Forensic Sketch To Mugshot Matching Algorithm Based On Dynamic Difference Of Gaussian Oriented Gradient Histogram

• 主要作者: Setumin, Samsul
• 格式: Thesis
• 語言: 英语
• 出版: 2019
• 主題: T Technology; TK Electrical Engineering. Electronics. Nuclear Engineering
• 在線閱讀: http://eprints.usm.my/55422/

An automatic photo retrieval system based on facial sketch has very useful application in criminal investigations. The face sketch and photograph are from different modality. In inter-modality matching approach, it is unclear which descriptor is the most modality-invariant. Next, the real-world photo may be exposed to lighting variation and the sketch may experience some degrees of shape exaggeration with very less accurate details. With these effects, the retrieval rate reduces significantly. In this research work, at the beginning, the most modality-invariant local hand craftedde scriptor is determined. Next, a new fiducial points for face alignment and a newdescriptor called Difference of Gaussian Oriented Gradient Histogram (DoGOGH) are introduced to reduce the factor of shape exaggeration and to minimize the illumination effects, respectively. It is followed by new feature extraction methods called Dynamic DoGOGH (D-DoGOGH) and Cascaded Static and Dynamic DoGOGH (C-DoGOGH) to really cater for the shape exaggeration effects. The accuracy and speed are improved further after incorporating feature fusion, Patch of Interest (PoI) and score fusion into the proposed method. The experimental results for CUHK Face Sketch Database (CUFS) and CUHK Face Sketch FERET Database (CUFSF) datasets demonstrate that the proposed method outperforms the state-of-the-art methods. It gives rank-1 accuracy of 100% and 95.48% for the CUFS and CUFSF datasets, respectively. The evaluation is extended further to semi-forensic and forensic sketch datasets to indicate that the proposed method is feasible to be used in the real-world criminal investigations. It gives rank-1 accuracy improvements of 28.56% and 66.77% for the semi-forensic and forensic sketch datasets, respectively.