Filtered local pattern descriptor for face recognition and infrared pedestrian detection

The Journal of Engineering

Ning Sun¹  Jixin Liu¹  Guang Han²  Hengchao Yan² 

[1] Engineering Research Center of Wideband Wireless Communication Technology, Ministry of Education, Nanjing University of Posts and Telecommunications, Nanjing 210003, People's Republic of China

关键词: traditional local pattern descriptor;    multitype filter banks;    feature representation;    multiscale filter banks;    face recognition;    FLPD;    filtered local pattern descriptor;    TLPD feature extraction;    infrared pedestrian detection;   

DOI  :  10.1049/joe.2016.0307

学科分类：工程和技术（综合）

来源: IET

【 摘 要 】

In recent decades, the local pattern descriptor has achieved tremendous success in the field of face recognition, pedestrian detection, and image texture analysis. This study presents a generic approach, called the filtered local pattern descriptor (FLPD), which expands the traditional local pattern descriptor (TLPD) by using multi-scale and multi-type filter banks. The FLPD encodes the local information of an image based on the convolutional sum of the sub-image blocks and the filter banks, instead of the original pixel values in the TLPD. This design can effectively increase the diversity of the TLPD feature extraction, thereby enhancing the ability of feature representation and its reliability. Two FLPD-based feature representation methods are proposed for the face image and the pedestrian image. To evaluate the performance of the proposed FLPD, extensive experiments on face recognition and infrared pedestrian detection are conducted using several benchmark image datasets. The experimental results illustrate that the FLPD has a significant advantage in the discrimination and stability of feature extraction, and is able to achieve a satisfactory accuracy in comparison with state-of-the-art methods. It is demonstrated that the FLPD is a powerful and convenient extension of the TLPD by filter banks, and suitable to be implemented as feature extraction into approaches to solve the binary or multi-class image classification problems.

【 授权许可】

CC BY