Dynamic fast local Laplacian completed local ternary pattern (dynamic FLapCLTP) for face recognition
Today, face recognition has become one of the typical biometric authentication systems used for high security. Some systems may use face recognition to enhance their security and provide high protection level. Feature extraction is considered to be one of the most important steps in face recognition...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/30383/1/Dynamic%20fast%20local%20laplacian%20completed%20local%20ternary%20pattern%20%28dynamic%20flapcltp%29.pdf http://umpir.ump.edu.my/id/eprint/30383/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Today, face recognition has become one of the typical biometric authentication systems used for high security. Some systems may use face recognition to enhance their security and provide high protection level. Feature extraction is considered to be one of the most important steps in face recognition systems. The important and interesting parts of the image in feature extraction are represented as a compact feature vector. Many features, such as texture, colour and shape, have been proposed in the image processing fields. These features can also be classified globally or locally depending on the image extraction area. Texture descriptors have recently played a crucial role as local descriptors. Different types of texture descriptors, such as local binary pattern (LBP), local ternary pattern (LTP), completed local binary pattern (CLBP) and completed local ternary pattern (CLTP), have been proposed and utilised for face recognition tasks. All these texture features have achieved good performance in terms of recognition accuracy. Although the LBP performed well in different tasks, it has two limitations. LBP is sensitive to noise and occasionally fails to clearly distinguish between two different texture patterns with the same LBP encoding code. Most of the texture descriptors inherited these limitations from LBP. CLTP is proposed to overcome the limitations of LBP. CLTP performed well with different image processing tasks, such as image classification and face recognition. However, CLTP suffers from two limitations that may affect its performance in these tasks: the fixed value of the threshold value that is used during the CLTP extraction process regardless of the type of dataset or system and the longer length of the CLTP histogram than that of previous descriptors. This study focused on handling the first limitation, which is the threshold selection. Firstly, a new texture descriptor is proposed by integrating the fast-local Laplacian filter and the CLTP descriptor, namely, fast-local Laplacian CLTP (FLapCLTP). The fast-local Laplacian filter can help in increasing the performance of the CLTP due to its extensive detail enhancements and tone mapping; this contribution is handled by the constant threshold value used in CLTP. A dynamic FLapCLTP is then proposed to address the aforementioned issue. Instead of using a fixed threshold value with all datasets, a dynamic value is selected based on the image pixel values. Therefore, each different texture pattern has different threshold values to extract FLapCLTP from the pattern. This dynamic value is automatically selected according to the centre value of the texture pattern. Therefore, a dynamic FLapCLTP is proposed in this study. Finally, the proposed FLapCLTP and dynamic FLapCLTP are evaluated for facial recognition systems using ORL Faces, Sheffield Face, Collection Facial Images, Georgia Tech Face, Caltech Pedestrian Faces 1999, JAFFE, FEI Face and YALE datasets. The results showed the priority of the proposed texture compared with previous texture descriptors. The dynamic FLapCLTP achieved the highest recognition accuracy rates with values of 100%, 99.96%, 99.75%, 99.69%, 94.86%, 90.33%, 86.86% and 82.43% using UMIST, Collection Facial Images, JAFFE, ORL, Georgia Tech, YALE, Caltech 1999 and FEI datasets, respectively. |
|---|