Senior Design Projects

ECS193 A/B Winter & Spring 2019

Image Classification using SVM and Neural Networks

Name Carl Stahmer
Affiliation Data Science Initiative

Project's details

Project title Image Classification using SVM and Neural Networks
Background The Data Science Initiative ( DSI ) has developed software that uses a variety of machine learning techniques to provide content based image recognition ( CBIR ) of an image library and also builds a similarity network of an entire image library. The current system relies on modified SURF feature point extraction to identify salient features from each image, and then uses a combination of clustering algorithms, feature collection distance calculations, and a neural network to estimate image similarity for search and retrieval and for building an image similarity network. The current system is currently being deployed in variety of contexts to help users find duplicate and/or near copies of the same image. ( See, for example, ).
Description The current software is extremely accurate when directed to find copies of "the same" image ( for example, multiple images of the same person holding the same flower, even if there are variations in angle, etc. ) This project involves the application of a variety of machine learning techniques, ranging from SVM to convolutional neural networks, to enhance the platform’s functionality so that it can also find "similar" images. For example, a user starting with an image of a tree should be able to find not only pictures of the same tree but of all trees, regardless of type, size, color, etc. Students working on the project will first implement a trained SVM classifier to provide baseline functionality, after which they will design, train, test and implement a neural network based solution to the problem. Students working on the project will work directly with the Associate Director of the DSI, DSI Graduate Student Researchers, and the DSI staff data scientist.
Deliverable C++ and/or R code that outputs 1 ) a data table containing data definitions for image classes found in the library of analysis; and 2 ) a data table that assigns probability assignments for each image in the library to each class in the class definition data table from item1; and 3 ) some system for visualizing the results of the classification for testing purposes only.
Skill set desirable C++ and/or R coding skills. Coursework in computer vision, SVM, and/or neural networks desirable but not required.
Phone number 15307521138
Client time availability 30-60 min weekly or more
IP requirement Open source project
Attachment N/A
Selected N/A
Stuff N/A