MU Data Science and Analytics professors Ilker Ersoy and Grant Scott Attend the 47th IEEE Applied Imagery Pattern Recognition Workshop at the Cosmos Club in Washington, DC
Drs. Ersoy and Scott snagged a picture together at the Cosmos Club in Washington, DC. They were both in attendance at the 47th Annual IEEE Applied Imagery Pattern Recognition Workshop for 2018. The theme of the conference this year was ubiquitous imaging and there was a heavy focus on deep learning applied to solving real image analysis challenges. Below is a synopsis of three papers presented at the conference.
Improving Nuclei Classification Performance in H&E Stained Tissue Images using Fully Connected Regression Network (FCRN) and Convolutional Neural Network (CNN)
- Ali Hamad, Ilker Ersoy, and Filiz Bunyak
The paper presents a detection classification approach in deep learning framework to accurately classify nuclei of cancer cells in histopathology images as part of an image analytics pipeline aiming to study and understand tumor microenvironment by quantifying morphologies and the spatial distribution of cancer cell groups.
Another Exciting Note: Nvidia tweets about Dr. Ersoy’s upcoming workshop in Madrid, Spain in Dec 2018 for the IEEE International Conference on Bioinformatics and Biomedicine. https://twitter.com/NVIDIAAIDev/status/1049698032487256064
Exploring the Effects of Class-Specific Augmentation and Class Coalescence on Deep Neural Network Performance Using a Novel Road Feature Datasets
- Tyler W. Nivin, Grant J. Scott, J. Alex Hurt, Raymond L. Chastain, Curt H. Davis
Insufficient training data is a well-known problem in machine learning when using deep convolutional neural network techniques. This work takes a look at class-specific augmentation and class coalescence as methods for improving DCNN performance in datasets that face training data quantity challenges.
Benchmark Meta-Dataset of High-Resolution Remote Sensing Imagery for Training Robust Deep Learning Models in Machine-Assisted Visual Analytics
- Alex Hurt, Grant J. Scott, Derek T. Anderson, and Curt H. Davis
The benchmark datasets released in recent years have many co-occurring object classes that can be combined into a Meta-Dataset (MDS) that is more generalizable in broad area scanning applications. Our research explores the advantages of using the Meta-Dataset in both cross validation and broad area scanning contexts, and explores the cost vs benefit of using the MDS compared to benchmark datasets for machine assisted visual analytics.