Skip to Main Content
Frequently Asked Questions
Submit an ETD
Global Search Box
Need Help?
Keyword Search
Participating Institutions
Advanced Search
School Logo
Files
File List
36371.pdf (4.8 MB)
ETD Abstract Container
Abstract Header
Semantic Segmentation Using Deep Learning Neural Architectures
Author Info
Sarpangala, Kishan
Permalink:
http://rave.ohiolink.edu/etdc/view?acc_num=ucin157106185092304
Abstract Details
Year and Degree
2019, MS, University of Cincinnati, Engineering and Applied Science: Computer Science.
Abstract
In many picture handling methods, the capacity to segment items of concern automatically is a helpful pre-processing phase. There are several algorithms segmenting pictures depending on one or more parameters or previous understanding of the necessary category of items. While these algorithms generate visually attractive segmentations, their complexities are making many of them computationally costly. One of the approaches which has received a lot of traction recently is semantic segmentation. It’s a process of associating each pixel of an image with a class label, (such as flower, person, road, sky, ocean, or car). Semantic segmentation of a picture has recently been addressed by deep end-to-end neural networks. One of the main problems among all architectures is to consider the global visual context of the input to improve this segmentation forecasting. State of the art drawings use architectures that try to link different components of the image in order to understand the interactions between the items. To solve this problem, the concept of KeyPoint recognition is introduced within semantic segmentation by researchers. A KeyPoint recognition can be described as a procedure to find the most repeatable object or key points for the specific target object during the training phase. Thus, a general semantic segmentation algorithm gets enhanced to a global level which ensures understanding of a scene at the top level. This approach moves much of the computational burden to the training phase, without sacrificing recognition efficiency. Therefore, the resulting algorithm is robust, precise and fast enough for frame rate efficiency. General global semantic segmentation algorithms generate super pixels based solely on color. However, limits of objects do not necessarily overlap with contours of color, allowing segmentation in different situations to fail. In such instances, information beyond color channels is needed to achieve significant and accurate segmentations. A trend in software for picture and video acquisition is not only recording color but also a range of extra data, such as infrared values and depth. This research discusses a completely new method in which pictures are segmented into a required number of super pixels. These super pixels are generated using global semantic segmentation over a whole image, to ensure an extremely precise comprehension of the environment a new concept of integrating various channels beyond color channels to use the extra data accessible to enhance the outcomes of segmentation is used. The accomplished pixel-wise prediction is helpful for certain scenarios like action recognition, video captioning, or visual issues answering, scene understanding is also addressed. Together with multi-task loss, this segmentation approach task should help improve the understanding of a scene globally. This new segmentation approach accomplishes an improvement of more than 4 percent in border precision over regular segmentation which uses only color-based segmentation. We discuss several deep learning models using variations of this segmentation approach and compare their results. One new model, using a traditional FCN model with new parameter choices, is compared with three models from the literature. The new model runs faster and uses fewer resources and gives results of comparable quality to previously published results.
Committee
Carla Purdy, Ph.D. (Committee Chair)
Yizong Cheng, Ph.D. (Committee Member)
Anca Ralescu, Ph.D. (Committee Member)
Pages
87 p.
Subject Headings
Artificial Intelligence
Keywords
Semantic Segmentation
;
Convolutional Neural Network
;
Computer Vision
;
Deep Learning Neural Network
;
Artificial Intelligence
;
Fully Convolutional Network
Recommended Citations
Refworks
EndNote
RIS
Mendeley
Citations
Sarpangala, K. (2019).
Semantic Segmentation Using Deep Learning Neural Architectures
[Master's thesis, University of Cincinnati]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ucin157106185092304
APA Style (7th edition)
Sarpangala, Kishan.
Semantic Segmentation Using Deep Learning Neural Architectures.
2019. University of Cincinnati, Master's thesis.
OhioLINK Electronic Theses and Dissertations Center
, http://rave.ohiolink.edu/etdc/view?acc_num=ucin157106185092304.
MLA Style (8th edition)
Sarpangala, Kishan. "Semantic Segmentation Using Deep Learning Neural Architectures." Master's thesis, University of Cincinnati, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=ucin157106185092304
Chicago Manual of Style (17th edition)
Abstract Footer
Document number:
ucin157106185092304
Download Count:
503
Copyright Info
© 2019, all rights reserved.
This open access ETD is published by University of Cincinnati and OhioLINK.