WO1991011783A1 - Reconnaissance de motifs dans les images - Google Patents

Reconnaissance de motifs dans les images Download PDF

Info

Publication number
WO1991011783A1
WO1991011783A1 PCT/US1991/000481 US9100481W WO9111783A1 WO 1991011783 A1 WO1991011783 A1 WO 1991011783A1 US 9100481 W US9100481 W US 9100481W WO 9111783 A1 WO9111783 A1 WO 9111783A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
input
index
panel
window
Prior art date
Application number
PCT/US1991/000481
Other languages
English (en)
Inventor
Robert L. Harvey
Paul N. Dicaprio
Karl G. Heinemann
Original Assignee
Massachusetts Institute Of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Massachusetts Institute Of Technology filed Critical Massachusetts Institute Of Technology
Publication of WO1991011783A1 publication Critical patent/WO1991011783A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/24—Aligning, centring, orientation detection or correction of the image
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N15/00—Investigating characteristics of particles; Investigating permeability, pore-volume or surface-area of porous materials
    • G01N15/10—Investigating individual particles
    • G01N15/14—Optical investigation techniques, e.g. flow cytometry
    • G01N15/1429—Signal processing
    • G01N15/1433—Signal processing using image recognition
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N15/00—Investigating characteristics of particles; Investigating permeability, pore-volume or surface-area of porous materials
    • G01N15/10—Investigating individual particles
    • G01N15/14—Optical investigation techniques, e.g. flow cytometry
    • G01N15/1468—Optical investigation techniques, e.g. flow cytometry with spatial resolution of the texture or inner structure of the particle

Definitions

  • This invention relates to recognition by machines of patterns in images.
  • recognizing an object requires determining whether a certain pattern (corresponding to the object) appears within a field-of-view (FOV) of an input image.
  • the pattern generally is defined by spatial gradients and discontinuities in luminance across the input image. Other types of gradients and discontinuities may also produce perceivable patterns. Perceivable patterns may occur in the presence of:
  • An input image is here meant to include any two-dimensional, spatially ordered array of signal intensities.
  • the signals may be of any frequency within the entire electromagnetic spectrum, such as infrared radiation signals and radar ranging signals.
  • visual recognition here denotes recognition of an object based on electromagnetic radiation received from the object. Humans easily recognize spatial gray-scale object patterns regardless of the patterns' location or rotational orientation within a FOV. In perceiving these patterns, the human visual recognition system operates in two stages, first locating patterns of interest within the FOV, and then classifying the
  • Biological vision systems can rapidly segment an input image in a manner described as "preattentive.” It has been found experimentally that segmentation is context-sensitive, i.e., what is perceived as a pattern at a given location can depend on patterns at nearby locations.
  • the invention features apparatus for recognizing a pattern within an input image based on visual characteristics of the pattern, the image being represented by signals whose values correspond to the visual characteristics.
  • the apparatus includes a location channel which determines the location of the pattern within the image based on the signal values, and a classification channel which categorizes the object based on the signal values, the location channel and the classification channel
  • the location channel includes a coarse locator which makes a coarse determination of the existence and location of the pattern within the image, and a fine locator, responsive to the coarse locator, which makes a fine determination of the location of the pattern within the image.
  • the coarse locator includes a neural network which compares the image with traces corresponding to general shapes of interest.
  • the coarse locator operates with respect to a field of view within the image and a feedback path from the classification channel to the locator channel controls the position of the field of view within the image.
  • the fine locator includes circuitry for responding to feedback from the classification channel in order to adjust the position of a field of view within the image in order to
  • the coarse locator provides a feedforward signal to the fine locator which also affects the fine position of the field of view.
  • the classification channel includes a signal processer for preprocessing the signal values, a signal analyzer responsive to the signal processor for
  • the signal analyzer includes edge detectors for detecting information about edges of the pattern. Some edge detectors are adapted to generate measures of the strengths of edges in predetermined orientations within portions of the image.
  • predetermined orientations include vertical, horizontal, and 45%.
  • Other edge detectors are adapted to generate measures of the existence of edges at the periphery of a portion of the image. The edges are detected at the top, bottom, and each side of the portion of the image.
  • the signal analyzer also includes a gross size detector for detecting the gross size of a pattern within a portion of the image.
  • the measures of the visual characteristics are arrayed as a spectrum for delivery to the classifier. Measures which correspond to coarser features appear in the lower end of the spectrum and measures which
  • the signal analyzer includes a feedback path for providing the measures of the visual
  • the invention features apparatus including an orientation analyzer adapted to analyze the orientations of edges of the pattern within subwindows of the image, and a strength analyzer adapted to analyze the strengths of edges of the pattern near the periphery of a portion of a window of the image.
  • the orientation analyzer includes detectors for detecting the strengths of orientation of edges in four different possible orientations: 0, 45, 90, and 135 degrees, respectively.
  • the apparatus also includes a classifier for processing the outputs of the
  • a mapper causes outputs corresponding to subwindows of the image to be treated in the spectrum in an order such that outputs of subwindows nearer to the center of the image are treated as appearing lower on the spectrum than outputs of subwindows nearer the periphery of the image.
  • Each analyzer includes neural networks.
  • the strength analyzer includes an averaging module for averaging elements of the window to derive an averaged window, and four neural networks for processing the averaged window to determine the strength of edges at the north, south, east, and west peripheries of the window.
  • the invention features apparatus for categorizing, among a set of user-specified categories, a pattern which appears in an image based on visual characteristics of the pattern, the image being represented by signals whose values correspond to the visual characteristics.
  • the apparatus includes an unsupervi ⁇ ed classifier adapted to define classes of patterns and to categorize the patterns based on the visual features and the classes, and a supervised classifier adapted to map the classes to the set of user-specified categories.
  • the unsupervised classifier is an ART2 classifier.
  • the invention features apparatus including a location channel which determines the location of the pattern within the image based on the signal values, a classification channel which categorizes the pattern based on the signal values, and a feedback path from the classification channel to the location channel to cause the location channel to adapt to classification results generated by the classification channel.
  • the invention provides a highly effective, efficient scheme for recognizing patterns.
  • Computer processing power is devoted more heavily to portions of the image which contain possible patterns.
  • the spectrum is arranged to place relatively gross features at the lower end and relatively detailed features at the upper end which aids analysis of the relationship between features and the resulting classification.
  • Fig. 1 is a diagram of an image pattern and windows and a subwindow of the image.
  • Fig. 2 is a functional block diagram of an object recognition system.
  • Fig. 3 is a diagram of a spectrum of pattern information.
  • Fig, 4 is a diagram of edge recognition networks.
  • Fig. 5 is a table of possible outputs for example input edge patterns.
  • Fig. 6 is a diagram of the effect of window averaging.
  • Fig. 7 is a diagram of edge recognition functions.
  • Fig. 8 is a table of edge recognition network outputs.
  • FIG. 1 consider, by way of example, an image 10 consisting of a 525 by 525 array of 8-bit pixel values.
  • the pixels are arrayed along the x and y axes and the z axis represents an 8-bit luminance value of each pixel.
  • a pattern 12 representing an object to be recognized within the image is defined by a collection of 8-bit pixels. The goal is to be able to recognize quickly and accurately the existence,
  • the recognition task is performed by a visual recognition system 8 which includes a collection of modules which roughly achieve the functions of their biological counterparts in recognizing, in a selected FOV within the image, gray-scale patterns having arbitrary shifts and
  • System 8 includes a location channel 9 which locates patterns of interest in the selected FOV and a classification channel II which classifies patterns (i.e., associates a name with each pattern) located in the FOV according to known classes of objects.
  • the location channel may detect the existence of a pattern in the lower left corner of the FOV and the classifier may identify the pattern as that of the class of objects known as an automobile.
  • the classification channel is the classification channel
  • the classification channel consists of a
  • Lateral Geniculate Nucleus (LGN) module 30 which receives the input image pixel values and performs initial processing of the image.
  • Module 30 feeds three other modules: a visual area 1 (V1) module 56, a visual area 2 (V2) module 32, and a sum module 54.
  • V1 visual area 1
  • V2 visual area 2
  • V2 visual area 2
  • sum module 54 a sum module 54.
  • These three modules perform further detailed processing and generate pattern size, orientation, and location information about the image which is conceptually arrayed along a "frequency" spectrum 72.
  • the information in the spectrum is passed to an Inferior Temporal Cortex 1 (ITC1) module 58 and then to an Inferior Temporal Cortex 2 (ITC2) module 66 which classify the pattern and provide the classification results to a store 68.
  • ITC1 Inferior Temporal Cortex 1
  • ITC2 Inferior Temporal Cortex 2
  • the modules of the classification channel are also assigned numbers on Fig. 2 (such as A
  • the array of image pixels is organized into 9 windows 14a ... I4i, each containing a 175 by 175 array of pixels. Processing proceeds window by window and each window represents a FOV within the image.
  • the location channel operates on one window at a time.
  • the location channel 9 determines the location within the window presently being processed (the active window) at which any pattern lies and conveys this location to the classification channel by a 10-bit value 21.
  • the 10-bit value includes 5 bits which provide a row index and 5 bits which provide a column index for positioning the 175 by 175 bit window within the 525 by 525 input image.
  • the 10-bit value specifies the center of the 175 by 175 window within the 525 by 525 input image. Five bits give the row coordinate and five bits give the column coordinate of the 525 by 525 image. The center is given to an accuracy of + or - 9 bits.
  • the 10-bit location value and a 1-bit window enable signal 23 cause a row and column select unit 25 to indicate to LGN 30 that a pattern has been found and is located in a window whose position is specified by the 10-bit value.
  • the active window 27 i.e., the FOV
  • the pixels within the shifted window are then processed by a calibrate unit 34 and a normalize unit 36 to distribute their intensities across a gray-scale.
  • the resulting preprocessed window 37 is then sent to the later modules.
  • the calibrate unit calculates a histogram of the pixel values of the pattern within the selected (active) window.
  • the histogram is typically concentrated in a sub-band within the total possible range of 0 to 255.
  • the calibration unit spreads the histogram over the entire 0 to 255 range by linearly mapping the histogram values in the sub-band to the values 0 to 255, with the lowest value in the sub-band being mapped to 0 and the highest value in the sub-band being mapped to 255.
  • the lowest value of the histogram sub-band is defined as the value where the number of pixels falls to 1% of the cumulative number.
  • the highest value of the histogram is defined as the value where the number of pixels first exceeds 99.25% of the cumulative number.
  • the normalize unit then rescales the pixel values by dividing each of them by 255 so that all pixel values leaving the LGN module are in the range from 0 to 1. In Fig. 2, the [0,1] indicates that the values lie between 0 and 1.
  • the active window is further subdivided into 625 subwindows 42 each having an array of 7 by 7 pixels (the subwindow 42 in Fig. 1 is shown at a much larger scale than the window 38 from which it came, for clarity).
  • the window is first fed to a spiral map module 62 which performs a spiral mapping of the 625 subwindows, taking the upper left hand subwindow first (i.e., subwindow 40 of Fig. 1), then the other subwindows in the top row from left to right, then the subwindows in the right column from top to bottom, then the bottom row, left column, second row, and so on, finally ending with the center
  • the subwindows are then delivered one by one in the spiral order to the visarea 1 unit 63.
  • each 7 by 7 pixel subwindow is processed to generate measures of the visual strengths of the edges of the patterns in the horizontal
  • visarea 1 generates measures of the magnitude of the luminance gradient in the four
  • each edge measurement is performed for each 7 by 7 subwindow by a
  • Visarea 1 thus includes four neural networks 202, 204, 206, 208, each of which receives the pixels of each subwindow 57 and generates one of the outputs 210.
  • each neuron can be either an excitatory-type or an inhibitory-type, but not both simultaneously.
  • a set of actual interconnection weights useful for the four networks for the example are set forth in Appendix A.
  • Each of the detectors is a three layer neural network having an input layer, a hidden layer, and a single output neuron.
  • Appendix A includes two sets of four matrices each. One set of four matrices (marked horizontal) is used for the horizontal and vertical detectors; the other set of four matrices (marked diagonal) is used for the 45 and 135 degree detectors.
  • the four matrices A, B, C, and D contain interconnection weight values respectively for interconnections within the hidden layer,
  • each row in a matrix represents all of the interconnections from a given neuron, and each column represents all of the interconnections to a given neuron.
  • the diagonal of the A matrix thus represents all of the interconnections of hidden layer neurons with themselves.
  • detectors are designed using the genetic algorithm in the manner described in the copending patent application cited below, for a particular orientation and gradient direction.
  • the responses to orientations of 90 degrees or larger and/or gradients in the opposite sense can use the same detector weights if the input 7 by 7 subwindow are properly rotated first. The rotations are performed in visarea 1.
  • luminance lines and gradient magnitudes model similar processing that occurs in biological visual systems.
  • a technique for determining the interconnection weights for the neural network is set forth in copending United
  • gray-scale patterns of the kinds shown in column 222 would produce visarea 1 outputs as shown.
  • each line represents pixels of constant value. The indicated gradient in the pattern can be reversed without
  • object classification is done in part on the basis of these orientation strengths over a set of subwindows.
  • the four orientation signals generated by visarea 1 for each of the 625 7 by 7 pixel subwindows yields a total of 2500 orientation values for the entire window.
  • the 2500 orientation signal values 71 generated by visarea 1 can be arrayed as lines on a spectrum 72 in which the length of each horizontal line represents the magnitude of the signal.
  • the positions along the spectrum may be thought of as corresponding to different "frequencies".
  • orientation signal lines for each window are arranged in order as shown, and the successive subwindows in the spiral order are arranged in order along the spectrum so that the first subwindow's lines appear first.
  • the outer subwindows of the windowed image are nearer the top of the spectrum (lower frequency) and the inner subwindows are nearer the bottom.
  • the second feature-generating module in the classification channel is visarea 2 module 32.
  • the function of this module is to detect edges near the perimeter of the 175 by 175 pixel window. Since only the outside edges of the pattern are of interest in this step, the window image is first defocused by a 25 by 25 average unit 49.
  • this averaging smears details of the pattern (the detail is captured by visarea 1), but retains the necessary outside edge information.
  • the averaging produces a single smeared 7 by 7 pixel image 230, 230' of the pattern in the 175 by 175 window 232, 232'. As shown, the averaging simplifies the pattern edges to enable them to be easily detected.
  • visarea 2 includes four neural networks, 234, 236, 238, 240, each of which detects the presence or absence of an edge.
  • Two 3 by 7 pixel detectors 234, 236 detect the presence of nearly horizontal edges respectively at the top and bottom of the window image.
  • Two 7 by 3 pixel detectors 238, 240 detect the presence of nearly vertical edges
  • edge detectors are like the ones in visarea 1 except the input images are now 7 by 3 or 3 by 7 instead of 7 by 7.
  • Each detector uses 25 neurons with fixed interconnection weights.
  • a set of actual interconnection weights for these four neural networks are set forth in Appendix B. Only one set of four matrices is provided; these may be used in all of the four different detectors simply by rotating the input 7 by 7 subwindow by 45, 90, or 135 degrees as the case may be.
  • the output of the visarea 2 unit is four spectrum lines 45 which measure the north, south, east, and west edge
  • the four network outputs of visarea 2 are all high, while for a pattern in the lower right corner, the north and west outputs are low while the south and east outputs are high.
  • the third feature-generating module is a sum module 54. This module sums the pixel values in the 175 by 175 pixel window. The computed sum is a measure of the gross size of the pattern in the window and it is used as one of the input spectrum values to the
  • classification is achieved by interpreting a combination of the visual feature measures discussed above.
  • these feature measures include some values which have been only slightly processed (the output of the sum module), some moderately processed (the output of the visarea 2 module), and some highly processed (the output of the visarea 1 module).
  • the magnitudes of the lines are adjusted by each module to ensure appropriate comparative weighting of each module's output.
  • the visarea 1 module outputs are adjusted by subtracting the minimum (usually negative) of all of the visarea 1 outputs from each of the visarea 1 outputs to ensure that the visarea 1 portion of the spectrum is entirely positive with a minimum value of zero.
  • the visarea 2 and sum outputs are multiplied by scale factors which depend on the window size used in LGN 30 (Fig. 2). For a window size of 175 by 175, the scale factors are 0.1 for the visarea 2 outputs and 0.01 for the sum module output. For a window size of 42 by 42, the factors are 1.5 and 0.3 respectively. This weighting ensures that the
  • classifier gives equal significance to information about size, edges, and detail structure.
  • classification is done, using the spectrum 72 of information, by an unsupervised classifier 58 followed by a supervised classifier 66.
  • the unsupervised classifier ITC1 module uses the ART 2 classifier technique discussed in G.
  • ART2 is a two-slab neural network. One slab is called F1 and consists of 3 interacting layers which perform noise filtering and signal enhancement. The second slab is called F2 and consists of a single interacting
  • the F2 neurons are used to indicate by their activity the category of the input pattern.
  • the input patterns, after processing by F1 are judged to be close or far from the LTM traces. If a new input spectrum is different from previous spectra, then a new category is defined for the input. If a new input spectrum is similar to a previous category class, then the existing category is updated with an additional example.
  • the classifier is 'trained' by presenting to it a sequence of example patterns which are then categorized by ITC1. In principle, if the examples are sufficiently
  • ART2 The definition of ART2 and its operating characteristics are well-known. It is selected over other classifiers such as Hopfield nets and perceptrons because of its feature enhancement, noise reduction, and stability properties.
  • the orient unit 250 determines the closeness of the match between the input and a stored pattern based on a positive number
  • the confidence unit 252 associates the closeness measure
  • 1.0, then the confidence level is 100% and if
  • 0.7, then the confidence level is 50%, with a linear interpolation for
  • the ITC1 module After training the ITC1 module, its output nodes 61 correspond to examples of input patterns from particular categories or classes. For example, if the first ten examples are trucks, the first ten ITC1 output nodes are in a category (say category 1) that
  • the function of the location channel is to isolate an individual pattern in the FOV so that the classification channel processing can be applied to that pattern.
  • the location channel includes a Superior
  • Colliculus (superc) module 18 includes the LGN, visarea 2, and Posterior Parietal Cortex (PPC) modules.
  • the location channel supports both feedforward and feedback flows of signals.
  • Locating individual patterns within the FOV involves a two-stage process consisting of coarse location followed by fine location and
  • the superc module performs the coarse location procedure.
  • a modified ART2 neural network is used to grossly locate objects of interest within the FOV.
  • the F2 slab of the ART2 is used to impress a stored LTM trace on the top layer of the F1 slab. LTM traces for the general shapes of interest are computed off-line and stored in the superc.
  • the system is 'primed' to locate a particular class of objects.
  • a 175 by 175 pixel window is extracted from the input image and impressed on the bottom layer of the ART2.
  • the pattern specified by the LTM trace 19 is compared to the windowed image.
  • the LTM trace is designed so that an object of the correct general size will cause a match, even if off-center, to indicate its presence.
  • a row map unit 24 is used to map the windowed input to the ART2 input. Because the input window is 175 by 175, there are 30,625 input pixels delivered to the ART2. If no match is found, then another
  • non-overlapping window in the image is input as the active window and evaluated for the presence of an object.
  • the degree of match between the image pattern and the LTM traces is used as an enable signal 23 to the LGN module.
  • the selection of the coarse window position from among the nine possible windows is done by a fovea move unit 20.
  • the coarse position 22 is sent to the row map unit, and to the PPC module for further adjustment.
  • the second stage of the location process is the fine adjustment and pull-in stage.
  • This pull-in stage is done by a feedback path which includes the LGN, visarea 2, and PPC modules.
  • the function of the LGN and visarea 2 modules was described above.
  • the center of attention, or fovea i.e., the location of the center of the active window
  • fovea is adjusted to center the window on the pattern of interest.
  • the object 12 is not centered in any of the nine original windows of the image.
  • the object pattern is made to lie in the center of the window as shown by reference numeral 50.
  • the centering function evaluates the outputs of visarea 2, i.e., the strength of the four edges of the window, which are sent to PPC on lines 81.
  • the fovea delta 1 unit 46 in the PPC implements the control law for moving the window.
  • One possible control law is a standard bang-bang rule with a
  • the window is moved a fixed amount vertically, up or down depending on the sign of the difference. For example, if north - south is positive and larger than the positive threshold, then the window is moved
  • the output of the fovea delta 1 box is the magnitude of adjustment for the location in the vertical and horizontal directions, and is fed to the fovea adjust unit 83.
  • the fovea adjust unit adjusts the value provided by the fovea move unit 20 and delivers the current location values in the horizontal and vertical directions on line 21. Adjustments may be made one pixel at a time in either direction.
  • a second pull-in path includes the LGN, visarea 2, ITC1, ITC2, and PPC modules. This path is used to take additional looks at an object when the confidence in pattern identification is low. If the confidence level is judged to be insufficient, then an enable signal 99 from ITC2 activates a fovea delta2 unit 68 in PPC. This unit generates a random adjustment of the window in the vertical and horizontal directions. This random adjustment gives the system a second chance to achieve a better pattern classification.
  • a counter in ITC2 (not shown) is used to limit the number of
  • the system After some preset number of retries, the system stores the object's conjectured identity together with the confidence level and location, and then goes on to search for other objects.
  • a slew enable signal 101 is used to activate the fovea move unit 20 to move to the next coarse position, i.e., to the next one of the nine windows in the original image.
  • the system functions are executed in a sequential manner.
  • the location channel finds and centers in a window an object of interest.
  • the choice between which window will be used for the analysis depends on numerical runoff errors and appears random to the user.
  • the classification channel identifies the object.
  • the modules would run simultaneously.
  • the sequencing of functions would be controlled by enable signals, as described above, and by properly selecting the neural network interconnection time constants. Time constants associated with the location channel's LTMs are short so that the channel will converge quickly to the location which is to be analyzed.
  • classification channel's LTM time constants are longer and the identification process is comparatively slow. This difference in the time constants ensures that classification is done on a centered object. Possible time constants would be such that the ratio of location time to classification time would be from 1:3 up to 1:10 or more. The exact time would depend on the nature of the application including the size of the input images, and grayness.
  • OBJS image util.o vfilter.o median. o IRP_histogram.o ⁇
  • IRP_edge_detector.o IRP_visar2.o IRP_LGN.o ART2.0 verrtool.o ALL_OBJS $(OBJS)
  • LIBS -lm -lsuntool -lsunwindow -lpixrect -lstd
  • ALL_LIBS $(LIBS)
  • cellview.o cellview. c cellview. h netparam.h ⁇ activation.h image_io.h LTM.h image_util .o: image_util.c image_io.h
  • verrtool.o verrtool.c cellview.h
  • vfilter.o vfilter.c cellview. h netparam.h ⁇ image_io.h activation. h
  • IRP_histogram.o IRP_histograB.c cellview.
  • IRP_edge_detector .o IRP edge_detector .c cellview.
  • IRP_LGN .o IRP _LGN.c image_io.h activation.h ART2.O: ART2.c activation. h LTM.h
  • HST_HEIGHT 64 #deflne HST_HEIGHT 64 /* Plot amplitude for largest peak */ #define PLOT_BORDER_WIDTH 16 /* Width of blank border around plot */ #define HST_WIN_WIDTH (VLT_SIZE + (2 * PLOT_BORDER_WIDTH))
  • EDF_WIN _WIDTH EDF_DISPLAY _WIDTH + (2 * PLOT_BORDER_WIDTH)
  • EDF_PLOT _ WIDTH ( EDr_SPECTRUM SIZE + (2 * PLOT_BORDER_WIDTH)
  • N_EDF_ P LOT_PIXELS (EDF_PLOT _WIDTH * HST _PLOT _HEIGHT)
  • V1 First location fo edge detector (V1) signals in the spectrum array */ #define EDF_SPECTRUM_OFFSET ( V2_SPECTRUM_OFFSET + V2_SPECTRUM_SIZE)
  • Use information in SunView include files to define a structure for accessing the base frame's color map */
  • BOX_STRUCT (int size_x, size_y, x0, y0, x1, y1, x, y;);
  • setup_windows (argc, argv, "IRP Interactive Image Analysis Software Testbed”); setup_ img_ menu();
  • window_create (base_frame, CANVAS,
  • window_create (base_frame , CANVAS ,
  • window_set (img_canvas, WIN_BELOW, vlt_canvas, 0);
  • window_set (hst_canvas, WIN_Y, top_y, 0);
  • window_set (edf_canvas, WIN_Y, top_y+HST WlN_HEIGHT+48, 0);
  • window_set ( edf_hdr_canvas, WIN_Y, top_y+HST_WIN_HEIGHT+23, 0);
  • hst_pw canvas_pixwin(hst_canvas);
  • edf_pw canvas_pixwin(edf_canvas);
  • edf_hdr pw canvas_pixwin(edf_ hdr_canvas);
  • V2_ wt 0.100
  • scroll_bar_thickness (int) scrollbar_get(SCROLLBAR, SCROLL_THICKNESS);
  • ART2_ hdr_ item panel_create_item(ART2_ panel, PANEL_ TEXT,
  • LGN_mult_ item panel_ create_item(ART2_panel, PANEL_TEXT,
  • V2_mult_item panel_create_item(ART2_panel, PANEL_TEXT,
  • LTM_input_ item panel_create_ item(ART2_ panel, PANEL_TEXT,
  • PANEL_LABEL_STRING "LTM input file:”, PANEL_VALUE, old_LTM_file, PANEL_VALUE_ DISPLAY_LENGTH, FNL,
  • LTM_output_item panel_create_ item(ART2 panel, PANEL_TEXT,
  • PANEE_LABEL STRING "LTM save file:”
  • PANEL_VALUE new_LTM_file
  • PANEL_VALUE_DISPLAY_LENGTH FNL
  • window_set (ART2_panel, WIN_ X, pnl_ x, 0);
  • strncpy (LGN_mstr, (char *)panel_ get_value( LGN_mult_item), NUM_STR_ LEN); sscanf (LGN_mstr, "#f", &LGN_wt);
  • strncpy V2_mstr, (char *)panel_ get_value(V2_mult_ item), NUM_STR_LEN; sscanf (V2_mstr, "#f", &V2_ wt);
  • LTM source_file fopen(old LTM_ file, "r");
  • err_str (char *)calloc(num_cFar, sizeof (char));
  • LTM_ output_ file fopen(new_ LTM_ file, "w” );
  • err_str (char * )calloc(num_ char, sizeof ( char));
  • num_vals 2 * TOT_SPECTRUM_SIZE * nF2;
  • err_str (char *)calloc(num_ char, sizeof (char)); strcpy(err_ str, "Problem reading LTM trace values from file ⁇ **); strcat(err_str, old_LTM_ file);
  • num_cha r 52 + s t r len ( new_LTM_ f i le ) ;
  • num_vals 2 * TOT SPECTRUM_ SIZE * nF2;
  • err_str (char *)calloc(num_cRar, sizeof (char)); strcpy(err_str, "Problem writing LTM trace values to file ⁇ ""); strcat ( err_str, new_LTM_file);
  • color_fetch_index bas_cms.cms_size
  • V2_hidden_layer 1 NULL
  • V2_hidden_layer NULL
  • pw_writebackground (hst_pw, 0, 24, hist_width, HST_PLOT_HEIGHT, PIX_SRC);
  • window_set (base_ frame, FRAME_NO_CONFIRM, TRUE, 0);
  • strncpy seq_ cwd, (char *)panel_get_ value(batch_cwd_ item), FNL
  • base_ frame window_ create(NULL, FRAME, FRAME LABEL, frlabstr,
  • num_ Item panel_ create_ item(control_ panel, PANEL TEXT,
  • hdr_item panel_ create_ item(control_panel, PANEL TEXT,
  • npnl_rows + aux_file_info( );
  • csr_item panel_create_item(control panel, PANEL TEXT,
  • PANEL_LABEL_STRING "Cursor position:”
  • PANEL_ VALUE_ DISPLAY_ LENGTH FNL
  • zoom_item panel_create_item(control_panel, PANEL TEXT,
  • PANEL_LABEL_STRING "Current Zoom Factors:”
  • PANEL_ VALUE_DISPLAY_ LENGTH FNL
  • img_box_item panel_create_item(control_panel, PANEL_TEXT,
  • PANEL_LABEL_STRING "Image box:”, PANEL_VALUE, box_str,

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

Un motif dans une image est reconnu en se basant sur des caractéristiques visuelles du motif, l'image étant représentée par des signaux dont les valeurs correspondent aux caractéristiques visuelles, en utilisant un canal de localisation (4) qui détermine la position du motif dans l'image en se basant sur les valeurs des signaux, et un canal de classification (11) qui classe par catégorie le motif en se basant sur les valeurs des signaux, le canal de localisation et le canal de classification (11) fonctionnant en parallèle et en coopération pour reconnaître le motif. Dans d'autres aspects, les orientations des bords du motif à l'intérieur des sous-fenêtres de l'image sont analysées tout comme le sont les densités du bord des motifs à proximité de la périphérie des portions de l'image; un classificateur non supervisé définit des classes d'objets de représentation interne, et un classificateur supervisé (66) répertorie les classes en catégories spécifiques aux utilisateurs; un circuit de rétro-action est prévu du canal de classification (11) au canal de localisation (9).
PCT/US1991/000481 1990-01-23 1991-01-23 Reconnaissance de motifs dans les images WO1991011783A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US46868190A 1990-01-23 1990-01-23
US468,681 1990-01-23

Publications (1)

Publication Number Publication Date
WO1991011783A1 true WO1991011783A1 (fr) 1991-08-08

Family

ID=23860799

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US1991/000481 WO1991011783A1 (fr) 1990-01-23 1991-01-23 Reconnaissance de motifs dans les images

Country Status (1)

Country Link
WO (1) WO1991011783A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115598025A (zh) * 2022-12-13 2023-01-13 四川亿欣新材料有限公司（Cn） 图像处理方法及使用该方法的碳酸钙粉质检系统
CN117649633A (zh) * 2024-01-30 2024-03-05 武汉纺织大学 一种用于高速公路巡检的路面坑洼检测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US3643215A (en) * 1967-11-15 1972-02-15 Emi Ltd A pattern recognition device in which allowance is made for pattern errors
US4242662A (en) * 1978-10-16 1980-12-30 Nippon Telegraph And Telephone Public Corporation Method and apparatus for pattern examination
US4685143A (en) * 1985-03-21 1987-08-04 Texas Instruments Incorporated Method and apparatus for detecting edge spectral features
US4965725A (en) * 1988-04-08 1990-10-23 Nueromedical Systems, Inc. Neural network based automated cytological specimen classification system and method
US4972499A (en) * 1988-03-29 1990-11-20 Kabushiki Kaisha Toshiba Pattern recognition apparatus
US4996593A (en) * 1988-12-01 1991-02-26 Westinghouse Electric Corp. A method of and apparatus for comparing the location of an extended feature within a field of view with a standard location

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US3643215A (en) * 1967-11-15 1972-02-15 Emi Ltd A pattern recognition device in which allowance is made for pattern errors
US4242662A (en) * 1978-10-16 1980-12-30 Nippon Telegraph And Telephone Public Corporation Method and apparatus for pattern examination
US4685143A (en) * 1985-03-21 1987-08-04 Texas Instruments Incorporated Method and apparatus for detecting edge spectral features
US4972499A (en) * 1988-03-29 1990-11-20 Kabushiki Kaisha Toshiba Pattern recognition apparatus
US4965725A (en) * 1988-04-08 1990-10-23 Nueromedical Systems, Inc. Neural network based automated cytological specimen classification system and method
US4965725B1 (en) * 1988-04-08 1996-05-07 Neuromedical Systems Inc Neural network based automated cytological specimen classification system and method
US4996593A (en) * 1988-12-01 1991-02-26 Westinghouse Electric Corp. A method of and apparatus for comparing the location of an extended feature within a field of view with a standard location

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115598025A (zh) * 2022-12-13 2023-01-13 四川亿欣新材料有限公司（Cn） 图像处理方法及使用该方法的碳酸钙粉质检系统
CN115598025B (zh) * 2022-12-13 2023-03-10 四川亿欣新材料有限公司 图像处理方法及使用该方法的碳酸钙粉质检系统
CN117649633A (zh) * 2024-01-30 2024-03-05 武汉纺织大学 一种用于高速公路巡检的路面坑洼检测方法
CN117649633B (zh) * 2024-01-30 2024-04-26 武汉纺织大学 一种用于高速公路巡检的路面坑洼检测方法

Similar Documents

Publication Publication Date Title
US5313532A (en) Recognition of patterns in images
Mardia et al. A spatial thresholding method for image segmentation
Rowe et al. Statistical mosaics for tracking
CN108830842B (zh) 一种基于角点检测的医学图像处理方法
CN109508636A (zh) 车辆属性识别方法、装置、存储介质及电子设备
CN113870157A (zh) 一种基于CycleGAN的SAR图像合成方法
CN113419623A (zh) 一种非标定眼动交互方法和装置
CN107423771B (zh) 一种两时相遥感图像变化检测方法
CN117557784B (zh) 目标检测方法、装置、电子设备及存储介质
Bock et al. Gray-scale ALIAS
WO1991011783A1 (fr) Reconnaissance de motifs dans les images
Akhmetshina et al. Improvement of Grayscale Images in Orthogonal Basis of the Type-2 Membership Function.
CN112580442B (zh) 一种基于多维金字塔层次模型的行为识别方法
WO1991014235A1 (fr) Reconnaissance de configurations dans des images
CN108288041A (zh) 一种行人目标错检去除的预处理方法
Tistarelli Computation of coherent optical flow by using multiple constraints
Du et al. A color projection for fast generic target tracking
Várkonyi-Kóczy New advances in digital image processing
Basavaraj et al. Real Time Object Distance and Dimension Measurement using Deep Learning and OpenCV
CN110136173A (zh) 一种目标定位处理方法及装置
Margner et al. On benchmarking of document analysis systems
Ahmad et al. Fuzzy Based Hybrid Focus Value Estimation for Multi Focus Image Fusion.
Silberberg Multiresolution aerial image interpretation
Doshi et al. Monochromatic Image Dehazing Using Enhanced Feature Extraction Techniques in Deep Learning
Scott Applied machine vision

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): CA JP

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH DE DK ES FR GB GR IT LU NL SE

COP Corrected version of pamphlet

Free format text: PAGES 1/8-8/8,DRAWINGS,REPLACED BY NEW PAGES 1/9-9/9;DUE TO LATE TRANSMITTAL BY THE RECEIVING OFFICE

NENP Non-entry into the national phase

Ref country code: CA

122 Ep: pct application non-entry in european phase