Computer_Vision

Software · Fri Oct 04 04:00:00 EDT 2002

OSTI ID:1230670

Justin Doak, LANL

The Computer_Vision software performs object recognition using a novel multi-scale characterization and matching algorithm. To understand the multi-scale characterization and matching software, it is first necessary to understand some details of the Computer Vision (CV) Project. This project has focused on providing algorithms and software that provide an end-to-end toolset for image processing applications. At a high-level, this end-to-end toolset focuses on 7 coy steps. The first steps are geometric transformations. 1) Image Segmentation. This step essentially classifies pixels in foe input image as either being of interest or not of interest. We have also used GENIE segmentation output for this Image Segmentation step. 2 Contour Extraction (patent submitted). This takes the output of Step I and extracts contours for the blobs consisting of pixels of interest. 3) Constrained Delaunay Triangulation. This is a well-known geometric transformation that creates triangles inside the contours. 4 Chordal Axis Transform (CAT) . This patented geometric transformation takes the triangulation output from Step 3 and creates a concise and accurate structural representation of a contour. From the CAT, we create a linguistic string, with associated metrical information, that provides a detailed structural representation of a contour. 5.) Normalization. This takes an attributed linguistic string output from Step 4 and balances it. This ensures that the linguistic representation accurately represents the major sections of the contour. Steps 6 and 7 are implemented by the multi-scale characterization and matching software. 6) Multi scale Characterization. This takes as input the attributed linguistic string output from Normalization. Rules from a context free grammar are applied in reverse to create a tree-like representation for each contour. For example, one of the grammar’s rules is L -> (LL ). When an (LL) is seen in a string, a parent node is created that points to the four child symbols ‘(‘ , ‘L’ , ‘L’, and ‘)‘ . Levels in the tree can then be thought of as coarser (towards the root) or finer (towards the leaves) representations of the same contours. 7.) Multi scale Matching. Having a multi-scale characterization allows us to compare objects at a coarser level before matching at finer levels of detail. Matching at a coarser level not only increases the speed of the matching process (you’re comparing fewer symbols) , but also increases accuracy since small variations along contours do not significantly detract from two objects’ similarity.

• Short Name / Acronym: CV; 001662MLTPL00
• Site Accession Number: LANL Copyright No. C-02, 039,
• Version: 00
• Programming Language(s): Medium: X; OS: No OS requirements.
• Research Organization: Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
• Sponsoring Organization: USDOE
• DOE Contract Number: W-7405-ENG-36 with DOE
• OSTI ID: 1230670
• Country of Origin: United States