Presenter(s)

Zhiyuan Xie

Download

Download Project (1.9 MB)

Description

In recent years, neural networks have become more and more popular because of their outstanding performance in the object classification area. The convolutional neural network (CNN) is a deep learning, feed-forward neural network that has excellent performance in visual imagery analysis area. The idea of the connectivity pattern between neurons of the CNN came from the organization of the animal visual cortex. For human vision, different observational directions of objects can get different views. Human can easily recognize objects in different observational directions, but machines cannot achieve this easily. Therefore, multi-view object classification has been researched for many years. To solve this problem, we design an efficient CNN architecture to perform classification of the multi-view images of objects by appropriately choosing the number of layers, the sequence of layers cascading, and size of the filters. Then, we improve the classification performance by adding image enhancement techniques before CNN as a preprocessing stage. CNN extracts various significant features of the image. It is expected that an enhanced image helps to extract stronger features. The training and testing input images of the CNN are original images or enhanced images. The image enhancement is performed by nonlinear enhancement techniques such as multilevel windowed inverse sigmoid (MWIS) function based technique or a locally tuned sine nonlinearity (LTSN) technique. It is observed that the preprocessing by image enhancement provides improved performance in the cases of the smaller training set. Research work is in progress to modify the CNN architecture to see the impact of recognition performance for multi-view object classification. Advanced non-linear enhancement technologies might also be investigated to see the effectiveness in classification.

Publication Date

4-18-2018

Project Designation

Graduate Research

Primary Advisor

Vijayan K. Asari

Primary Advisor's Department

Electrical and Computer Engineering

Keywords

Stander Symposium project

Convolutional Neural Network Based Multi-view Object Classification

Share

COinS