- 學位論文
利用芮氏指標最小化進行非督導式指紋辨識
Unsupervised Fingerprint Recognition Based on the Minimization of Rand Index
摘要
近年來，指紋辨識技術的發展已日益成熟，不論是在刑事鑑識科學、門禁安全管制或是個人資訊的隱私安全防護方面都有其精確性與可靠性，然而現行的指紋辨識屬於「督導式」的辨識模式，它必須先行收集已知人士的指紋，並建立範本資料庫，才能辨識未知的指紋。這種模式並無法處理未登錄於資料庫中之人的指紋，因此不符合實際的犯罪現場使用需求，因為資料庫中很可能並不曾登錄犯罪者的指紋。本論文因而嘗試探討一種「非督導式」的指紋辨識技術，在不需範本資料庫的情況下，將所採集的未知指紋進行自動分群，使其達到每一群內只包含同一人的指紋。則若在不同犯罪現場所採集到許多無法辨識的指紋時，我們仍有機會藉由分群技術了解是否有某些指紋屬於相同的人，提供更多有助於刑案偵查的線索。同時，若我們想標示每一指紋所屬何人時，只需以群組的方式進行即可，毋須逐一標示指紋。 本論文利用芮氏指標來做為指紋分群的準則，該指標反應分群的錯誤情形，當分群數目與實際人數相同，且每群內指紋皆屬同一人時，該指標值將為零。我們透過兩兩指紋的相似性量測值來估算分群的芮氏指標，並利用遺傳演算法找出能使芮氏指標達到最小的分群組合。
並列摘要
With rapidly developing techniques, automatic fingerprint recognition has shown its potential in criminal investigation, security systems, and privacy applications. However, current existing fingerprint recognition techniques operates in a supervised manner, which requires that the fingerprints of known persons be collected in advance to establish the template patterns. Such a supervised framework cannot handle the fingerprints from the persons not enrolled in the database. To solve this problem, this thesis studies an unsupervised fingerprinting framework. It is aimed to partition a set of unknown fingerprints into several clusters, such that each cluster contains fingerprints exclusively from only one person. This framework is particularly advantageous for criminal investigation, since fingerprints collected from multiple criminal spots could be grouped together to infer that the criminals of different cases may belong to the same persons, even though the persons are not enrolled in the database. By clustering fingerprints from the same person, the human efforts require to check each fingerprint can be reduced to only checking each cluster. The proposed system uses Rand index as a criteria for cluster generation. The index represents the miss-clustering errors and its value is zero only when all the fingerprints from the same person are grouped into a cluster and when the number of generated clusters equals the number of involved persons. We use the inter-fingerprint similarities to estimate the Rand index, and then apply genetic algorithm to minimize the estimated Rand index, so that the resulting clustering can be optimized. Our experiments conducted using FVC2002 DB3 show the feasibility of the proposed fingerprint clustering system.