Zhankun Luo

Ph.D. Student

Elmore Family School of Electrical and Computer Engineering, Purdue University

Biography

I joined the Elmore Family School of Electrical and Computer Engineering at Purdue University West Lafayette as a Ph.D. student in Fall 2021. Prior to that, I obtained my MS degree in the Department of Electrical and Computer Engineering at PNW in 2021. Before that I received my BS degree in Telecommunication Engineering from Beijing Institute of Technology in 2019. My research interests include machine learning, computer vision and image processing.

Download my résumé.

Interests

Machine Learning
Statistical Learning
Deep Learning
Computer Vision
Image Processing

Education

Ph.D. student in Electrical and Computer Engineering (Present)
Purdue University West Lafayette
Ms in Electrical and Computer Engineering, 2021
Purdue University Northwest
BSc in Telecommunication Engineering, 2019
Beijing Institute of Technology

Publications

Characterizing Evolution in Expectation-Maximization Estimates for Overspecified Mixed Linear Regression

In this paper, we thoroughly investigated the EM’s behavior in overspecified two-component Mixed Linear Regression (2MLR) models. We rigorously characterized the EM estimates for both regression parameters and mixing weights by providing the approximate dynamic equations for the evolution of EM estimates and establishing the convergence guarantees for the final accuracy, time complexity, and sample complexity at population and finite-sample levels, respectively. Additionally, our novel analysis sharpens bounds for statistical error, time complexity, and sample complexity with fixed sufficiently balanced mixing weights. Furthermore, we discussed the differences between results of 2MLR and 2GMM, and extended our analysis from the overspecified setting to the finite low SNR regime.

Zhankun Luo, Abolfazl Hashemi

Characterizing Evolution in Expectation-Maximization Estimates for Overspecified Mixed Linear Regression

Unveiling the Cycloid Trajectory of EM Iterations in Mixed Linear Regression

In this paper, we derived closed-form expressions for the EM updates in the 2MLR problem. Notably, in the noiseless setting we first showed and then analyzed the cycloid trajectory of EM updates. Additionally, we demonstrated the quadratic convergence rate for regression parameters, which is independent of mixing weights. We emphasized that errors in mixing weights primarily arise from the angle formed between true and estimated regression parameters. Finally, we conducted a detailed analysis of the statistical errors in the estimation of regression parameters and mixing weights.

Zhankun Luo, Abolfazl Hashemi

Unveiling the Cycloid Trajectory of EM Iterations in Mixed Linear Regression

High-Resolution UAV Image Generation for Sorghum Panicle Detection

In this paper, we presented the use of GAN generated synthetic images to augment the training data for panicle detection and counting. We examined two image-to-image translation GANs and showed that their use can improve the performance of panicle detection and counting. We did not use the temporal information available in our real Sorghum UAV dataset during training due to the limitation of the network structures. Future work includes developing multi-temporal methods that can generate synthetic plant images in a temporally consistent style. This will also us to estimate phenotypic traits as the plant grows. We will also examine our approach for estimating traits of other plant such as maize tassels.

Enyu Cai, Zhankun Luo, Sriram Baireddy, Jiaqi Guo, Changye Yang, Edward J. Delp

High-Resolution UAV Image Generation for Sorghum Panicle Detection

Structured Light Vision Systems Using a Robust Laser Stripe Segmentation Method

In the thesis, we propose a structured light vision system equipped with multi-cameras and multi-laser emitters for object height measurement or 3D reconstruction. The proposed method offers a better accuracy performance over a single camera system. The structured light method may fail the interference of reflection and scattering of light. We use U-Net to extract the laser region, obtain the laser stripe center after erosion and dilation, and finally reconstruct the point cloud corresponding to the laser stripe. Our experiments demonstrate that our structured light system with the U-Net can perform effectively and robustly in a complex environment.

Zhankun Luo

Structured Light Vision Systems Using a Robust Laser Stripe Segmentation Method

Smart Ladle: AI-Based Tool for Optimizing Caster Temperature

The paper describes the work of the authors in making the Smart Ladle, an ongoing project to develop a machine-learning tool that uses industry process data to predict future behavior in the steel refining and casting processes. Using information gathered by SDI Butler Division, the Smart Ladle builds connections between the different inputs (such as a ladle’s history) and the temperature of the steel in the continuous caster.

Nicholas J. Walla, Zhankun Luo, Bin Chen, Yury Krotov, Chenn Q. Zhou

Smart Ladle: AI-Based Tool for Optimizing Caster Temperature

Multi-level Random Sample Consensus Method for Improving Structured Light Vision Systems

The paper proposes a structured light vision system equipped with multi-cameras and multi-laser emitters for object height measurement …

Zhankun Luo, Yaan Zhang, Lizhe Tan

Multi-level Random Sample Consensus Method for Improving Structured Light Vision Systems

Computer Vision Techniques for Improving Structured Light Vision Systems

Compared the proposed MLRANSAC method with time-division and color division methods, the experiment results validated the effectiveness of the MLRANSAC algorithm.

Yaan Zhang, Zhankun Luo, Jintao Hou, Lizhe Tan, Xinnian Guo

Projects

Automated Fetal Brain Segmentation Using Deep Convolutional Neural Network

Fetal MRI is widely used in normal/abnormal development diagnosis for the gestational age (GA) over 20 weeks. Meanwhile, 3D volume registration reconstruction depends on the performance of fetal brain segmentation. We implemented a customized U-Net for automated fetal brain segmentation. The experiment results were evaluated by commonly used metrics in image segmentation.

Comparison of Capsule Networks and Other Networks for Object Segmentation Tasks

A new architecture was introduced by Hinton, named Capsule Networks. A capsule is a group of neurons whose activity vectorsrepresent the parameters of an entity. Hinton uses the length of the activity vector to represent the probability of an entity’s existence, and uses its direction to represent the parameters. The first-level active capsule predicts the higher-level capsule through the parameters of the transformation matrix pair. The objective of this study is to compare Capsule networks with conventional networks, i.e. U-Net and DenseNet for object segmentation Tasks. Using PyTorch library, a python program is implemented to achievethese goals. These models ware tested on Drive dataset.

Parking Occupancy Detection Using Computer Vision

Developed the program using Convoluional Neural Networks for parking space classification and counting occupied parking spaces, whose accuracy reached 90%, precision was 96%. We incorporated the mutual information method for image registration and the affine transformation to eliminate the impact caused by camera shake, the robustness of detection was enhanced.

Re-implementation of Hierarchy Based Classification and Embedding Method

Re-implemented the hierarchy based classification and embedding method for encoding of categories to decrease average hierarchical distance at top 1 by 3%, and that at top 5 by 10% on our VIPER-FoodNet dataset with 82 food categories, 15 thousand images.

Experience

Research Assistant

Machine Intelligence & Networked Data Science (MINDS) Group, Purdue University West Lafayette

August 2025 – Present West Lafayette, IN

Design efficient algorithms with mathematical guarantees to render practical deployment of learning-based systems possible under a variety of considerations such as limited resources, robustness, and adversarial behaviors. Recent applications include Federated Learning and and Cyber-Physical Systems, see Mission of the project for more information.

Teaching Assitant

ECE department, Purdue University West Lafayette

May 2022 – August 2025 West Lafayette, IN

TA for ECE69500 Optimization for Deep Learning (Fall 2022)

Hold office hours and help sessions, grade homeworks and exams
This course introduces students to the theoretical principles behind stochastic, gradient-based algorithms for DL as well as practical considerations such as adaptivity, generalization, distributed learning, and non-convex loss surfaces typically present in modern DL problems
The concatenated slides are available or download and public use under CC BY-NC-SA license

TA for ECE20002 Electrical Engineering Fundamentals II (Fall 2023, 2024; Spring 2023,2024,2025; Summer 2023,2024,2025)

Hold office hours, grade quizzes and coordinate TA group activities
The course addresses mathematical and computational foundations of circuit analysis (differential equations, Laplace Transform techniques) with a focus on application to linear circuits having variable behavior as a function of frequency, with emphasis on filtering

TA for ECE20001 Electrical Engineering Fundamentals I (Summer 2022)

Hold office hours, grade exams and compose Collection of Practice Problems of homeworks
The course covers the basic concepts of current and voltage, devices, theorems, and applications of direct-current (DC), 1st order, alternating-current (AC) and basic electronic components including diodes and transistors

Research Assistant

Video and Image Processing Laboratory (VIPER), Purdue University West Lafayette

August 2021 – May 2022 West Lafayette, IN

Worked on the project Image Based Plant Phenotyping: The PhenoSorg Project

Generated 1 thousand synthetic high-resolution UAV RGB images with panicle labels by using image-to-image translation GANs with a ground truth dataset of 400 real UAV RGB images
Improved mean average precision with Intersection over Union from 0.5 to 0.95 (mAP[.5, .95]) for panicle detection task from 72% to 79%, and reduced Mean Absolute Percent Error (MAPE) for panicle counting task from 11.6% to 7.2%
Created labels for panicles in PhenoRover RGB images to test our approach on PhenoRover data

Worked on the project Technology Assisted Dietary Assessment (TADA)

Investigated reliable and effective methods for Fine-Grained Visual Classification (FGVC)
Re-implemented a hierarchy-based embedding method for encoding of categories to decrease average hierarchical distance at top 1 by 3%, and that at top 5 by 10% on our VIPER-FoodNet dataset with 82 food categories, 15 thousand images
Corrected the incorrect labels and bounding boxes of our VIPER-FoodNet dataset

Research Assistant

Center for Innovation through Visualization and Simulation (CIVS), Purdue University Northwest

February 2020 – May 2021 Hammond, IN

Worked on the project Smart Ladle: Al-Based Tool for Optimizing Casting Temperature

Developed a machine learning application using DNN, lightGBM to provide steel casting temperature predictions
Reduced Root Mean Square Error (RMSE) of predicted casting temperature to 3 degrees Fahrenheit
Collaborated application with SQL database and GUI using Unity (C#) to display predictions and parameters
Tested and deployed this tool at Steel Dynamics Inc (SDI) Butler Division, awarded AIST 2022 Hunt-Kelly Outstanding Paper Award – third place (AIME) and AIST 2021 Digitalization Applications Technology Best Paper Award