Implementation of Image Recognition System based on the Common Butterflies in Taiwan

• Main Authors: Wei-Pin Wang, 王維彬
• Other Authors: Jen-Chang Liu
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/90673155037932379927

碩士 === 國立暨南國際大學 === 資訊工程學系 === 102 === Due to the rising awareness of eco-tourism, many people go outdoor for nature sightseeing especially for butterfly species recent years. Butterflies are commonly seen in Taiwan and hence become a good topic for science education. The common way people observe butterflies is through butterfly illustrated handbook which is not easy to carry and time-wasting. In this thesis, we propose a butterfly image recognition system on the Android platform. Within the system, the users can easily identify the butterflies on smartphones or tablets. Through a few simple operation steps, the system will use image features to recognize butterfly species and return similar butterfly species pictures and detailed information. In this thesis, 94 kinds of common butterflies in Taiwan were used in the experiments, each butterfly type has 10 pictures, a total of 940 pictures are included in the database. After normalizing the butterfly images, a custom mask will be trained by using butterfly images in the database. Then we remove the background through GrabCut algorithm with the custom mask. Feature extraction includes the 2-D Fourier Descriptors, Local Color Histogram, and SIFT feature, in the single feature butterfly identification experiments. Besides, 2-D Fourier Descriptors and Local Color Histogram were combined in the experiments. We discovered that using combination of multi-feature is better than using only a single feature on the butterfly image recognition; moreover, Linear Evidence Combination reached 77.02% Top 1 average precision. Concluding from the experimental results and considering the computing speed of mobile devices, we apply custom mask in GrabCut and take the Local Color Histogram method in the implementation. For the databases of 94 species, the average computational time is 6.92 seconds.