{"title": "High-Speed Airborne Particle Monitoring Using Artificial Neural Networks", "book": "Advances in Neural Information Processing Systems", "page_first": 980, "page_last": 986, "abstract": null, "full_text": "High-Speed Airborne Particle Monitoring \n\nUsing Artificial Neural Networks \n\nAlistair Ferguson \n\nTheo Sabisch \n\nERDC, Univ. of Hertfordshire \n\nDept. Electrical and Electronic Eng. \n\nA.Ferguson@herts.ac.uk \n\nU niv. of Hertfordshire \n\nPaul Kaye \n\nERDC, Univ. of Hertfordshire \n\nLaurence C. Dixon \n\nNOC, Univ. of Hertfordshire \n\nERDC, Univ. of Hertfordshire, Herts, ALtO 9AB, UK \n\nHamid Bolouri \n\nAbstract \n\nCurrent environmental monitoring systems assume particles to be \nspherical, and do not attempt to classify them. A laser-based sys(cid:173)\ntem developed at the University of Hertfordshire aims at classify(cid:173)\ning airborne particles through the generation of two-dimensional \nscattering profiles. The pedormances of template matching, and \ntwo types of neural network (HyperNet and semi-linear units) are \ncompared for image classification. The neural network approach is \nshown to be capable of comparable recognition pedormance, while \noffering a number of advantages over template matching. \n\n1 \n\nIntroduction \n\nReliable identification of low concentrations of airborne particles requires high speed \nmonitoring of large volumes of air, and incurs heavy computational overheads. An \ninstrument to detect particle shape and size from spatial light scattering profiles has \n\n\fHigh-speed Airborne Particle Monitoring Using Artificial Neural Networks \n\n981 \n\npreviously been described [6]. The system constrains individual particles to traverse \na laser beam. Thus, spatial distributions of the light scattered by individual particles \nmay be recorded as two dimensional grey-scale images. \n\nDue to their highly distributed nature, Artificial Neural Networks (ANNs) offer the \npossibility of high-speed non-linear pattern classification. Their use in particulate \nclassification has already been investigated. The work by Kohlus [7] used contour \ndata extracted from microscopic images of particles, and so was not real-time. While \nusing laser scattering data to allow real-time analysis, Bevan [2] used only three \nphotomultipliers, from which very little shape information can be collected. \n\nThis paper demonstrates the plausibility of particle classification based on shape \nrecognition using an ANN. While capable of similar recognition rates, the neural \nnetworks are shown to offer a number of advantages over template matching. \n\n2 The Hyper N et Architecture \n\nHyperNet is the term used to denote the hardware model of a RAM-based sigma-pi \nneural architecture developed by Gurney [5]. The architecture is similar in nature \nto the pRAM of Gorse and Taylor (references in [4]). The amenability of these \nnodes to hardware realisation has been extensively investigated, leading to custom \nVLSI implementations of both nodes [3, 4]. Each HyperNet node is termed a multi(cid:173)\ncube unit (MeU), and consists of a number of subunits, each with an arbitrary \nnumber of inputs. j references the nodes, with i = 1, ... ,Ii indexing the subunits. \nIJ denotes the site addresses, and is the set of bit strings 1J1, ... ,lJn wl1ere n denotes \nthe number of inputs to the subunit. Zc refers to the cth real-valued input, with \nZc E [0,1] and Zc == (1 - zc). For each of the 2n site store locations, two sets are \ndefined: c E M:!o if IJc = 0; c E M:!t if IJc = 1. The access probability p(lJii) for \nlocation IJ in subunit i of hidden layer node j is therefore \n\n(1) \n\nThe activation (ai ) is formed by accumulating the proportional site values (SIS';) \nfrom every subunit. The activation is then passed through a sigmoidal transfer \nfunction to yield the node output (yi). \n\n1 \n\n. \n1 + ea1 / p \n\n(2) \n\n(3) \n\n. \n\n. \n\ny' = u(a1 ) = \n\nwhere p is a positive parameter determining the steepness of the sigmoidal curve. \nBy combining equations (1) and (2), it becomes apparent that the node is a higher(cid:173)\norder or sigma-pi node [9]. A wide variety of learning algorithms have been tailored \nfor these nodes, notably reward-penalty and back-propagation [5]. \n\n\f982 \n\nA. FERGUSON, T. SABISCH, P. KAYE. L. C. DIXON. H. BOWURJ \n\n3 Description of the Particle Monitoring System \n\nThe instrument draws air through the laser scattering chamber at approximately \n1.5 min-1 , and is constrained to a column of approximately 0.8mm diameter at the \nintersection with the laser beam. Light scattered into angles between 300 and 141 0 \nto the beam direction is reflected through the optics and onto the photocathode of \nan intensified CCD (charge-coupled device), thus giving rise to the scattering profile. \nThe imaging device used has a pixel resolution of 385 x 288, which is quantised into \n2562 8-bit pixels by the frame grabbing processor card of the host computer. \n\nData was collected on eight particle types, namely: long and short caffeine fibres; \n31lm and 121lm micro-machined silicon dioxide fibres; copper flakes (2- 5Ilm in length \nand O.lllm thick); 31lm and 4.31lm polystyrene spheres; and salt crystals. An exem(cid:173)\nplar profile for each class is given in figure 1. Almost all the image types are highly \nvariable. In particular, the scattering profile obtained for a fibrous particle is af(cid:173)\nfected by its orientation as it passes through the laser beam. The scattering profiles \nare intrinsically centred, with the scaling giving important information regarding \nthe size of the particle. The experiments reported here use 100 example scattering \nprofiles for each of the eight particle classes. For each class, 50 randomly selected \nimages were used to construct the templates or train the neural network (training \nset), and the remainder used to test the performance of the pattern classifiers. \n\n4 Experimental Results \n\nThe performance of template matching is compared to both HyperNet and networks \nof semi-linear units. In all experiments, high-speed classification is emphasised by \n\n~ \n\n. \n\n\u2022~. ?'., \n\n.. ~.., \n. ' , \nt \n. \n\n, \n\n\" , \n\nl t.. \n: ~ \n. , '. \n'l.J; \n\n, \n\nFigure 1: Exemplar Image Profile For Each Of The Eight Benchmark Classes \n\n\fHigh-speed Airborne Particle Monitoring Using Artificial Neural Networks \n\n983 \n\navoiding image preprocessing operations such as transformation to the frequency \ndomain, histogram equalisation, and other filtering operations. Furthermore, all \nexperiments use the scatter profile image as input, and include no other information. \n\nThe current monitoring system produces a 2562 8-bit pixel image. The sensitivity of \nthe camera is such that a single pixel can represent the registration of a single photon \nof light. Two possible methods of reducing computation, implementable through \nthe use of a cheaper, less sensitive camera were investigated. The first grouped \nneighbouring pixels to form a single average intensity value. The neighbourhood \nsize was restricted to powers of two, producing images ranging in size from 2562 to \n42 pixels. The second banded grey levels into groups, again in powers of two. Each \npixel could therefore range from eight bits down to one. \n\n4.1 Template Matching Results \n\nThe construction of reference templates is crucial to successful classification. Two \napproaches to template construction were investigated \n\n<D Single reference image for each class. Various techniques were applied rang(cid:173)\ning from individual images, to mode, median, and mean averaged templates. \nMean averaged templates were found to lead to the highest classification \nrates. In this approach, each pixel location in the template takes on the \naveraged value of that location across the 50 training images. \n\n\u00ae Multiple templates per class. A K-means clustering algorithm [1] was used \nto identify clusters of highly correlated images within each class. The initial \ncluster centres were hand selected. The maximum number of clusters within \neach class was limited to six. For each cluster, the reference template was \nconstructed using the mean averaging approach above. \n\nTables 1 and 2 summarise the recognition rates achieved using single, and multiple \nmean averaged templates for each particle class. In both cases, the best average \nrecognition rate using this approach was gained with 1282 3-bit pixel images. With \na single template this lead to a recognition rate of 78.2%, increasing to 85.2% for \nmultiple templates. However, the results for both 162 and 82 pixel images are \nreasonable approximations of the best performance, and represent an acceptable \ntrade-off between computational cost and performance. With few exceptions, mul(cid:173)\ntiple templates per class led to higher recognition rates than for the corresponding \nsingle template results. This is attributable to the variability of the particles within \na class. As expected, the effect of grey level quantisation is inversely proportional \nto that of local averaging. \n\nIn order to evaluate the efficiency of the template construction methods, every image \nin the training set was used as a reference template. 2562 8-bit, 1282 3-bit, and 642 \n2-bit pixel images were used for these experiments. However, the recognition rate \ndid not exceed 85%, demonstrating the success of the template generation schemes \npreviously employed. \n\n\f984 \n\nA. FERGUSON, T. SABISCH, P. KAYE, L. C. DIXON, H. BOLOURI \n\nTable 1: Single Template Per Class % Recognition Rates \n\ngrey levels \n\n256 \n128 \n64 \n32 \n16 \n8 \n4 \n2 \n\nimage size \n\n322 \n74.7 \n74.7 \n74.5 \n75.5 \n76.0 \n\n642 \n74.7 \n74.7 \n74.5 \n75.2 \n76.7 \n\n1282 \n75.0 \n75.0 \n75.0 \n74.7 \n76.0 \n\n2562 \n73.5 \n73.5 \n73.0 \n73.0 \n74.0 \n15.5 18.2 11.5 11.5 16.0 73.7 \n58.5 \n68.4 \n23.0 \n69.7 \n\n42 \n82 \n162 \n75.0 \n67.2 \n74.7 \n75.0 68.5 \n74.5 \n66.2 \n74.7 \n74.2 \n74.2 \n66.5 \n74.7 \n75.0 15.5 56.0 \n38.7 \n18.7 \n16.6 \n\n71.0 \n65.5 \n\n69.7 \n46.2 \n\n70.7 \n66.2 \n\n69.7 \n68.7 \n\nTable 2: Multiple Templates Per Class % Recognition Rates \ngrey levels \n\nimage size \n\n256 \n128 \n64 \n32 \n16 \n8 \n4 \n2 \n\n42 \n\n642 \n80.2 \n80.5 \n80.2 \n81.7 \n83.0 \n\n322 \n80.5 \n80.5 \n80.5 \n80.0 \n81.2 \n\n1282 \n80.0 \n80.2 \n80.2 \n81.2 \n83.5 \n\n2562 \n82 \n76.7 10.2 \n78.0 \n69.7 \n77.0 \n78.5 \n69.2 \n76.0 \n78.7 \n67.7 \n76.7 \n78.2 \n56.0 \n80.2 \n78.5 \n82.2 85.2 84.5 84.1 81.0 80.0 43.5 \n72.7 \n39.2 \n0.03 \n69.7 \n\n162 \n79.0 \n79.0 \n79.2 \n78.7 \n79.5 \n\n72.2 \n62.7 \n\n72.2 \n70.7 \n\n74.5 \n70.2 \n\n69.5 \n51.7 \n\n61.2 \n51.7 \n\n4.2 Neural Network Results \n\nA fully connected three layer feed-forward network was used in all experiments. The \nnumber of hidden layer neurons was equal to the square root of the number of pixels. \nThe target patterns were chosen to minimise the number of output layer nodes, while \nensuring an equitable distribution of zeros and ones. Six output layer neurons were \nused to give a minimum Hamming distance of two between target patterns. The \nclassification of a pattern was judged to be the particle class whose target pattern \nwas closest (lowest difference error). The HyperNet architecture was trained using \nsteepest descent, though the line search was hardware based and inexact. The semi(cid:173)\nlinear network was trained using a variety of back-propagation type algorithms, with \nthe best results obtained reported. Both networks were randomly initialised. Due \nto the enormous training overhead, only 162 and 82 pixel images were tried. The \nrecognition rates achieved are given in table 3. \n\nBoth neural networks are significantly better than the single, and some of the mul(cid:173)\ntiple template matching results. With optimisation of the network structures, it is \nlikely that the ANNs could exceed the performance of multiple templates. \n\n\fHigh-speed Airborne Particle Monitoring Using Artificial Neural Networks \n\n985 \n\nTable 3: Neural Network % Recognition Rates \n\nQuantisation Levels \n\n162 4 bit 1162 3-bit 82 4-bit 82 3-bit \n\n83.8 I 82.3 \n\n86.3 \n\n84.5 \n\n83.0 \n77.8 \n\n76.8 \n76.0 \n\n, , , , , \n\n, \n\nClassifier \n\nHyperNet \nSemi-linear \n\nle+09 \n\no~ \n\n,-.. le+08 \n'\" c:: \n'-\" \n~ Ie+{)? \nll! 0;; \n'\" 8 \n\nle+06 \n\n~ \n\nle+05 \n\nle+04 \n\n82 \n\n162 \n\n322 \n\n642 \n\n1282 \n\n2562 \n\nNumber of pixels \n\nFigure 2: Hardware classification speeds for a single pattern against image size \n\n5 Speed Considerations \n\nSingle processor, pipelined hardware implementations of the three classification \ntechniques have been considered. A fast (45ns) multiply-accumulate chip (Logic \nDevices Ltd, LMA201O) was utilised for semi-linear units. Both template matching \nand HyperNet were implemented using the Logic Devices LGC381 ALU (26ns per \naccumulate). The cost of these devices is approximately the same (\u00a310-20). The \nHyperNet implementation uses a bit-stream approach to eliminate the probability \nmultiplications [8], with a stream length of 256 bits. Figure 2 plots single pattern \nprocessing time for each classifier against image size. \n\nFor small image resolutions, the semi-linear network offers the best performance, be(cid:173)\ning almost three times faster than template matching. However, template matching \nand HyperNet yield faster performance at higher image resolutions. At the op(cid:173)\ntimum (indicated by template matching results (\u00a74.1); 1282 pixels), HyperNet is \nalmost seven times faster than the comparable implementation of semi-linear units. \nWhile the hardware performance of template matching is similar to HyperNet, it \nsuffers from a number of disadvantages to which the neural approaches are immune \n\nCD Recognition rate is dependent on the choice of reference images. \n\n\u00ae Multiple reference images must be used to achieve good recognition rates \n\n\f986 \n\nA. FERGUSON, T. SABISCH, P. KAYE, L. C. DIXON, H. BOLOURI \n\nwhich drastically increases the amount of computation required. \n\n@ New reference images must be found whenever a new class is introduced. \n\n@ Difficult to make behaviour adaptive, ie. respond to changing conditions. \n\n6 Conclusions \n\nThe feasibility of constructing an airborne particle monitoring system capable of \nreliable particle identification at high speeds has been demonstrated. Template \nmatching requires multiple reference images and is cumbersome to develop. The \nneural networks offer easier training procedures and equivalent recognition rates. In \naddition, HyperNet has the advantage of high speed operation at large image sizes. \n\nAcknowledgements \n\nThe authors would like to thank Dr. Eric Dykes and Dr. Edwin Hirst at the Uni(cid:173)\nversity of Hertfordshire, Dr. Kevin Gurney at BruneI University, and the EPSRC \nand the Royal Society for financial support. \n\nReferences \n\n[1] Stephen Banks. Signal Processing, Image Processing, and Pattern Recognition. \n\nPrentice Hall, 1990. \n\n[2] A V Bevan et al. The application of neural networks to particle shape classifi(cid:173)\n\ncation. Journal of Aerosol Science, 23(Suppl. 1):329-332, 1992. \n\n[3] Hamid Bolouri et al. Design, manufacture, and evaluation of a scalable high(cid:173)\n\nperformance neural system. Electronics Letters, 30(5):426-427, 3 March 1994. \n\n[4] T G Clarkson et al. The pRAM: An adaptive VLSI chip. IEEE 'Ihmsactions \n\non Neural Networks, 4(3):408-412, May 1993. \n\n[5] Kevin N Gurney. Learning in networks of structured hypercubes. PhD thesis, \n\nDepartment of Electrical Engineering, UK, 1995. \n\n[6] Paul H Kaye et al. Airborne particle shape and size classification from spatial \n\nlight scattering profiles. Journal of Aerosol Science, 23(6):597--611, 1992. \n\n[7] R Kohlus et al. Particle shape analysis as an example of knowledge extraction \n\nby neural nets. Part. Part. Syst. Charact., 10:275-278, 1993. \n\n[8] Paul Morgan et al. Hardware implementation of a real-valued sigma-pi network. \nIn Artificial Neural Networks 5, volume 2, pages 351-356, North-Holland, 1995. \n\n[9] David E Rumelhart et al. Parallel Distributed Processing: Explorations in the \n\nMacrostructure of Cognition, volume 1. MIT Press, 1986. \n\n\fPART IX \nCONTROL \n\n\f\f", "award": [], "sourceid": 1084, "authors": [{"given_name": "Alistair", "family_name": "Ferguson", "institution": null}, {"given_name": "Theo", "family_name": "Sabisch", "institution": null}, {"given_name": "Paul", "family_name": "Kaye", "institution": null}, {"given_name": "Laurence", "family_name": "Dixon", "institution": null}, {"given_name": "Hamid", "family_name": "Bolouri", "institution": null}]}