Breaking News: You Can Count On Us Moderator: Byron Jacobs Matthews Sejeso Micheal Olusanya Jordan Masakuna Thato Mashigo Nkosi Nathi Industry Representative: Dario Punucchi 1 / 26 Introduction Problem Statement: Cracking the mystery To design an image processing tool that can localize and count cells in a fracture pattern. To better understand the mechanics of how such fracture patterns form, partially in order to aid the tool in the point above, but also to develop new insights that may lead to new quality control procedures. 2 / 26 Methods Methods: It’s not all black and white Segmentation of the image is the main goal of the work. Watershed is a commonly used segmentation algorithm that is modelled on distributing water across the image and segmenting the image according to the pools that form at local minima. Binarization is desirable to obtain a schematic representation of the image. However, the image cells may not be ‘sealed’ entirely making segmentation algorithms perform poorly. 3 / 26 Methods Binarization Technique This byrinization technique combines a diffusion equation, to smooth noise and outliers, with a cubic source term with a binarizing effect. The model is given by ∂u = cd ∇2 u + cs u(1 − u)(u − a) (1) ∂t where a is a threshold parameter calculated based on local statistical properties of the image. The coefficients of diffusion and the source term can be balanced to show preference to each of the processes. The model is subjected to Neumann boundary conditions along the edges and the initial state is given by the input image, u(x, y , 0) = Image(x, y ). This is then discretized using an explicit finite difference scheme and implemented on a massively parallel GPU architecture. 4 / 26 Numerical Experiments and Results Binarization Figure : Input image of shattered glass. 5 / 26 Numerical Experiments and Results Binarization Figure : Binarized image of shattered glass. 6 / 26 Numerical Experiments and Results Binarization Figure : Input image of shattered glass. 7 / 26 Numerical Experiments and Results Binarization Figure : Binarized image of shattered glass. 8 / 26 Numerical Experiments and Results Delauny Mesh Generation The shattered glass is very similar to Voronoi diagrams. We can use a corner detection to isolate the intersections between the cells and then fit a Delauny Mesh to these points, since Delauny meshes are the duality of the Voronoi diagram. 9 / 26 Numerical Experiments and Results Binarization Figure : Delauny Mesh Connected to Corners. 10 / 26 Numerical Experiments and Results A Morphological Approach This approach uses a morphological opening on the binarized image. Then the connected morphological components are computed and their centres are found. Each component is taken to be a glass cell. 11 / 26 Numerical Experiments and Results Binarization Figure : Morphological Components. 12 / 26 Numerical Experiments and Results Delauny Mesh overestimates 410 cells, where length times breadth estimates is 240. Potential upper and lower bound. Morphological components count 168 cells in agreement with results obtained later. 13 / 26 Numerical Experiments and Results Filtering Gaussian Filter Gaussian filter smooths an image by computing the weighted averages in a filter box. g (x, y ) = x2 + y2 1 exp − where (x, y ) is position of each pixel. 2πσ 2 2σ 2 Median Filter Median filter replaces each entry by the median of neighboring entries. 14 / 26 Numerical Experiments and Results Seperation of Background and Foreground Otsu’s algorithm If g (x, y ) is a thresholded version of f (x, y ) of a threshold t then : ( 1 if f (x, y ) ≥ t g (x, y ) = 0 otherwise Minimize intraclass variance. Maximize interclass variance. t−1 n−1 X X wb (t) = p(i) and wa (t) = p(i) i=0 2 σin (t) = wb (t)σb2 (t) i=t + wa (t)σa2 (t) 2 σout (t) = wa (t)[µb (t) − µ]2 + wa (t)[µa (t) − µ]2 15 / 26 Numerical Experiments and Results Otsu’s Threshold Algorithm Compute histogram and probabilities of each intensity level (pixel) Initialize wi (0) and µi (0) Step through all possible threshold t = 1, · · · , maximum intensity. 2 (t) and σ 2 (t). Then update wi and µi and compute σin out 2 (t) and to the Desired threshold corresponds to the maximum of σout 2 minimum of σin (t) Compute two different optimal thresholds a, b desired threshold is d = a+b 2 16 / 26 Numerical Experiments and Results Segmentation: The Random Walker Algorithm This algorithm computes the probability that each node sends a random walker to the seeds. The image is represented by a graph G = (V , E ), where vi ∈ V (set of the pixels) and eij = (vi , vj ) ∈ E . The edge Gaussian weighting function is given by : (gi − gj )2 wij = exp − σ where gi is the image intensity at node vi and σ is the standard deviation With xi real-valued variable associated with each node in G , the Laplacian matrix L which is formed by V , E and wij , the random walker tries to optimize the following energy : Q(x) = x t Lx = X wij (xi − xj )2 and L = D − A = (`i,j )n×n eij 17 / 26 Numerical Experiments and Results Segmentation: Ambling Arbitrarily Where `i,j deg(vi ) if i = j := −1 if i 6= j and vi is adjacent to vj 0 otherwise where deg (vi ) is degree of the vertex i. with F and B the sets of foreground and background seeds, the optimization is constrained by xi = 1 for vi ∈ F and xi = 0 for vi ∈ B. 18 / 26 Numerical Experiments and Results Counting: What’s on the tally? Since our image is already labelled (segmented), it is easy to count the number of labels (cells). To get the centroid of each region. Given a sequence of masses and points like mi and (xi , yi ) for i = 1, · · · , p mx = p X mi × xi and my = i=1 p X mi × yi i=1 m= p X mi i=1 And then the centroid will be : (x̂, ŷ ) = mx my , m m 19 / 26 Numerical Experiments and Results Figure : Image Processing : Tempered Glass 20 / 26 Numerical Experiments and Results Supperpixels approach What is superpixel? ... Superpixels is a group of pixels which have similar characteristics. Graph based algorithms - Each pixels is treated as a node in the graph and the edges represent the similarity between the pixels. Gradient-ascent based algorithms - Iteratively uses the gradient ascent methods to refine the clusters until convergence. 21 / 26 Numerical Experiments and Results Superpixels SLIC Algorithm... SLIC is a simple and efficient method to decompose an image into visually homogeneous regions. 1 SLIC performs a local clustering of pixels in 5-D space defined by the L, a, b values of the CIELAB colorspace and x, y coordinates of the pixels. 2 It has a different distance measurement which enables compactness and regularity in the superpixel shapes, Ds = dlab + 3 m dxy . S SLIC generates superpixels by clustering pixels based on their color similarity and proximity in the image plane. 22 / 26 Numerical Experiments and Results Superpixel Results The segmentation traces edges of a shattered glass. 23 / 26 Numerical Experiments and Results Superpixel approach Further work... Merge superpixels which are positioned within the region of the shattered glass. Heuristic Approach Bayesian Approach 24 / 26 Conclusion Conclusion and Further Suggestions: We cracked it! Physical models: Learning algorithm to learn the initial stress tensor that minimizes final result when compared with binary image. Segmentation by weighted aggregation: A multiscale segmentation approach that coarsens a fine set of segments abiding by the image edges. Superpixels: Adaptive merging of cells to fit the glass cells. Image guided fracture generation: Using the binary image as an underlying guide on which to generate fracture. 25 / 26 Conclusion Thank you!! Questions? and Comments... 26 / 26 References S. S. Gleason, K. W. Tobin, Directional dilation for the connection of piece-wise objects: A semiconductor manufacturing case study, in: Image Processing, 1996. Proceedings., International Conference on, Vol. 3, IEEE, 1996, pp. 9–12. Rashmi, Mukesh, Kumar and Rohini Saxena . Algorithm and technique on various edge detection : A survey. Signal and Image Processing : An international journal (SIPIJ) vol4, No 3, June 2013. Leo Grady. Random walks for image segmentation. IEEE Transactions and pattern analysis and Machine Intelligence, vol 28, No 11, Pp 1768-1783. November 2006. Bryan S. Morse. Lecture 4 : Thresholding. Brigham Young University 1998-2000 26 / 26
© Copyright 2026 Paperzz