<strong>IEEE</strong> COMSOC MMTC R-LetterExtending Signal Processing Techniques to Graph DomainA short review for “Perfect Reconstruction Two-Channel Wavelet Filter Banks for Graph Structured Data”Edited by Jun ZhouSunil K. Narang and Antonio Ortega, “Perfect Reconstruction Two-Channel Wavelet FilterBanks for Graph Structured Data”, <strong>IEEE</strong> Transactions on Signal Processing, Vol. 60, No. 6,pages 2786-2799, 2012.Graph theory has been successfully adopted inmany computer vision and pattern recognitionapplications. When dealing with large scale data,one <strong>of</strong> the problems that hinders the wideadoption <strong>of</strong> graphical models is the very highcomputational complexity caused by largenumber <strong>of</strong> nodes and vertices in graph. Toaddress this challenge, one would expect thatonly a few nodes in the graph be used to form acompacted representation <strong>of</strong> the original graph.Then data processing can be performed only on asmall neighborhood <strong>of</strong> each node. Some recentefforts in this direction have explored traditionalsignal processing techniques, such as wavelettransform, as possible solutions.The paper published by Narang and Ortega in<strong>IEEE</strong> TSP is a seminal work on graph samplingand design <strong>of</strong> critically sampled wavelet filterbanks on graphs. It not only provides acomprehensive review <strong>of</strong> the spatial/spectralrepresentation <strong>of</strong> graph signals and existing workon two-channel filter banks, but also proposesthe important characteristics <strong>of</strong> the samplingstrategy and filter banks for perfectreconstruction <strong>of</strong> bipartite graphs.The key idea behind this method is applying atwo-channel filter banks that decompose a graphinto high-pass and low-pass channels, eachcontaining only part <strong>of</strong> the nodes in the graphafter downsampling and following upsamplingoperations. When these two channels arecombined, they form a perfect reconstruction <strong>of</strong>the original graph representation. In order toachieve such distortion-free reconstruction, analiasing component, which is composed <strong>of</strong> filterbanks and downsampling functions, shall be setto zero. Therefore, the goal <strong>of</strong> this research is t<strong>of</strong>igure out what are the proper filter banks anddownsampling functions to meet the aboverequirement.To develop the downsampleing strategy, theauthors proposed that the decomposed high-passthe low-pass channels shall contain complementnode sets <strong>of</strong> the original graph. This leads to thebuilding <strong>of</strong> a bipartition <strong>of</strong> the graph nodes [1].Based on the graph spectral theory, this strategygenerates spectral coefficients at symmetricgraph frequencies around a central frequency,which is equivalent to the aliasing component <strong>of</strong>the reconstruction function.To design the filter banks, the authors pointedout that they shall meet three conditions, i.e.,aliasing cancellation, perfect reconstruction, andorthogonality. Therefore, a quadrature mirrorfilter bank method [2] (wavelet is one <strong>of</strong> suchmethod) was chosen and extended to bipartitegraph. This method allows a single basis spectralkernel be created, while all other kernels are builton top <strong>of</strong> the basis kernel.Whilst it is straightforward to adopt the waveletfilter banks on bipartite graph, the application <strong>of</strong>this framework to arbitrary graph requiresgenerating a series <strong>of</strong> bipartite subgraphs fromthe original graph. Then each subgraph can beprocessed independently with a cascadedtransform being implemented at the end. In thispaper, the authors proposed to use the biparticitymethod from Harary et al [3] for subgraphgeneration.Two experiments have been performed todemonstrate how the proposed method can beapplied to image processing (as an example <strong>of</strong>regular graph) and traffic graph analysis (as anexample <strong>of</strong> irregular graph). These examplesshow that the two-channel wavelet filter banksand the sampling method form a practicalsolution for graph decomposition andreconstruction. It enables efficient graphcomputation, which has been expected by theresearch community. I believe this work willgenerate long-term impact to the development <strong>of</strong>graph theory because it provides an elegant wayhttp://committees.comsoc.org/mmc 8/22 Vol.4, No.4, <strong>August</strong> <strong>2013</strong>
<strong>IEEE</strong> COMSOC MMTC R-Letter<strong>of</strong> applying signal processing techniques to solvestructured pattern recognition problems.References:[2] S. Narang and A. Ortega, “Downsamplinggraphs using spectral theory,” Proceedings<strong>of</strong> the International Conference onAcoustics, Speech and Signal Processing,pages 4208-4211, 2011.[3] J. Johnston. “A filter family designed foruse in quadrature mirror filter banks,”Proceedings <strong>of</strong> the <strong>IEEE</strong> InternationalConference on Acoustics, Speech andSignal, pages 291-294, 1980.[4] F. Harary, D. Hsu, and Z.Miller, “Thebiparticity <strong>of</strong> a graph,” Journal <strong>of</strong> GraphTheory, vol. 1, no. 2, pp. 131–133, 1977.Jun Zhou received the B.S.degree in computer scienceand the B.E. degree ininternational business fromNanjing University <strong>of</strong>Science and Technology,China, in 1996 and 1998,respectively. He receivedthe M.S. degree in computerscience from ConcordiaUniversity, Canada, in 2002,and the Ph.D. degree incomputing science from University <strong>of</strong> Alberta, Canada,in 2006.He joined the School <strong>of</strong> Information andCommunication Technology in Griffith University asa lecturer in June 2012. Prior to this appointment, hehad been a research fellow in the Australian NationalUniversity, and a researcher at NICTA. His researchinterests are in statistical pattern recognition,interactive computer vision, and their applications tohyperspectral imaging and environmental informatics.http://committees.comsoc.org/mmc 9/22 Vol.4, No.4, <strong>August</strong> <strong>2013</strong>