Computer Vision for Industrial Applications PDF

Computer Vision for Industrial Applications APPS  Artificial Vision § Vision is one of the most powerful senses of human beings From vision, we extract an extremely high informative content, contactless § The knowledge of biological vision systems is still fragme...

Computer Vision for Industrial Applications APPS  Artificial Vision § Vision is one of the most powerful senses of human beings From vision, we extract an extremely high informative content, contactless § The knowledge of biological vision systems is still fragmented § Excellent results in industrial applications, where tasks are well defined and the environment is in some way structured § Open problems in the realization of vision systems with human features APPS 2  Artificial Vision Artificial vision in robotics is defined as the process of extraction, characterization and interpretation of information coming from images of a tridimensional world APPS 3 Tasks of a Vision System § A vision system, essentially, takes images as input and provides their description § The concept of description is broad, and can encompass potentially infinite levels of detail Fundamental requirements: – It must have a relationship with the input image (in turn, with the 3D object – It must contain all the information useful to accomplish a given task – It must operate a compression of the information APPS 4  Affine Paradigms § Image processing § Pattern classification § Scene analysis APPS 5 Hierarchical Organization § Perception is the process that provides a computer image § Pre-processing deals with noise reduction and detail improvement § Segmentation divides the image in objects of interest § Description computes characteristics (such as dimensions, shapes) that are useful to differentiate one object from another § Recognition is the process that identifies such objects § Interpretation gives a meaning to the recognized objects APPS 6 PERCEPTION APPS 7  Perception § VIDICON tube camera APPS 8  Perception § CCD sensors APPS 9  Standard video § CCIR 625 lines per frame - 25 frames per second (Europe, Australia) § RS170 525 lines per frame - 30 frames per second (USA, Japan) APPS 10 Creation of a digital image APPS 11  Digital Image Digitalization of spatial coordinates (x, y)= image sampling Digitalization of amplitudes = intensity or grey level quantization (min. two levels => binary image) APPS 12 Example: image sampling 256 x 256 128 x 128 64 x 64 32 x 32 APPS 13 Example: image quantization 32 16 8 4 APPS 14 Lightning techniques APPS 15 Perspective transformation APPS 16 Matrix of perspective transformation APPS 17  Inverse transformation Inverse transformation wh = P-1ch APPS 18 Indetermination in the inverse transformation Let us assume that a point in the image has coordinate (x0, y0, 0), where the «0» in z coordinate indicates that we are in the image plane z = 0. The image point with coordinates (x0, y0) corresponds to the set of the 3D points aligned along a line that joins (x0, y0, 0) and (0, 0,l) APPS 19  Solution The reconstruction of a 3D point from a bidimensional image requires the knowledge of at least one of the coordinates of the point in the 3D system APPS 20 Camera model § These equations (and their inverse) characterize the formation of an image through the projecton of 3D points on the camera plane § This model assumes that the camera and the world coordinate systems coincide § In reality, these systems can be distinct APPS 21 Camera model (1) Displacement w0 of the origin of the camera reference system; (2) Pan of x axis, (3) Tilt of z axis; (4) Displacement r of the image plane with respect to the center of the joint, on which the camera is mounted and around which can be rotated. APPS 22  Camera model (1) Displacement of the origin of the camera reference system; (2) Pan of x axis, (3) Tilt of z axis; (4) Displacement of the image plane with respect to the center of the joint, on which the camera is mounted and around which can be rotated APPS 23 Camera model Perspective Transformation Total transformation ch = PCRGwh APPS 24 Camera calibration ch = Awh A = PCRG APPS 25  Camera calibration we omit the formula for ch3 since it corresponds to the z axis The calibration procedures is carried out as follows: Obtain m ³ 6 reference points of known coordinates (Xi, Yi, Zi) with i = 1, 2, 3, …, m (there are two equations that comprise the coordinates of two points, thus, we need at least six points to solve the equation system) Acquire the representation of the points in the camera plane, obtaining the corresponding points with coordinates (xi, yi), i = 1, 2, 3, …, m Use the data derived above to solve the system and compute the unknown coefficients APPS 26 Stereoscopic Vision APPS 27 Stereoscopic vision APPS 28 PRE-PROCESSING APPS 29  Filtering § Average of the neighborhood § Median filtering § Average of multiple images § Filtering of binary images APPS 30  Neighborhood average § Simple technique in the spatial domain § Given an image f(x,y), we generate a filtered image g(x,y) where the intensity of each pixel is obtained through the average of the intensities of pixels of f in a pretedermined neighborood of (x,y) § The filtering effect is strongly influenced by the neighborhood size APPS 31  Median filtering § The neighborhood average tends to blur margins and sharp contours (useful for segmentation) § We can mitigate this inconvenient using the median value instead of the average one § The median value tends to force pixels with very different intensities to be more similar to their neighboring ones, eliminating isolated transients § Higher computational effort (requires sorting) APPS 32  Filtering examples Original Altered Image image Average of neighborhood Median (5x5) filtering (5x5) APPS 33  Stencils Some important preprocessing operations can be executed through the application of so- called stencils) They are small bidimensional sets, whose coefficients are selected to reveal a given property or feature in an image APPS 34  Stencils § Highlight isolated -1 -1 -1 points (with intensity -1 8 -1 very different from -1 -1 -1 the background) § Neighborhood 1/9 1/9 1/9 average 1/9 1/9 1/9 1/9 1/9 1/9 APPS 35 Average of multiple images § Let us consider a noisy image g(x,y) characterized by an additive noise n(x,y) to a nominal (unaltered) image f(x,y). Let us consider the nois as uncorrelated and with zero average § Let us construct a new image acquiring K frames from the same scene and let us perform a point-by-point average APPS 36 Average of multiple images § We obtain § This implies that we will get closer to the nominal image as long as K increases § We should make sure that we can approximate the scene as still during the acquisition of the K frames APPS 37  Enhancement § Main difficulty in low-level vision algorithms: adapting automatically to changes in lightning in the scene § Solution: enhancement techniques § They must be low cost (hardware and computational) – Histogram equalization – Enhancement based on local properties APPS 38  Image histogram § Given an image, we can collect the intensity values of its pixels as realizations of a random variable § Let us assume for the moment that these values are continuous and the image points are infinite § Under these assumptions, we will have a continuous random variable, to which we can associate a probability density function APPS 39 Probability density function § The probability density function p(r) tells us that the probability that a realization of the random variable is between r* and r*+δr tends to p(r*)δr when δr tends to 0 § This implies that the probability that r