Faculty Recruiting Support CICS

Data-Driven Approach for Image classification

24 Jan
Thursday, 01/24/2019 9:00am to 11:00am
PhD Seminar


Image classification has been a core topic in the computer vision community, it's recent success with convolutional neural network (CNN) algorithm has led to various real world applications such as large scale management of photos/videos on cloud/social-media, image based search for online retailers, self-driving cars, building robots and healthcare. Image classification can be broadly categorized into binary, multi-class and multi-label classification problems. Binary classification involves assigning one of the two class labels to an instance. In multi-class classification problem, an instance should be categorized into one of more than two classes. Multi-label classification is a generalized version of the multi-class classification problem where each image is assigned multiple labels as opposed to a single label. 

In this work, we first present various methods that take advantage of deep learning representations (fully connected layer of pre-trained CNN on the ImageNet dataset) and yields better performance on multi-label classification when compared to methods that use over a dozen conventional visual features. Following the success of deep learning representations, we intend to build a generic end-to-end deep learning framework to address all three problem categories of image classification. However, there are still no well established guidelines to train an efficient deep neural network. Hence, we present some initial efforts in building a novel computational framework called Deep Decision Network (DDN) which is completely data-driven. DDN is a tree-like structured built stage-wise. During the learning phase, starting from the root network node, DDN automatically builds a network that splits the data into disjoint clusters of classes which would be handled by the subsequent expert networks. This results in a tree-like structured network driven by the data. The proposed approach provides an insight into the data by identifying the group of classes that are hard to classify and require more attention when compared to others. This feature is crucial for people trying to solve the problem with little or no domain knowledge and for applications that deal with big data.  Initially, we evaluate DDN on a binary classification problem and later extend it to more challenging  multi-class and multi-label classification problems. The extension of DDN to multi-class and multi-label involves some changes but they still operate under the same underlying principle. In all the three cases, the proposed approach is tested for it's recognition performance and scalability on publicly available datasets providing comparison to other methods.

Advisor: Raghavan Manmatha