Japanese
Welcome to UWICL!
University of Washington Image Computing Library on Texas Instruments
TMS320C80 (MVP).
What is UWICL?
-
Goals of UWICL.
-
Architecture of UWICL.
-
UWICL 3.0 function list and performance figures.
-
New on Release 3.1.
-
Benefits of the UWICL consortium.
-
The consortium.
-
How to join the consortium? (fees and licence)
-
History of UWICL.
-
The UWICL Team
-
The UWICL guides (design, style...) ICSL access only.
-
Miscellaneous Information
Copyright © 1995 Image Computing Systems Laboratory, University of Washington
cdeforge@icsl.ee.washington.edu
The library provides a portable infrastructure of low-level C80 routines so that higher level algorithms and applications can be developed and quickly integrated on the target C80-based system with minimal custom coding by using the the library routi
nes as building blocks. A highly modularized design is used for all the functions supported by the UWICL. The UWICL Design Guide documents this architecture in detail while the UWICL Style Guides specify the coding standards. Also, the Quality Assurance G
uide establishes the QA program and bug handling procedures for the UWICL.
Each function has
its own README file in which we describe how to make the file, how to use the simulator, what are the restrictions, what are the performance on MediaStation 5000 (GSP5). It has also a UNIX
MAN page including the algorithm specifics, synopsis, server to MP parameters, MP to PP-C parameters, path, author, bug report. We provide also a server.c and a Script file to test the function on GSP5 hardware.
Top
Next
The goal in developing the UWICL is to provide a portable and efficient library of core image computing algorithms to the MVP user community. The UWICL project attempts to provide many low-level image processing functions. The long-term goal is to pr
ovide an infrastructure of low-level MVP routines so that higher-level algorithms and applications can be developed and quickly integrated on the MVP using the UWICL routines and minimal custom code. This way, a system integrator or application programmer
can mix and match different data flows with different image processing functions in various architectures based on the MVP(s). Currently, UWICL is actively expanding, upgrading and supporting the library to facilitate the widespread use of the C80.
Top
Next
A UWICL function is to be divided into 3 hierarchical segments, the MP-module, the PP-C module, and the PP tight-loop module. The overall program structure is shown in Figure 1. Such a modular design provides the UWICL user with a very flexible archi
tecture. The user may choose to use the functions provided by the UWICL by calling any of these hierarchical modules depending on the required application and on the user's familiarity with MVP programming. A novice user may choose to use a packaged MP mo
dule which uses all the MVP resources in a predetermined way and with little modification in C language apply it to his application. A more experienced user who would like to customize the use of some of the MVP resources may choose to use the PP-C module
only. Still, a more advanced user of the MVP who understands the intricacies of the MVP's data transfer protocol and PP programming may choose to use only the PP tight-loop module or even modify it to optimize the PP tight-loop further. This approach wil
l give the greatest flexibility in the usage of the MVP's resources, but the greatest knowledge and understanding of the MVP is required.
Top
Next
Arithmetic
Offset 3.5 ms Scale 5.5 ms Normalize 6.4 ms
Floor 3.6 ms Clip 3.6 ms Invert 3.5 ms
Addition(8) 5.4 ms Addition(16) 17.2 ms Signed Addition(8) 7.5 ms
Subtraction(8) 5.4 ms Subtraction(16) 17.2 ms Signed Subtraction(8) 7.5 ms
Multiplication 7.5 ms Division 16.3 ms Absolute Value 3.6 ms
Absolute Difference 5.4 ms Right Shift 3.5 ms Left Shift 3.6 ms
Maximum Image 5.2 ms
Matrix Mult. 1.1 ms
Matrix x Vector 1.1 ms ( matrix 256 x 256, vector 256 with 32 or 64-bit output. Fixed point reprentation of floats )
Logical
AND 5.4 ms OR 5.4 ms XOR 5.4 ms
Morphology
Binary:
Erosion 12.7 ms Dilation 12.7 ms Opening 21.2 ms
Closing 21.2 ms
Grayscale:
Erosion 32.7 ms Dilation 32.7 ms Opening 65.2 ms
Closing 65.2 ms
Spatial Filters
Median 10.7 ms Sobel 18.9 ms
Gradient 28.3 ms Laplacian 16.1 ms
Unsharp masking 19.4 ms Correlation 19.4 ms
8-Bit Convolution
2D Convolution(NxN) 19.4 ms (112.1 ms for 9x9)
2D Signed Convolution(NxN) 19.4 ms (112.1 ms for 9x9)
2D Convolution(MxN) 19.4 ms (112.1 ms for 9x9)
2D Padded Convolution(MxN) 20.0 ms (114.6 ms for 9x9)
Boxcar Filter 21.4 ms ( 46.1 ms for 9x9)
16-Bit Convolution
2D convolution 20.7 ms (16 bit output)
Signal Processing
FIR Filters 13.0 ms
Transforms
FFT/IFFT (16-bit) 85.0 ms FFT/IFFT (32-bit, 256x256) 35.0 ms
General FFT/IFFT (16-bit) 95.0 ms Real FFT/IFFT (16-bit) 50.0 ms
FFT Shift/Inverse FFT Shift 14.3 ms General Freq. Domain Filter 200.0 ms
DCT/IDCT (8x8 blocks) 9.5 ms Frequency-Domain Filters 187.5 ms
Wavelets (D4) 18.1 ms Wavelets (D6) 26.2 ms
Compression
Full-Search VQ Enccode 574.5 ms VQ Decode 9.3 ms
Huffman Encode 24.4 ms Huffman Decode 40.5 ms
Run-Length Encode 12.1 ms Run-Length Decode 2.3 ms
Histograms
Equalize 8.3 ms Stretch 9.0 ms
Normalize 6.4 ms Histogram Generate 4.3 ms
Auto Threshold(8) 0.5 ms Auto Threshold(12) 5.8 ms
Geometric Manipulation
Shrink Image (by 2) 2.2 ms Arbitrary Shrink 3.0 ms (for 128x128 output)
Arbitrary Magnify (1 factor) 20.9 ms Arbitrary Magnify (2 factors) 21.0 ms
Zoom Image (by 2) 9.3 ms Translate Image 9.4 ms (VRAM To VRAM)
90 Degree Rotate 3.5 ms Flip Image 3.5 ms
Interpolate 2 Images (8-bit) 5.3 ms
Orthogonal Slicing of Volume 16.1 ms (256x256xz256)
Volume Rotation by 90 deg. 265.0 ms (256x256xz256)
Warping
General Warp 34.3 ms (with 4 control points on 8-bit image)
41.9 ms (with 4 control points on 16-bit image)
56.6 ms (color with 4 control points)
Perspective Warp 46.2 ms ( 8-bit image, bilinear interpolation)
46.3 ms (16-bit image, bilinear interpolation)
37.1 ms ( 8-bit image, nearest neighbor interpolation)
37.2 ms (16-bit image, nearest neighbor interpolation)
Affine Warp 14.0 ms (on 8-bit image)
14.8 ms (on 16-bit image)
26.8 ms (color)
Rotate 29.0 ms (on 8-bit image)
36.3 ms (on 16-bit image)
50.8 ms (color)
Statistics
Mean 1.8 ms Variance 3.8 ms
Moments (1, 2, 3) 3.8 ms Contrast 4.4 ms
Entropy 5.9 ms Image Correlation 105.4 ms (with 9x9)
Mean square error 3.5 ms Count Pixels of Gray Level 3.3 ms
Min/Max Gray Level Search 2.0 ms
Image Statistics ( 1st, 2nd moment, bounding box) 4.2 ms, 3.2 ms, 4.9 ms
Features
Gray Level Difference stat 14.6 ms Gray Level Co-occurrence stat 78.4 ms
Hurst 5.8 ms Range 5.8 ms
Projection Histogram 8.9 ms
Template Matching 115.4 ms (for 9x9 kernel)
Template Matching 283.0 ms (for 9x9 kernel, with normalized correlation coefficient)
Graphics
Line Drawing 18.0 ms Drawing Filled Rectangles 4.4 ms
Filled Circle Drawing 2.8 ms Drawing Filled Polygons 5.3 ms
Meshplot 23.6 ms
Nota bene: The graphics timing represent the drawing of several objects.
Image Registration
with Coincident Bit Counting 47.1 ms (with 100x100 ROI)
with Absolute Difference Criterion 1079.0 ms
with rotation and Absolute Difference 1246.0 ms
with Outlier Minimal Number Estimator 68.3 ms (with 100x100 ROI)
with full search 75.0 s ( two 512 x 512 images)
with multiresolution full search 3954.8 ms ( two 512 x 512 images)
with multiresolution translate and rotate 569.0 ms ( two 512 x 512 images)
Color Image Processing
Color Conversions :
RGB to HSI 15.1 ms HSI to RGB 7.6 ms (720 x 200 x 32-bit image)
kYCbCr to RGB 10.4 ms Color Format Conversion 4.3 ms
Pack YUV -> YUYV 9.5 ms Unpack YUYV -> YUV 7.5 ms
32-bit RGB -> 16-bit 565 RGB 5.9 ms 16-bit 565 RGB -> 32-bit RGB 5.7 ms
32-bit RGB -> 16-bit 555 RGB 5.9 ms 16-bit 555 RGB -> 32-bit RGB 5.7 ms
Color Geometric Manipulation :
Affine Warp of Color Image 26.8 ms Rotate RGB Image 50.9 ms
Affine Mapping of RGB image 50.9 ms General Warp of RGB Image 56.6 ms (4 control points)
Perspective Warp of RGB image 67.6 ms
Color Segmentation :
Color Format Conversion 4.3 ms
HSV Color Segmentation 23.4 ms (on 700 x 200 x 32-bit)
Other Image Processing
Contouring 12.0 ms Connected Components 35.0 ms
Distance Transform 40.0 ms Canny's Edge Detection 152.5 ms
Arithmetic Encoder 110.0 ms Arithmetic Decoder 200.0 ms
Image Composition 7.3 ms YUV to (X) RGB 10.4 ms (for 720x240)
Image Threshold 3.6 ms
Segmentation with Snakes 250.0 ms
Pack 2.2 ms Unpack 2.0 ms
Quadratic Facet Model 51.9 ms (produces 6 images)
Radial Search 2.4 ms (on a 110 x 128 area 16-bit)
Convex Hull 3.5 ms
All Numbers are based on a 50 Mhz MVP and
2 cycles/column DRAM with 8-bit 512 x 512 Images (Unless Specified).
Top
Next
Release 3.1 of the UW Image Computing Library was made available on
October 17, 1997. There are now a total of 194 functions contained in the
library, including twenty new functions in the current release. The new
functions and key features in this release include:
- 20 new functions, including improved generalized warping,
image padding, more 16 bit functions (unary and logical functions, threshold),
10-bit histogram, 8-bit look-up table, text printing, and color conversion.
- Granularity restrictions have been removed in more functions as of the
current release. These functions include threshold, mean, mean-square-error,
moments, entropy, contrast, count8, and min_max8.
- Improvements to the convolution functions to support up to 25x25
sized kernels.
- Inclusion of more convolution kernel data files for different filters and
different sizes.
- Improved generalized warping to support color images and different sized
interpolation kernels.
- Inclusion of a Visual Basic program to generate the spatial transformations
required by the generalized warping function.
- Inclusion of software and documentation on making UWICL compatible with
big-endian C80 boards in the Windows NT operating system environment.
Top
Next
Fast implementation using the UWICL routines as building blocks. Programmers do not need to understand the details of the DSP architecture to be able to maximize the DSP's processing power.
Easy customized development
- Multi-level hierarchical organization.
- Standardized design and styles.
- High-level interaction.
University of Washington
- Innovation in imaging algorithms and systems since 1984.
- Experience in MVP programming and MVP based systems since 1990.
- Independemce and stability.
Top
Next
The Consortium Advisory Board (CAB) provides advice regarding the choice of
algorithms, recommends areas of interest, and gives feedback on the overall Consortium operation. Each Company will have one vote. The CAB meeting wil be held twice a year at the University of Washington.
- Bianual (April and October) expansion and upgrade
- Consortium will cease its operation on Dec. 31, 1999
- Consortium Director and Contact Person:
- Dr. Yongmin Kim
- Director of the UWICL Consortium
- Department of Electrical Engineering
- University of Washington, Box 352500
- Seattle, WA 98195-2500 U.S.A.
- (206) 685-2271 FAX: (206) 543-3842 Email: kim@ee.washington.edu
Top
Next
There is an Annual Sponsorship Fee of $20,000 to support the Consortium. The Company gets the software site license to use the UWICL source code right away in R&D.
When the company is ready to introduce products using the UWICL, they have two solutions for the Commercial License:
- Either paid up non-exclusive license: $80,000 one time payment OR
- Royalty bearing non-exclusive license:
- 1 to 1000: $30
- 1001 to 10000: $10
- More than 10000: $4
The sooner the company joins, the more advantageous the licensing terms become. For every year Company has been a member of the Consortium before a license agreement is executed, there is 20% reduction in paid-up license fee or royalty. The maximum r
eduction is 80%.
Top
Next
June 1994, the ICSL started the project UWICL.
January 1995, the pre-release of the UWICL is ready.
March 1995, release 1.0 with about 60 functions.
August 1995, pre-release 1.1 of UWICL with improved functions, more
flexibility.
September 1995, UWICL successfully hosts a conference for the TMS320C80 Developers.
October 1995, release 1.1 of UWICL with about 94 functions. In the release 1.1 the functions are more robust, handle a very large variety of image sizes and have be through a tough Quality Assurance program. An online support has been created.
Top
Copyright © 1995 Image Computing Systems Laboratory, University of Washington
cdeforge@icsl.ee.washington.edu
Questions?