Butterfly Images Recognition using Global Color Feature and Multiple Local Descriptors Matching

• Main Authors: Ting-Ying Lin, 林廷穎
• Other Authors: Jen-Chang Liu
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/37529h

碩士 === 國立暨南國際大學 === 資訊工程學系 === 105 === As tourism in Taiwan continues to thrive in recent years, many people have the chance to see countless beautiful butterflies while doing excursions in the outdoors or relaxing in parks. Most people are drawn by the eye-catching colorful wings of butterflies and become curious, wanting to know more about them. Traditionally, most people look to paper-based illustrated guides to butterflies or search websites about butterflies on the internet for information query. However, these two query approaches rely on the five butterfly families as indexes in almost all cases. This means people who are unfamiliar with butterflies need to know which family a specific butterfly belongs to before they can make a query. This is very inconvenient and time consuming. Some users will use an image to look for a matching image by using the “Google Images” website. However, the query results have more overseas butterfly species than Taiwanese ones and the correct answer rate is extremely low. This thesis studied the content-based image search for butterfly identification. The experimental dataset included 94 most common butterfly species in Taiwan. The butterfly images used for searching consisted of 10 images per species, amounting to a total of 940 images. These images underwent pre-processing before feature extraction and final feature matching, which was divided into two models. The first model used a global feature and then matched by using histogram intersection, the K-dimensional tree and the support vector machine. The second model used the local features including five single descriptors, which were SIFT, LIOP, RI, DAISY and GB. Single descriptor matching and multiple descriptors matching methods, including Density, CAT, Ratio and Ranking were experimented. The two models were applied in the experiments for butterfly image matching. The extracted global features underwent 5-fold cross validation. According to the experimental results, histogram intersection achieved an average precision up to 90.11% in the Top10 and an average query time of 0.000036s. To accelerate the search speed, a K-dimensional tree was used in the experiment, which achieved an average precision up to 89.68% in the Top 10 and an average query time of 0.000007s. A conclusion could then be drawn that the K-dimensional tree, despite 0.43% lower in the recognition rate, was 5.14 times faster in the average query time than the histogram intersection. A support vector machine in machine learning was also used in the experiment, which achieved an average precision up to 95.31% in the Top 10 and thus provided the best recognition rate among the three methods. According to the experimental results from the local feature matching with multiple descriptors, Density achieved a Mean Average Precision (mAP) higher than the average achieved by the other three methods, which was about 0.2484, making it a more appropriate method for butterfly identification for further investigation.