IEEJ TRANSACTIONS ON ELECTRICAL AND ELECTRONIC ENGINEERING
IEEJ Trans 2014; 9: 31–38
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI:10.1002/tee.21933
Paper
Spatial-temporal Fusion for Flotation Froth Image Denoising Based
on BLS-GSM Method in Curvelet Domain
Jinping Liu
∗
, Non-member
Weihua Gui
∗
, Non-member
Zhaohui Tang
∗a
, Non-member
Qing Chen
∗,∗∗
, Non-member
The visual appearance of flotation froth surface involves a major cue about the flotation performance, which is significant
for machine-vision-based flotation process monitoring and control. However, the froth image suffers from noise contamination
inevitably, which incurs serious negative effects on the visual feature extraction of froth images. This paper presents a spatial-
temporal image denoising scheme based on statistical modeling of the froth image in the Curvelet domain and weighted processing
of the relative patches of adjacent image sequences. First, the Gaussian scale-mixture model of the image coefficients in the
local spatial neighborhood is investigated according to their statistical distribution property to get the clean coefficients based on
inner-frame content by using Bayesian least-squares estimation. Then, the temporal patches from adjacent image sequences are
performed with weighted impact factors according to the similarities of the relative patches after motion compensation. Thus,
clean image coefficients based on spatial-temporal content are achieved. This method is validated by the simulated additive noise
removal and the real industrial image processing. The results of simulation and real application of image noise elimination reveal
the excellent performance of this method, which can effectively reconstruct the froth image while protecting more bubble details
for the following froth description. © 2013 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.
Keywords: Gaussian scale-mixture model, temporal-spatial fusion, Curvelet transformation, motion compensation, weighted impact factor
Received 16 May 2012; Revised 14 September 2012
1. Introduction
Froth flotation is an ore dressing technology used worldwide by
utilizing air bubbles to separate mineral materials based on their
relative affinity to water. In terms of the daily tonnages of ores
processed globally, froth flotation is the most important technology
for separating valuable minerals from crude ores. It is widely
accepted that the visual appearance of the upper surface of the froth
layer is closely related to the mineral separation performance [1].
In order to improve the flotation automatic control level for
achieving the ultimate higher production indices, a flourishing
literature can be found in both academia and industrial research
investigating machine-vision-based flotation process monitoring
and control. Some satisfactory experimental reports can be found
in the Refs [2,3]. An elaborate survey of the existing froth image
processing and froth feature extraction is reported in Ref. [4],
which reveals that the visual appearance of the froth is an effective
complement to production performance evaluation and flotation
process control.
In terms of the froth-image-based flotation process monitoring,
the froth image processing and the corresponding visual feature
extraction are very important. Unfortunately, the froth images
acquired from the flotation plant are inevitably corrupted by serious
noise under poor illumination, accompanied by the influence of
the serious dusts and mists and the electromagnetic effect of the
electronic equipment in the plants. The existing noise blurs and
a
Correspondence to: Zhaohui Tang. E-mail: ljp202518@163.com
* School of Information Science & Engineering, Central South University,
Changsha 410083, Hunan, P.R. China
** School of Computer and Communication, Hunan University of Tech-
nology, Zhuzhou 412008, Hunan, P.R. China
declines the visual effect of the froth images, which will of course
bring a series of problems to the visual feature extraction of froth,
and even mislead the successive flotation process monitoring and
control. Hence, froth image denoising is an essential preprocessing
step for the subsequent froth visual feature extraction.
Image restoration from a noisy image has long been one of
the hottest issues of image processing. The commonly used image
denoising methods can be divided into three categories, generally,
according to the pixel processing modes, namely spatial domain
denoising [5,6], temporal domain denoising [7,8], and spatial-
temporal domain denoising [9,10]. The simple temporal or spatial
domain denoising involves the design of proper filters to restore
or improve the visual appearance of the image. With better under-
standing of image noise models, a bunch of image denoising algo-
rithms has recently emerged, such as median filtering [11], mean
filtering [12], Fourier transform-based low-pass filtering [13],
partial differential equation (PDE)-based image denoising [14],
wavelet-transformation-based denoising [5–8], and so on. Most of
these are designed to process special types of images or remove
noise with special statistical distributions. With further inves-
tigation of the statistical distribution of the image coefficients
in the transform domain, the most accepted and widely applied
image denoising methods currently are the Bayesian-inference-
based noise-free image estimations, which are usually carried out
by transforming the image to special transformation domains and
making use of the Bayesian inference to get the optimal coeffi-
cient estimation by modeling the image coefficients with a proper
statistical model.
As it has the merits of multiresolution representation, wavelet
transformation has been widely used in image and digital signal
processing. Donoho [5] presented the wavelet threshold denoising
algorithm, given the universal threshold (hard threshold), and
proved its optimality in the progressive sense. However, the
© 2013 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.