Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
116 views

Real Time Image Processing

The document discusses using Python for real-time image processing for robot vision. It describes using Pygame for frame grabbing and basic image processing. NumPy arrays can be used for faster operations but Cython is recommended for compiled speed. Benchmarking shows Cython code runs 10x faster than Python for thresholding images. Key steps include converting images to NumPy arrays, thresholding values, finding connected components to detect objects, and using Cython type declarations.

Uploaded by

Anik Purwa
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
116 views

Real Time Image Processing

The document discusses using Python for real-time image processing for robot vision. It describes using Pygame for frame grabbing and basic image processing. NumPy arrays can be used for faster operations but Cython is recommended for compiled speed. Benchmarking shows Cython code runs 10x faster than Python for thresholding images. Key steps include converting images to NumPy arrays, thresholding values, finding connected components to detect objects, and using Cython type declarations.

Uploaded by

Anik Purwa
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 32

Real-Time Image Processing in Python

Gabriel J. Ferrer Hendrix College

The Situation

Using netbook to control a robot

Supports research and teaching in computer vision Rapid prototyping of algorithms Easy frame-grabbing with the pygame library Notoriously bad performance in tight loops Use Cython and Numpy

Why Python?

Why not Python?

Solution

Just how slow is Python, really?

Benchmarking web site

http://shootout.alioth.debian.org C is about 50 times faster than Python

Median Python vs. C

Robot Vision Processing

Goal

Use image sequences to guide robot behavior Acquire an image Transform it into a data structure Select a robot action based on data structure Rate of action selection is bounded by rate of image acquisition and transformation

Process

Performance consideration

Real-Time Image Processing

Real-time systems

Correctness of code depends on whether deadlines are met Efficiency is helpful Only necessary for meeting a deadline Implies a soft deadline for the image computations Ideal is 10 frames/second Performance degrades below this point

Need for prompt action selection by the robot


Performance issues in image processing

Images are arrays


Must visit every array element Need fast array access Need fast looping Image subtraction Edge detection Color matching Connected components

Typical operations

Pygame Library

As of version 1.9, includes frame grabbing Includes several important modules:

pygame.display

Handles rendering to a window Represents an image Grabs images from a camera

pygame.surface

pygame.camera

Initialization
import pygame import pygame.camera from pygame.locals import * pygame.init() pygame.camera.init() size = (640, 480)

Setting up
d = pygame.display.set_mode(size, 0) s = pygame.surface.Surface(size, 0, d) c = pygame.camera.list_cameras() cam = pygame.camera.Camera(c[0], size) cam.start() going = True

Main Loop
while going: if cam.query_image(): s = cam.get_image(s) d.blit(s, (0, 0)) pygame.display.flip() for e in pygame.event.get(): if e.type == QUIT: cam.stop() going = False

Processing Images

Pygame surfarray library

Converts pygame surfaces to numpy arrays High-speed n-dimensional arrays (ndarray) All elements have the same data type Tight two-dimensional loop Each inner loop iteration involves a type check!

Numpy (1.3)

Why is the same data type important?


Detecting Moving Objects

For each frame:


Convert image to an array Subtract the previous array Find the non-zero regions of nontrivial size

Applying numpy (1)

Add before the start of the loop: last_array = None diffs = None s = pygame.surface.Surface(size)

Applying numpy (2)

Inner loop, if statement body:


s = cam.get_image(s) s2d = pygame.surfarray.array2d(s) diffs = s2d if last_array != None: diffs = s2d last_array last_array = s2d pygame.surfarray.blit_array(s, diffs)

Problem

Excessive background noise Solution: Hue, Saturation, Value (HSV)


Hue: The type of a color Saturation: The strength of a color Value: The whiteness of a color

Disregard H and S; just use V

Extracting the Value

Each color is 24 bits:


Bits 23-16 are Red (RGB) or Hue (HSV) Bits 15-8 are Green (RGB) or Saturation (HSV) Bits 7-0 are Blue (RGB) or Value (HSV)

Mask all but the lower 8 bits to get V NB: Display is still RGB

The V will look blue

Code Alterations

When initializing cam: cam = pygame.camera.Camera(c[0], size, HSV)

Immediately after creating s2d: s2d = numpy.bitwise_and(s2d, 0xFF)

Finding the Blobs

Connected components (blobs)

Islands in an image that share a characteristic First, threshold the image

Blob finding:

Useful pixels will be high values Returns a list of the blobs

Second, find the blobs

Implementing Blob Finding

Useful variables to initialize at the start: b = (0, 0, 0xFF) r = (0xFF, 0, 0) t = (0x5A, 0xAA, 0xAA)

Additional Code

Inside the if statement, after blit_array:


m = pygame.mask.from_threshold(s, b, t) for blob in m.connected_components(10): coord = blob.centroid() pygame.draw.circle(s, r, coord, 50, 5)

Image Shrinking

Often an effective technique to boost frame rate Into our original program, insert at the top: shrunken = (320, 240) Then replace the blit() call with: p = pygame.transform.scale(s, size) d.blit(p, (0, 0))

Writing Custom Routines


Numpy is very nice, but it doesn't do everything Basic threshold function: def threshold(img, value, hi, lo): for i in range(img.shape[0]) for j in range(img.shape[1]): if img[i,j] < value: img[i,j] = lo else: img[i,j] = hi

Cython

Compiles Python programs to C

From C, compiles to binary object code Still very experimental Will compile any Python program For best results, augment with type declarations

Using version 0.11

Superset of Python

Cython Initialization
cimport numpy cimport cython @cython.boundscheck(False) @cython.wraparound(False)

Cython Type Declarations


def threshold (numpy.ndarray[numpy.int32_t, ndim=2] img, numpy.int32_t value, numpy.int32_t hi, numpy.int32_t lo): cdef Py_ssize_t i, j

Performance Difference

Interpreted Python

0.85 frames/second 10.81 frames/second

Cython with type declarations

Creating a setup.py script

Program name: filters.pyx


from distutils.core import setup from distutils.extension import Extension from Cython.Distutils import build_ext ext_modules = [Extension(filters, [filters.pyx])] setup(name = 'Img proc filters', cmdclass = {'build_ext': build_ext}, ext_modules = ext_modules)

Compiling the program


python setup.py build_ext --inplace

Generated C Code

Lots of setup code at function start


Checks expected vs. actual arguments Creates lots of temporary variables Use cdef to minimize this cdef functions are not callable from Python

Inefficient function calls inside tight loops


Code is otherwise a direct translation into C Numpy arrays are not C arrays

Array accesses use a macro for pointer arithmetic

Generated C Code
for (__pyx_t_1 = 0; __pyx_t_1 < (__pyx_v_img->dimensions[0]); __pyx_t_1+=1) { __pyx_v_i = __pyx_t_1; for (__pyx_t_2 = 0; __pyx_t_2 < (__pyx_v_img->dimensions[1]); __pyx_t_2+=1) { __pyx_v_j = __pyx_t_2; __pyx_t_3 = __pyx_v_i; __pyx_t_4 = __pyx_v_j; if (((*__Pyx_BufPtrStrided2d(__pyx_t_5numpy_int32_t *, __pyx_bstruct_img.b uf, __pyx_t_3, __pyx_bstride_0_img, __pyx_t_4, __pyx_bstride_1_img)) < __pyx_v_v alue)) { __pyx_1 = __pyx_v_lo; } else { __pyx_1 = __pyx_v_hi; } __pyx_t_5 = __pyx_v_i; __pyx_t_6 = __pyx_v_j; *__Pyx_BufPtrStrided2d(__pyx_t_5numpy_int32_t *, __pyx_bstruct_img.buf, __ pyx_t_5, __pyx_bstride_0_img, __pyx_t_6, __pyx_bstride_1_img) = __pyx_1; } }

Conclusion

You can do robot vision in Python! Pygame


Frame grabbing Some image processing High-performance arrays Matrix arithmetic Compilation; high-performance object code

numpy

cython

Resources

http://www.pygame.org http://www.cython.org http://wiki.cython.org/tutorials/numpy http://ozark.hendrix.edu/~ferrer/presentations/


These slides Sample code

You might also like