碩士 / 國立成功大學 / 系統及船舶機電工程學系 / 104 / In order to investigate fish detection and recognition, we choose the method of convolutional neural network. This research has two parts. The first part using small amount of data to build and test computing architecture. The Second part using large amounts of data to improve architecture based on data-driven.
In the first part, we capture fish and non-fish image from the video on vessel. Considering how to build architecture of convolutional neural network and which image has better effect. Result of the experiment shows that adding gradient strength on input layer can improve fish detection rate to 95%; In fish detection, there are five kinds of result: three labels fishes, other fish and non-fish. The result is better than two labels include target fish and non-target fish. Fish recognition rate of dolphin is 73.6% in two targets and 78.2% in five targets. We build several architectures. The best fish detection rate is 96% and the best recognition rate is 92.5% in tuna and 93.7% in dolphin. Using the best fish detection CNN to detect images on real vessel still have a lot of false alarms.
In second part, we collect more images and have more output labels of non-fish, which reduces fish detection false alarm. We also build several architectures in second parts. The best fish detection rate is 85% and the best detection rate is 99% in non-fish and 83% in tuna.
| Identifer | oai:union.ndltd.org:TW/104NCKU5345021 |
| Date | January 2016 |
| Creators | Tzu-LiangWei, 魏子量 |
| Contributors | Chung-Horng Lin, 林忠宏 |
| Source Sets | National Digital Library of Theses and Dissertations in Taiwan |
| Language | zh-TW |
| Detected Language | English |
| Type | 學位論文 ; thesis |
| Format | 66 |
Page generated in 0.0106 seconds