Deep image retrieval of large-scale vessels images based on BoW model

• Published in: Multimedia tools and applications Vol. 79; no. 13-14; pp. 9387 - 9401
• Main Authors: Tian, Chi, Xia, Jinfeng, Tang, Ji, Yin, Hui
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2020; Springer Nature B.V
• Subjects: Algorithms; Computer Communication Networks; Computer Science; Convolution; Data Structures and Information Theory; Dictionaries; Feature extraction; Image management; Image retrieval; Multimedia Information Systems; Neural networks; Special Purpose and Application-Based Systems; Vessels; Deep feature extraction; Vessel recognition; Image retrieval; Bag-of-words model
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-019-7725-y

This paper focuses on the vessel image retrieval from massive data, whose goal is to identify relevant records quickly and accurately when new images are given. Noteworthy, it is necessary to find features with high representativeness under the impact of moisture. Traditional features are extracted on the basis of single convolution feature and manual feature coding. However, only a few can express key features of vessels’ images due to the incomplete or redundant information. In order to solve this problem, this paper proposes a new strategy. Two dictionary databases are constructed using different convolution layers in VGG16 network; then they are merged to one database that can strongly express the vessel image. Materially, the combined dictionary database consists of two-layer convolution features, which express the original image well with strengthening key information and less redundant information. The algorithm uses BoW (Bag-of-Words) expression of VGG16 neural network in the domain of image retrieval. Compared with traditional methods using SIFT or SUFT features as BoW, experiments on self-build database shows that the proposed algorithm performs better and achieves higher accuracy.