台灣鄰近海域常見軍艦影像辨識系統之建構

AN IMAGE RECOGNITION SYSTEM FOR NAVAL SHIPS IN SEAS AROUND TAIWAN

近年來中國解放軍船艦常於台灣附近海域演練，面對持續增加的威脅，台灣對於軍艦辨識需求提升。本論文之研究目的是利用卷積神經網路（CNN）來建構一套能辨識周遭海域可能出現的大型軍艦之辨識系統。本實驗使用實船影像共1743張構建資料庫，分屬8國20艘航母或登陸作戰相關船艦。由於軍艦影像取得不易，因此使用平移、對比增強、加入高斯雜訊、影像加霧等四種資料增強的方式來擴充影像資料庫，擴建後的影像資料庫共8715張影像。本研究之船艦辨識系統以卷積神經網路建構，共選用四種網路GoogLeNet（Inception-v1）、Inception-v3、ResNet-18、ResNet-101訓練，並比對這四種CNN網路對於正常情況與雲霧繚繞的情況下的實船影像辨識率。研究結果顯示，本研究之船艦辨識系統在天氣良好的情況下的辨識率可達96.3%。在雲霧繚繞的情況下的辨識率可達89.5%，特別是對於美中兩國的航空母艦都有極佳的成功率，但對於日韓的大型兩棲登陸艦的效果稍差。

Taiwan has been facing the threats from China and need to improve the capability of identification of naval ships. This research use convolution neural network (CNN) to establish a system for this purpose. A database of total 1743 images of real ships were established including 20 types of aircraft carrier or landing craft related ships from 8 countries. Image processing techniques, such as shifting, contrast enhancing, adding Gaussian noise, and fogging, were used to expand the image database to 8715 images. Four networks were used for training, including GoogLeNet (Inception-v1), Inception-v3, ResNet-18, and ResNet-101. The identification rate were compared for normal and foggy weather using these four networks. The results show that the system's best identification rate is 96.3% in normal weather condition, and 89.5% in foggy condition. It is especially good for identifying American and China's aircraft carries. But poor at Japan and S. Korea's large amphibious assault ships.