Hits:

Indexed by:会议论文

Date of Publication:2019-01-01

Included Journals:EI

Volume:11854

Page Number:379-389

Key Words:Shoeprint; Semantic segmentation; Generative adversarial nets

Abstract:Shoeprints reflect some physiological characteristics of human beings, similar to fingerprints, which are important clues for criminal investigations. Extraction of shoeprints from images taken in crime scenes is a key preprocessing of the shoeprint image retrieval. It can be seen as a binary semantic image segmentation problem. Both traditional algorithms and existing deep learning approaches perform poorly for this problem, since shoeprint images contain various and strong background textures, are contaminated by serious noises and incomplete usually. This paper innovatively presents a framework with generative adversarial net (GAN) for the problem. Multiple generative networks, loses and adversarial networks are designed and compared within the framework. We also compared our method with the-state-of-the-art deep learning approaches on the professional shoeprint dataset with the evaluation criterion called MSS. The MSS of FCN, Deeplab-v3 and our method are 50.3%, 61.5%, and 75% on the dataset.