PoLin Lai

PoLin Lai
polinlai AT usc DOT edu

PoLin Lai
polinlai AT usc DOT edu

Course works / Projects

(Interesting works done during my M.S./B.S. years)

Spatially Scalable Stereo Image Coding using Wavelet-based Disparity Prediction
(Fall 2003, Term project of Wavelets, USC)
Stereo image pairs are highly correlated because of cross-view redundancy. Instead of transmitting the disparity fields and the residual, this term paper proposed a method called Wavelet-based Disparity Prediction. In this approach, the very low frequency wavelet coefficients of the image pair are used to predict the disparity fields of the remaining high frequency wavelet bands, and then the reconstruction residual is compensated. Because the encoder and decoder follow the same procedure when doing disparity prediction, we don¡¦t have to transmit the entire disparity fields. This scheme brings other benefits such as spatial scalability and suffering less from block effect.

Analysis and Cancellation of Environmental Noise
(Fall 2003, Term project of Immersive Signal Processing, USC)
There are basically two type of noise, those are streamy and those are bursty. Examples of the formal case are noises from air-condition, refrigerator, and computer. This kind of noises is incessant, and usually much more stationary than the bursty ones. In this project I investigated the idea of using adaptive signal processing techniques to track the noise, than instantly play the opposite phase sound to cancel the original noise, in a MATLAB simulated environment. The test subject is the noise comes from my refrigerator. The first step is to analysis the noise signal and determine how the adaptive techniques work on it. In the LMS adaptive structure, the predicted signal is used to cancel the input signal and the output serves as the feedback to the adaptive filter. The next stage is to simulate the noise canceling system which is placed near the listener. The final stage simulate the noise canceling system is placed just near by the noise source.

Survey of Fast Gray-Level Image Thresholding Algorithms
(Spring 2003, Term project of Selected Topics in Digital Image Processing - Multimedia Data Compression, USC)
Rerepresenting a gray-level image with only a few levels can be used to reduce storage memory or to the applications for image segmentation. To achieve this, we have to find thresholds and representing values. The Max Quantizer could a employed, but it will be extremely computation consuming because the images can have any kind of histograms. There are many approaches developed since around 1980. A survey paper comparing different bi-level representations, not multilevel thresholding, can be found while it contains no materials later then 1988. In this project I implemented some fast multilevel thresholding algorithms and compare their performance, in terms of visual properties and more importantly, the ability to preserve information. There is a famous method named Moment Preserved Multilevel Threshold, which will preserve the lower order moments of images after quantization. Another way is through Histogram Equalization, making the result images have the largest entropy. The third method I tried is Peak Detection algorithm to find the thresholds.

Audio Compression Using a Simple Psychoacoustics Model
(Fall 2002, Group project of Multimedia Systems Design, USC)
Exploit the frequency masking effect in human hearing to achieve audio compression. We designed and simulated the analysis/synthesis filter banks, then focus on psychoacoustics and dynamic bits allocation for different subbands, working primarily on frequency analysis and approximating the critical bands of the human auditory model. We then integrated our modules into a complete audio compression system.

Correlation Analysis of Speckle Distribution in Ultrasonic Images
(Fall 2000~Spring 2001, Research group, Ultrasonic Imaging Laboratory, NTU)
My research team's goal was to design a 3D ultrasonic image rendering system. By comparing speckle distribution in ultrasonic images, we calculated the relative position of the pulser/receiver and used those positions to reconstruct 3D images from consecutive 2D ultrasonic images. I selected a block in one image and used correlation analysis to find the location of this block in other images, so that I could estimate the displacement of the pulser/receiver. To speed up the block matching process, I tried different block shapes based on the direction of displacement (perpendicular and/or transverse). This method provided high accuracy and reduced the calculation complexity.