.png){kind=icon}
.png){kind=icon}
Return
Approximate object location deep visual representations for image retrieval
.png){kind=icon}
PRE
AI
DOI：10.1016/j.displa.2023.102376.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Image descriptor based on Convolution Neural Network (CNN) has been presented to represent image in recent lots of works, outperforming the traditional feature for image retrieval. In this paper, a new method is proposed to capture real object regions in images for better search. In contrast to the existing method that process image as a whole, our method focuses on searching for approximate object region in an image. The proposed method mainly has the following two aspects: (i) we employ the sliding windows with the multiple scales over the image, to extract corresponding local CNN features. After processing feature, an optimized image representation with several categories of weights has been prepared for image retrieval and re-rank. (ii) On the basis of (i), we construct a new framework called approximate object location to search out the most similar domain, which can be applied in renewed retrieval, in an image for the query. This proposed framework provides a vector consisting of many locals in different scales for an image. According to experiments, our proposed method outperforms most current approaches based on CNN, and excels many previous algorithms based on costly bag-of-words.
Keywords:
Image retrieval
Approximate object location
Deep learning