Home // EMERGING 2022, The Fourteenth International Conference on Emerging Networks and Systems Intelligence // View article
Authors:
Shin Kajihara
Masato Okazaki
Chika Oshima
Koichi Nakayama
Keywords: Difference detection; deep learning; physical search system.
Abstract:
This paper introduced the physical search system (PSS), which detects all objects displaced in a physical space with two cameras and a computer, based on image difference detection technology. A linking method (LM) improved the image clustering accuracy in the PSS. Using the PSS, displaced objects are detected by the differences between two images ordered in a time series photographed by two cameras each. In a two-step feature-clustering algorithm (TFA), features of the displaced object images are extracted from the deep residual network ResNet-50 pre-trained on ImageNet. Then, the displaced object images with features are grouped into clusters using a two-pass process. However, these clusters can include a few noisy images. In this paper, therefore, in parallel with the execution of TFA, the displaced object images were paired by the LM, as follows. Some cropped displaced object images were photographed by two cameras at the same time. By calculating the similarities between the displaced object images, pairs of images of the same displaced objects could be created. Namely, the noisy images do not pair with other images. Finally, the displaced object images in the clusters created via TFA that did not overlap with the images of pairs of groups created by LM were deleted. The results of the experiment showed that the method of combining the two algorithms indicated higher clustering accuracy than using TFA alone.
Pages: 13 to 18
Copyright: Copyright (c) IARIA, 2022
Publication date: November 13, 2022
Published in: conference
ISSN: 2326-9383
ISBN: 978-1-61208-993-5
Location: Valencia, Spain
Dates: from November 13, 2022 to November 17, 2022