Butterfly larvae recognition and App development based on texture features - A case study of common species for Yangming Mountain in Taiwan

• Main Authors: Chang, Kuan-Chung, 張冠中
• Other Authors: Chen, Ching-Yung
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/4254g5

碩士 === 國立高雄科技大學 === 電腦與通訊工程系 === 107 === This paper, based on pattern recognition, develops and implements a butterfly larvae identification system for common species in Yangming Mountain, Taiwan. In light of that texture features of butterfly larvae are special to individual species and only a limited number of training images can be collected, this paper tends to design the recognition algorithm using traditional machine learning methods with texture feature adopted. Thanks to the repeated texture feature in the same image, a segmentation pre-processing is proposed to obtain multiple sub-images for increasing training samples. The classification accuracy is further improved by a ballot box scheme for all sub-images of a test image. Among the others, the Histogram of Gradients (HOG) is selected as the texture feature and the Support Vector Machine (SVM) is chosen as the classifier. For those species having either similar or unobvious texture, using only texture feature might cause classification error. In order to improve the recognition accuracy, this paper further proposes a modified algorithm supplemented by color features. Among them, the statistics of hue and saturation are selected as color features and the K-Nearest Neighbors method (KNN) is chosen as the classifier. As regards the implementation, the proposed system is implemented on a smart phone by developing an application (App) of the Android operating system. An interactive user interface is provided to guide the user to shoot with the built-in rear camera and assist in cutting out the foreground part as the image to be recognized. Some experiments are performed to show that the proposed butterfly larvae identification system can yield good recognition results.