Student Project Ideas
Creating Super-Resolution ImagesImages are one of the fastest growing forms of data, with millions being taken every day. Unfortunately, many of these are not of the resolution required by their owner, particularly those taken using older devices. Recovering a high-resolution image from a single low-resolution image is a classical problem in computer vision, and many approaches have been proposed. Recently, attention has turned to the use of deep machine learning methods such as Dong et al's SRCNN (https://arxiv.org/pdf/1501.00092.pdf). The goal of this project is to develop a software tool capable of creating super-resolution colour images, ideally using deep machine learning, and to evaluate that tool on a range of images.
What's that Chord?
There have been many attempts to analyse images of musical notation, e.g. to produce MIDI files that can be played or integrated into recordings. As a guitarist with limited music reading ability I have a very specific problem: I can read single notes at a reasonable speed but I find it hard to recognise chords when they are presented in standard notation. The goal of this project is to produce a smartphone app that can capture an image of a single chord expressed in standard notation (and maybe the key signature to provide context) and recognise which chord it is. Having recognised the chord, the app might play it, show alternative fingering diagrams, etc.
Much computer graphics work is aimed at producing photo-realistic images of artificial objects. Painterly rendering takes a different approach and aims to produce images which appear to be have been painted, by a particular artist, or in a particular artistic style. This is done by taking an initial, real image (or set of images) and manipulating it to produce a new image that appears painted. This project involves selecting an artist or style that interests you, and developing a software tool capable of rendering input images in that style.
Active Visual Inspection
The goal of image analysis and computer vision is to extract information about viewed objects from images. In many cases those images are captured by equipment outside the control of the image analysis code, e.g. they might be gathered from surveillance cameras or photographic libraries. Allowing the vision system to control the position, etc of the camera gives it more flexibility: some tasks can be greatly eased, or simply made possible, by acquiring a different images. This leads to the idea of active vision systems, in which there is a feedback loop between the image analysis software and some form of robot system which sets the camera location and captures images. The aim of this project is to create an active vision system that drives a 2-axis, arduino-controlled camera placement system to gather information on objects placed inside the system.
The Leaf Segmentation Challenge/Leaf Counting ChallengeThe image processing and computer vision literature contains many segmentation algorithms, each aiming to break an input image into regions corresponding to individual objects, or parts of objects. Climate change and a growing population have recently put huge demands on the world’s food supply, creating a pressing need for methods that can measure various properties of plants - the idea being that if your goal is to create new plant varieties that produce more food you need to be able to assess them in various ways. Measurements of leaves are very important, and so a number of leaf segmentation methods have been proposed. To help this, the Leaf Segmentation Challenge (http://www.plant-phenotyping.org/CVPPP2014-challenge) was recently set up. This provides a large set of plant images, with manual segmentations giving the idea result. The goal of this project is to attempt the leaf segmentation challenge - to develop an image segmentation tool that can identify leaves in the challenge images. An alternative project, based on the same data, would be to count the leaves, which might be done without a complete segmentation.
Root Phenomics Hub
Phenotyping, or phenomics, is the science of determining the structure and function of living things. At Nottingham, a significant body of work has been done on the development of methods for phenotyping plant roots, usually through the analysis of various types of image, including laser microscopes, standard colour images, and X-ray tomography. A number of root measurement tools have been developed and are in wide use in the biological community. Deployment of these tools has now generated a sizeable and rich body of data, which is growing quickly. The goal of this project is to design and develop a web portal which will allow users to store, browse, access and compare the various datasets available. The long-term here goal is to create a resource for root biologists along the lines of the Ionomics Hub (http://www.ionomicshub.org/home/PiiMS) or MorphoSource (http://morphosource.org/index.php).
3D Shape from Image Pairs
Binocular stereo - the recovery of 3D models from pairs of images - is a longstanding problem in computer vision. The Middlebury Stereo site (http://vision.middlebury.edu/stereo/) provides several sets of image pairs, along with ground truth data showing what the true 3D structure of the scene is. This is a valuable resource for those attempting to develop new stereo methods. The goal of this project is to atempt the Middlebury challenge - to develop software capable of extracting some 3D measurements from a stereo image pair, and compare it to the ground truth.
Input image Ground truth depth map