- 學位論文
選擇性還原角之常見鞘翅目昆蟲辨識系統
A Common Coleoptera Recognition System based on Feeler Removal and Recovery
摘要
昆蟲是地球上數量、種類最多且分布最廣泛的生物，昆蟲中的鞘翅目因為具備特殊之形態外觀，近來成為最受大眾喜愛之昆蟲。目前鞘翅目昆蟲(俗稱甲蟲)的線上教學系統，然而受限於使用方式，無法進一步提升系統可用性。我們希望建立一套鞘翅目昆蟲的影像自動化辨識系統，以物種影像作為系統檢索的依據，系統對物種影像進行自動辨識，再依據辨識結果回傳資料庫中相關資料，該系統可提供方便的使用介面。然而，昆蟲不同的足部姿態，會嚴重影響辨識結果，成為建立物種影像自動辨識系統的主要障礙。 為了降低足部姿態對辨識之干擾，本論文提出利用昆蟲足部的幾何特徵移除影蟲體足部之機制，移除足部之影像可大幅提升辨識率。但由於部份雄性昆蟲外型特徵(觭角、螯夾)在去足過程中會一併被去除，因而被誤判為雌性昆蟲，所以去足後須要再還原角的部份。對還原角後的影像，重新進行辨識可以因分辨雌雄而再提升整體辨識率。由於此還原角程序消耗大量運算資源，為了減少不必要的還原角程序，我們初步判斷若昆蟲落入可能雌雄混淆之種類後，透過分析昆蟲角所在位置的物件面積比例，判定該影像是否需要進行還原角處理。藉由此的分析，可減少進行還原角處理昆蟲的數量。 本系統根據昆蟲外形特徵的差異，擷取CCD、Fourier descriptor、YCbCr直方圖作為特徵向量，並以K-Nearest Neighbor分類法建構辨識引擎。目前系統中建立20種常見鞘翅目昆蟲資料庫，辨識率為96.43%；透過事先分析是否進行還原角處理，相較於對全部測試影像進行還原角平均可節省35.47%的辨識時間。 我們未來將持續擴充鞘翅目昆蟲資料，並使用腹面影像以期進一步提升辨識率。
並列摘要
Feeding coleopetra becomes a popular activity for elementary school students in Taiwan. An user-friendly automatic coleopetra recognition and teaching system can provide a self-learning mechanism for students. However, without an unfriendly user interface, the system sometimes discourages the user and results in a valueless system. A well-suited way is directly providing the interested coleopetra image to the system with image recognition ability. However, the main obstacle of coleopetra recognition is the irregular pose of legs. We have proposed a leg clipping mechanism to remove the legs of the coleopetra in the image. Although using the leg-clipped coleopetra images can achieve a better recognition result, the leg-clipping process also lops the feelers which are important features to distinguish male and female for some species of coleopetra. A feller recovery process is applied to recover the fellers. However, this process is time consuming. In order to save the recognition time, we propose an optional species feeler recovery process. The leg-clipped images are recognized first, then the feller recovery process are only applied to those gender confusion species. In our system, twenty species of coleopetra which are common in Taiwan are incorporated in our system. The information of shapes (CCD and Fourier Descriptor) and colors (CbCr) are extracted as the features. K-nearest neighbor method is applied to be the recognizer. Five samples of each species- total one hundred samples, organize the training set. Due to the difficulty of collection, number of samples of each species in the testing set is seven. Our system achieves 96.43% recognition rate. In addition, the optional species feeler recovery process saves up to 35.47% of the recognition time comparing to the time consumed for all species feeler recovery.