Target detection with morphological shared-weight neural network : different update approaches
Abstract
Neural networks are widely used for image processing. Of these, the convolutional neural network (CNN) is one of the most popular. However, the CNN needs a large amount of training data to improve its accuracy. If training data is limited, a morphological shared-weight neural network (MSNN) can be a better choice. In this thesis, two different update approaches based on an evolutionary algorithm are proposed and compared to each other for target detection based on the MSNN. Another network training, based on back propagation, is used for comparisons in this thesis, which was proposed by Yongwan Won and applied by my colleague and fellow graduate student, Shuxian Shen and Anes Ouadou. Single-layer and multiple-layer MSNNs are both presented with different approaches. For a dataset, the author created part of a dataset for this thesis and used another dataset created by Shen to make comparisons with her network. Results of the MSNN are compared with CNN results to show the performance. Experiments show that for a single-layer MSNN, the performance of an evolutionary algorithm with partial backpropagation is the best. For a multiple layer MSNN, backpropagation performs better, although the MSNN still has a better performance than the CNN.
Degree
M.S.
Thesis Department
Rights
OpenAccess.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License.