A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks
碩士 === 國立臺灣科技大學 === 資訊工程系 === 106 === Ship detection is a major issue in video surveillance, analyzing the location of the ship through computer vision. The traditional detection method is judged by supervised learning, but the real world situation is more complicated and cannot be detected by ships...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | en_US |
Published: |
2018
|
Online Access: | http://ndltd.ncl.edu.tw/handle/sb2q49 |
id |
ndltd-TW-106NTUS5392037 |
---|---|
record_format |
oai_dc |
spelling |
ndltd-TW-106NTUS53920372019-11-28T05:22:05Z http://ndltd.ncl.edu.tw/handle/sb2q49 A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks 一個應用卷積式類神經網路於多角度航拍影像的即時船隻偵測方法 Fei-Ping Hu 胡斐評 碩士 國立臺灣科技大學 資訊工程系 106 Ship detection is a major issue in video surveillance, analyzing the location of the ship through computer vision. The traditional detection method is judged by supervised learning, but the real world situation is more complicated and cannot be detected by ships of every angle or size. In view of the shortcomings of the above methods, we have designed a method of ship detection that can be used to detect images at different angles (including elevation, depression, and azimuth), different colors, and different shapes. This paper proposes a method that can be applied to automatically detect ships in various sea areas. First we will train a ship model that can detect all angles and all colors. Through a machine learning method, convolutional neural network, a ship model capable of detecting multiple angles is obtained. In the ship detection process, we use the gray-scale, Gaussian matrix application, we achieve the effect of binarization; cut the foreground into individual objects, and input the regions of interest to classify the ships by the pre-trained model. Then we output our detection video at last. In the experimental part, we first analyze the accuracy of the model, and the final accuracy is about 98%. In addition, we analyze the different environments, such as the sea with a lot of white smoke, the sea with strong illumination, and the sea where the ship will rotate, the field of many colors of the ship, and the field of the land. The method proposed in this thesis can correctly detect the ship and distinguish between land and ships. The accuracy in the white smoke area is 96.6%, 96% in the sea with strong illumination, 89.8% in the field where the ship rotates, the color of the ship's field is 96.1%, and the land field is 91.1%. The overall execution time is very fast, and the processing time of an aerial image is about 0.05 to 0.06 seconds to achieve real-time multi-angle ship detection. Chin-Shyurng Fahn 范欽雄 2018 學位論文 ; thesis 66 en_US |
collection |
NDLTD |
language |
en_US |
format |
Others
|
sources |
NDLTD |
description |
碩士 === 國立臺灣科技大學 === 資訊工程系 === 106 === Ship detection is a major issue in video surveillance, analyzing the location of the ship through computer vision. The traditional detection method is judged by supervised learning, but the real world situation is more complicated and cannot be detected by ships of every angle or size. In view of the shortcomings of the above methods, we have designed a method of ship detection that can be used to detect images at different angles (including elevation, depression, and azimuth), different colors, and different shapes.
This paper proposes a method that can be applied to automatically detect ships in various sea areas. First we will train a ship model that can detect all angles and all colors. Through a machine learning method, convolutional neural network, a ship model capable of detecting multiple angles is obtained. In the ship detection process, we use the gray-scale, Gaussian matrix application, we achieve the effect of binarization; cut the foreground into individual objects, and input the regions of interest to classify the ships by the pre-trained model. Then we output our detection video at last.
In the experimental part, we first analyze the accuracy of the model, and the final accuracy is about 98%. In addition, we analyze the different environments, such as the sea with a lot of white smoke, the sea with strong illumination, and the sea where the ship will rotate, the field of many colors of the ship, and the field of the land. The method proposed in this thesis can correctly detect the ship and distinguish between land and ships. The accuracy in the white smoke area is 96.6%, 96% in the sea with strong illumination, 89.8% in the field where the ship rotates, the color of the ship's field is 96.1%, and the land field is 91.1%. The overall execution time is very fast, and the processing time of an aerial image is about 0.05 to 0.06 seconds to achieve real-time multi-angle ship detection.
|
author2 |
Chin-Shyurng Fahn |
author_facet |
Chin-Shyurng Fahn Fei-Ping Hu 胡斐評 |
author |
Fei-Ping Hu 胡斐評 |
spellingShingle |
Fei-Ping Hu 胡斐評 A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
author_sort |
Fei-Ping Hu |
title |
A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
title_short |
A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
title_full |
A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
title_fullStr |
A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
title_full_unstemmed |
A Real-time Multi-angle Ship Detection Method for Aerial Image Sequences Using Convolutional Neural Networks |
title_sort |
real-time multi-angle ship detection method for aerial image sequences using convolutional neural networks |
publishDate |
2018 |
url |
http://ndltd.ncl.edu.tw/handle/sb2q49 |
work_keys_str_mv |
AT feipinghu arealtimemultiangleshipdetectionmethodforaerialimagesequencesusingconvolutionalneuralnetworks AT húfěipíng arealtimemultiangleshipdetectionmethodforaerialimagesequencesusingconvolutionalneuralnetworks AT feipinghu yīgèyīngyòngjuǎnjīshìlèishénjīngwǎnglùyúduōjiǎodùhángpāiyǐngxiàngdejíshíchuánzhīzhēncèfāngfǎ AT húfěipíng yīgèyīngyòngjuǎnjīshìlèishénjīngwǎnglùyúduōjiǎodùhángpāiyǐngxiàngdejíshíchuánzhīzhēncèfāngfǎ AT feipinghu realtimemultiangleshipdetectionmethodforaerialimagesequencesusingconvolutionalneuralnetworks AT húfěipíng realtimemultiangleshipdetectionmethodforaerialimagesequencesusingconvolutionalneuralnetworks |
_version_ |
1719297016594432000 |