing on invariant features and robust homologous point... Read more
2. What are the efficacies and methodological innovations of template matching and statistical learning models for handwritten and gesture pattern recognition?
This theme examines the effectiveness of template matching techniques and statistical models such as Hidden Markov Models (HMMs) in recognizing time-varying and spatially variant patterns including handwritten characters and gesture sequences. Research centers on creating adaptable, user-specific, and computationally efficient models that address the challenges of variability inherent in handwriting and human motion, supported by advances in feature extraction, model structuring, and sequence segmentation.
Key finding: This paper proposes a novel template matching approach that maintains a low number of training samples yet achieves high recognition accuracy by dynamically adapting templates to users' handwriting styles. The method uses... Read more
Key finding: This work reviews the application of HMMs to gesture and sign language recognition, showing their robustness for modeling spatiotemporal variability and simultaneous segmentation and recognition of key and transition... Read more
Key finding: Comparative evaluation of classical machine learning algorithms (SVM, KNN, RFC) against deep learning-based convolutional neural networks (CNN) for handwritten digit classification demonstrates remarkable accuracy... Read more
Key finding: The paper formulates a morphological template matching algorithm for gray-scale images using elementary operators such as dilations, erosions, and anti-dilations. This provides a lattice-theory-based framework for inexact... Read more
3. How can nonparametric and kernel-based statistical approaches improve pattern recognition accuracy and modality fusion?
This research area explores advancements in nonparametric methods such as compact pattern synthesis and kernel-based fusion techniques to enhance the accuracy, robustness, and adaptability of pattern recognition systems. It spans density estimation improvements for high-dimensional data, statistical learning integration across multiple sensing modalities, and techniques that yield interpretable, efficient recognition in complex scenarios.
Key finding: The paper introduces overlap-based and partition-based pattern synthesis techniques to compactly and generally represent large, high-dimensional pattern datasets. These synthesized patterns reduce computational overhead and... Read more
Key finding: This study presents a kernel-based framework for multimodal pattern recognition by defining combined kernels as Cartesian products of single modality kernels. The approach enables sensor-level fusion, encompassing relevance... Read more
Key finding: The research identifies that a carefully selected subset of Local Binary Patterns (LBP) provides a more robust and discriminative texture representation than the full 256-bin histogram, particularly under geometric... Read more
Key finding: The paper proposes the color cooccurrence histogram which incorporates spatial relationships by counting pairs of color pixels at various distances, thus integrating geometric information into histogram-based object... Read more
Key finding: Two accelerated template matching algorithms based on φ-correlation coefficients and binary logical circuits demonstrate increased speed and accuracy compared to conventional grayscale and color image processing. Experimental... Read more