Article: A novel approach to retrieve unlabelled images Journal: International Journal of Business Intelligence and Data Mining (IJBIDM) 2022 Vol.20 No.1 pp.56 - 76 Abstract: In this paper an attempt has been made to propose an architecture of the photo search engine to retrieve unlabelled images. In order to have fast execution the images are transformed into space using data processing pipeline. The proposed pipeline consists of HOG face detection and extraction, face landmark estimation, indexer and transformer. The image is passed through the data pipeline and the encoded faces in the input image are compared with other vectors by computing l2 norm distance, resulting in top N results to the users. The performance on faces94, faces95, faces96 and grimace public datasets has been observed individually and on superset of all the datasets. It has been observed that face94 has the best, whereas Grimaceface96 has the worst performance. Further it has been observed that proposed system witnessed an accuracy of 99% and precision of 94.7% when the search is performed on a combination of all datasets. Inderscience Publishers - linking academia, business and industry through research

Title: A novel approach to retrieve unlabelled images

Authors: Ashish Pahwa; Deepali Kamthania; Aayush Gupta; Chirag Jain

Addresses: Vivekananda School of Information Technology, VIPS, AU Block (Outer Ring Road), Pitampura, GGSIP University, Delhi 110034, India ' Vivekananda School of Information Technology, VIPS, AU Block (Outer Ring Road), Pitampura, GGSIP University, Delhi 110034, India ' Vivekananda School of Information Technology, VIPS, AU Block (Outer Ring Road), Pitampura, GGSIP University, Delhi 110034, India ' Vivekananda School of Information Technology, VIPS, AU Block (Outer Ring Road), Pitampura, GGSIP University, Delhi 110034, India

Abstract: In this paper an attempt has been made to propose an architecture of the photo search engine to retrieve unlabelled images. In order to have fast execution the images are transformed into space using data processing pipeline. The proposed pipeline consists of HOG face detection and extraction, face landmark estimation, indexer and transformer. The image is passed through the data pipeline and the encoded faces in the input image are compared with other vectors by computing l2 norm distance, resulting in top N results to the users. The performance on faces94, faces95, faces96 and grimace public datasets has been observed individually and on superset of all the datasets. It has been observed that face94 has the best, whereas Grimaceface96 has the worst performance. Further it has been observed that proposed system witnessed an accuracy of 99% and precision of 94.7% when the search is performed on a combination of all datasets.

Keywords: face recognition; FR; deep learning; histogram of oriented gradients; HOG; FaceNet architecture; machine learning; support vector machine; SVM.

DOI: 10.1504/IJBIDM.2022.119948

International Journal of Business Intelligence and Data Mining, 2022 Vol.20 No.1, pp.56 - 76

Received: 09 Oct 2019
Accepted: 23 Dec 2019
Published online: 04 Jan 2022 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: A novel approach to retrieve unlabelled images

Keep up-to-date