Scalable Object Detection

Project PI: Garret Vo, NGA

This project will use the robust regression algorithm to perform object detection among millions or billions of images. The robust regression method detects objects in a single image with heavy noise and non-uniform background.  The algorithm was originally developed to detect objects in electron microscope images with heavy noise and a non-uniform background. To detect objects in the input image, the method performs image binarization to turn an image into black and white using a two-stage approach. First, the method estimates the image background and removes it from the original image. Then global thresholding is applied to obtain the binary image.

This project will apply the method at a large scale to detect objects in millions or billions of images with the Blue Waters system. Applications include detecting biological cells in large number of electron microscope images or isolating animals in remote sensing images with low resolution and heavy noise.  With this effort, NGA can potentially detect objects in commercial images at large-scale to inform policymakers.