Object Detection Using Structure-Preserving Wavelet Pyramid Reflection Removal Network

Images shot through glass will produce reflections. However, these reflections will cause the objects in the images to be unrecognizable. Hence, to improve the recognition rate, the reflections in the images should be removed. Reflection removal has been widely used in the field of deep learning. Al...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on instrumentation and measurement 2022, Vol.71, p.1-11
Main Authors: Hsu, Wei-Yen, Wu, Wan-Jia
Format: Article
Language:English
Subjects:
Deep learning
Feature extraction
Glass
Image enhancement
Image restoration
Machine learning
Object detection
Object recognition
Reflection
reflection removal
structure preservation
Wave reflection
wavelet pyramid
Wavelet transforms
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Images shot through glass will produce reflections. However, these reflections will cause the objects in the images to be unrecognizable. Hence, to improve the recognition rate, the reflections in the images should be removed. Reflection removal has been widely used in the field of deep learning. Although these methods have good results, they all assume that reflection removal is performed in a specific situation and provide their own datasets for research, such as strong reflections in some local areas, and this limitation will lead to the inability to effectively remove the reflections in the real world, which will affect the recognition of objects. To address this issue, we propose a novel structure-preserving wavelet pyramid reflection removal network (SpWPRRNet) to achieve effective background structure preservation and reflection removal to further improve the object recognition rate in images. After wavelet decomposition, we put the high- and low-frequency images of each level into reflection layer removal and detail enhancement subnetwork (RDSn) and structure preservation subnetwork (SPSn), respectively, and then use structure-level fusion (SLF) and inverse stationary wavelet transform (ISWT) to restore clean images recursively. In addition, to further separate the reflection layer and the transmission layer, we propose the reflection layer information transmission (RLIT), through which the reflection layer features of the high-frequency images can be extracted to help the SPSn to effectively separate these two layers to achieve the results of reflection removal and improve object recognition rate. The experimental results indicate that the proposed method can greatly improve the object recognition rate in images.
ISSN:0018-9456
1557-9662
DOI:10.1109/TIM.2022.3204081