(43b) Extracting Relevant Quality Information through Image Data Analysis | AIChE

(43b) Extracting Relevant Quality Information through Image Data Analysis


Stefanov, Z., The Dow Chemical Company
As more powerful image analysis algorithms emerge from computer science and related fields, the application of digital image processing in manufacturing industries has been gaining popularity. There have been several reported case studies of image analysis for control and quality monitoring of industrial processes, such as detecting froth levels in flotation plants [1], industrial flare combustion control through multivariate image analysis [2], and spectral analysis using PCA techniques [3], [4]. However, most of these applications are proof of concepts in niche areas that monitor a single quality (i.e. position of the froth layer) using fast, efficient but relatively simple image analysis methods.

Inspired by these efforts, we addressed the challenge of measuring large amount of particle size distributions data through analyzing images of product samples. Using Hough transform based search algorithms [5], accurate quantitative information about the particle size distribution could be obtained. Scaling up the analysis algorithm to other product images allows us to create a particle evolution profile (particle size distribution over a set of experimental conditions) that could be used to discriminate against products of poor quality.  This approach was applied to two industrial case studies.

In the first case, the product sample consists of spherical droplets in solution. Images of the droplets suspended in solution are taken while heating the solution to the desired temperature. The expansion and contraction profile of the droplets within the microscope image was hypothesized to be a key quality control parameter that needs to be investigated.  In the second case, the quality of the crystalline structures is evaluated through image analysis as the process conditions of the reactor and centrifuge have been modified in order to reduce cycle batch time.  The crystal length distribution was of interest and compared with length size distributions collected before these changes.

The image analysis based approach yielded successful results at a fraction of the cost of other conventional methods. Through analyzing the particle size distribution data, standard multivariate analysis techniques identified the most relevant information in the profile that could be used for future development of online in-situ monitoring techniques.


[1]            S. Africa, P. O. Wits, and S. Africa, “Digital Image Processing As A Tool for On-Line Monitoring of Froth in Floatation Plants,” Miner. Eng., vol. 7, no. 9, pp. 1149–1164, 1994.

[2]            D. Castin, B. C. Rawlings, and T. F. Edgar, “Multivariate Image Analysis ( MIA ) for Industrial Flare Combustion Control,” 2012.

[3]            L. L. Simon, K. Abbou, Z. K. Nagy, and K. Hungerbuhler, “Bulk video imaging based multivariate image analysis , process control chart and acoustic signal assisted nucleation detection,” Chem. Eng. Sci., vol. 65, no. 17, pp. 4983–4995, 2010.

[4]            M. H. Bharati and J. F. Macgregor, “Multivariate Image Analysis for Real-Time Process Monitoring and Control,” no. Figure 2, pp. 4715–4724, 1998.

[5]            T. J. Atherton and D. J. Kerbyson, “Size invariant circle detection,” vol. 17, no. February 1997, pp. 795–803, 1999.