I am trying to identify connected regions of pixels in an image stack. Since it is a stack, the input is quite large (on the order of 10 million pixels, although only about 1 million are bright), and looping through it pixel by pixel is extremely slow. Nonetheless, my first attempt works as follows:
1) Check if current pixel is bright, if so go to 2. If not go to 4.
2) Obtain slice of image containing all 26 neighbors. Set all neighbors that have not yet been scanned to be dark. Set all neighbors that fall outside of the image (e.g. indices of -1) to be dark. Set current pixel to dark.
3) If all pixels in slice are dark, assign current pixel a new label. If exactly one pixel in slice is bright, set current pixel label to be equal to this bright pixel's label. If more than one pixel is bright, set current pixel to lowest of bright pixel labels, and record the equivalence of these labels.
4) Advance one row. Once all rows in the column are scanned, advance one column. Once every pixel is in the slice is scanned, advance one slice in the stack. Once the entire image is scanned, advance to 5.
5) Loop through the list of equivalencies, reassigning labels in the output.
I don't think the processing in steps 2 and 3 are much slower than other, similar algorithms, but I could be wrong. I think the major bottleneck is simply looping through each pixel. I know the bottleneck is not #5 because I have never been patient enough to reach that step. Even just looping over the image indices with a print statement is quite slow. I know that with the same data I can do this analysis quickly, so I suspect there must be a smarter way to do this with a very different approach. I have read something vague about performing dilation and using intersections, but I can't imagine what the mechanism they were alluding to was.
Is there something fast and clever out there, or is this basically the only way to do it? How is the MATLAB algorithm so much faster?
for i in xrange(0,nPix):
for j in xrange(0,nPix):
for k in xrange(0,nFrames):
if img[i,j,k] != 0: #If the current pixel is bright
#Construct an array of already scanned
#neighbors. Pixels are scanned first in z, then y,
#then x.
#Construct indices allowing for periodic boundary
ir = np.array(range(i-1,i+2)).reshape(-1,1,1)%nPix
jr = np.array(range(j-1,j+2)).reshape(1,-1,1)%nPix
kr = np.array(range(k-1,k+2)).reshape(1,1,-1)%nFrames
pastNeighbors = img[ir,jr,kr] #Includes i-1:i+1
#Set all indices outside image and "future neighbors" to 0
pastNeighbors[2,:,:] = 0
pastNeighbors[1,2,:] = 0
pastNeighbors[1,1,2] = 0
pastNeighbors[1,1,1] = 0 #Pixel itself; not a neighbor
#Eliminate periodic boundary
if i-1 < 0: pastNeighbors[0,:,:] = 0
if i+1 == nPix: pastNeighbors[2,:,:] = 0
if j-1 < 0: pastNeighbors[:,0,:] = 0
if j+1 == nPix: pastNeighbors[:,2,:] = 0
if k-1 < 0: pastNeighbors[:,:,0] = 0
if k+1 == nFrames: pastNeighbors[:,:,2] = 0
if np.count_nonzero(pastNeighbors) == 0: #If all dark neighbors
label = label + 1 #Assign new label to pixel
out[i,j,k] = label
elif np.count_nonzero(pastNeighbors) == 1: #One bright neighbor
relX, relY, relZ = np.nonzero(pastNeighbors)
relX = relX - 1
relY = relY - 1
relZ = relZ - 1
out[i,j,k] = out[i + relX, j + relY, k + relZ]
else: #Mutliple bright neighbors
relX, relY, relZ = np.nonzero(pastNeighbors)
relX = relX - 1
relY = relY - 1
relZ = relZ - 1
equiv = out[i + relX, j + relY, k + relZ]
out[i, j, k] = np.min(equiv) #Assign one of the
#multiple labels
for i in xrange(0,equiv.size):
equivList = np.append(equivList, [[np.min(equiv),equiv[i]]], axis = 0)
Related
I have an image (saved as a variable called canny_image) and it looks like this after preprocessing.
I am basically trying to find the distance between the first two vertical lines. I tried using the hough_line function from skimage, but it's unable to find the first line, so I thought it might be easier to solve this manually.
I am basically trying to solve this by going through each row in the image until I get to the first pixel with a value of 255, (the lines have a value of 255, while everything else is zero), and then I store the location of that pixel in an array. And I take the mode of the values in the array as the x location of the first line. I'll do the same for the 2nd line by using the first x-value as a starting point.
def find_lines(canny_image):
threshold = 255
for y in range(canny_image.shape[0]):
for x in range(canny_image.shape[1]):
if canny_image[x, y] == threshold:
return x
This is the code I wrote to get the x-location of the first line, however, I'm not getting the desired output. Any help on how to solve this will be much appreciated. Thanks!
Perhaps try something like this
# Returns an array of lines x positions in an image
def find_line_x_positions(image, lines_to_detect: int, buffer_zone: int):
threshold = 255
(height, width) = image.shape
x_position_sums = np.zeros((lines_to_detect, 2), np.double) # For each line, store x_pos sum[i,0] and point count[i,1]
for y in range(height):
buffer = 0
line_index = 0
for x in range(width):
if buffer > 0:
buffer -= 1
if (image[y, x] >= threshold) and (buffer == 0):
buffer = buffer_zone
x_position_sums[line_index, 0] += x
x_position_sums[line_index, 1] += 1
line_index += 1
if ((line_index) == lines_to_detect):
break
# Divide the x position sums by the point counts to get the average x position for each line
results = x_position_sums[np.all(x_position_sums,axis=1)]
results = np.divide(results[:,0],results[:,1])
return results
You can also try OpenCV's HoughLines() function which is simpler to implement than scikit lib. When I tested OpenCV implementation out, it seems to have a hard time finding vertical lines(within 10 degrees from vertical) but you can solve this by rotating your image X degrees and look for lines within that range of rotation.
I have a big fits file I'm working with in astropy, and I'm randomising 100 sets of coordinates to work with (calculating sky noise), but because the fits file I'm working with is an image from a telescope, the image is skewed. Because of this, large sections of the file are just NaN values. I want to avoid the areas with NaNs so I can only analyse areas with actual pixel values. How do I do this in the loop I've written?
N = 100
x = np.random.uniform(60, og.shape[0]-60, N).astype(int) # random co-rdinate within a boarder
y = np.random.uniform(60, og.shape[1]-60, N).astype(int)
n = np.zeros(N)
def dntp(x, y): # 'og' is my big fits file
box = og[x-30:x+30, y-31:y+30] # creating a box the same size as mask
pht = mask*box
mjy = ((photfnu*pht)/et)*1e6 # converting to microjanksies
return mjy
for i in range(N):
if np.sum(np.isnan(og[x[i]])) == 0 & np.sum(np.isnan(og[y[i]])) == 0:
n[i] = np.sum(dntp(x[i],y[i]))
When looking at the values of n the loop gives me, I get all 0 values, any ideas?
I can compute the SLIC boundaries using skimage as follows:
def compute_superpixels(frame, num_pixels=100, std=5, iter_max=10,
connectivity=False, compactness=10.0):
return slic(frame, n_segments=num_pixels, sigma=std, max_iter=iter_max,
enforce_connectivity=connectivity, compactness=compactness)
Now, what I would like to do is get the index of pixels which form the boundary of each label. So my idea was to get all pixels belonging to a given segment and then check which pixels have a change in all two directions
def boundary_pixels(segments, index):
# Get all pixels having a given index
x, y = np.where(segments == index)
right = x + 1
# check we are in bounds
right_mask = right < segments.shape[0]
down = y + 1
down_mask = down < segments.shape[1]
left = x - 1
left_mask = left >= 0
up = y - 1
up_mask = up >= 0
neighbors_1 = np.union1d(right_n, down_n)
neighbors_2 = np.union1d(left_n, up_n)
neighbors = np.union1d(neighbors_1, neighbors_2)
# Not neighbours to ourselves
neighbors = np.delete(neighbors, np.where(neighbors == i))
However, with this all I managed to do was to get the neighbours in the 4 directions of a given label. Can someone suggest some way to actually get all pixels on the border of the label.
I found an answer to my own question. The mark_boundaries in the skimage.segmentation package does exactly what I needed.
Usage:
processed = mark_boundaries(frame, segments==some_segment)
Here frame is he current image frame and segments is the label array. some_segment is the label integer index whose boundaries we are interested in.
You can make use of the find_contours function available in skimage.measure module to find the co-ordinates of the pixels along the boundary. An example is available at find_contours.. Next, you can change for change in both directions as needed.
I have some code for calculating missing values in an image, based on neighbouring values in a 2D circular window. It also uses the values from one or more temporally-adjacent images at the same locations (i.e. the same 2D window shifted in the 3rd dimension).
For each position that is missing, I need to calculate the value based not necessarily on all the values available in the whole window, but only on the spatially-nearest n cells that do have values (in both images / Z-axis positions), where n is some value less than the total number of cells in the 2D window.
At the minute, it's much quicker to calculate for everything in the window, because my means of sorting to get the nearest n cells with data is the slowest part of the function as it has to be repeated each time even though the distances in terms of window coordinates do not change. I'm not sure this is necessary and feel I must be able to get the sorted distances once, and then mask those in the process of only selecting available cells.
Here's my code for selecting the data to use within a window of the gap cell location:
# radius will in reality be ~100
radius = 2
y,x = np.ogrid[-radius:radius+1, -radius:radius+1]
dist = np.sqrt(x**2 + y**2)
circle_template = dist > radius
# this will in reality be a very large 3 dimensional array
# representing daily images with some gaps, indicated by 0s
dataStack = np.zeros((2,5,5))
dataStack[1] = (np.random.random(25) * 100).reshape(dist.shape)
dataStack[0] = (np.random.random(25) * 100).reshape(dist.shape)
testdata = dataStack[1]
alternatedata = dataStack[0]
random_gap_locations = (np.random.random(25) * 30).reshape(dist.shape) > testdata
testdata[random_gap_locations] = 0
testdata[radius, radius] = 0
# in reality we will go through every gap (zero) location in the data
# for each image and for each gap use slicing to get a window of
# size (radius*2+1, radius*2+1) around it from each image, with the
# gap being at the centre i.e.
# testgaplocation = [radius, radius]
# and the variables testdata, alternatedata below will refer to these
# slices
locations_to_exclude = np.logical_or(circle_template, np.logical_or
(testdata==0, alternatedata==0))
# the places that are inside the circular mask and where both images
# have data
locations_to_include = ~locations_to_exclude
number_available = np.count_nonzero(locations_to_include)
# we only want to do the interpolation calculations from the nearest n
# locations that have data available, n will be ~100 in reality
number_required = 3
available_distances = dist[locations_to_include]
available_data = testdata[locations_to_include]
available_alternates = alternatedata[locations_to_include]
if number_available > number_required:
# In this case we need to find the closest number_required of elements, based
# on distances recorded in dist, from available_data and available_alternates
# Having to repeat this argsort for each gap cell calculation is slow and feels
# like it should be avoidable
sortedDistanceIndices = available_distances.argsort(kind = 'mergesort',axis=None)
requiredIndices = sortedDistanceIndices[0:number_required]
selected_data = np.take(available_data, requiredIndices)
selected_alternates = np.take(available_alternates , requiredIndices)
else:
# we just use available_data and available_alternates as they are...
# now do stuff with the selected data to calculate a value for the gap cell
This works, but over half of the total time of the function is taken in the argsort of the masked spatial distance data. (~900uS of a total 1.4mS - and this function will be running tens of billions of times, so this is an important difference!)
I am sure that I must be able to just do this argsort once outside of the function, when the spatial distance window is originally set up, and then include those sort indices in the masking, to get the first howManyToCalculate indices without having to re-do the sort. The answer might involve putting the various bits that we are extracting from, into a record array - but I can't figure out how, if so. Can anyone see how I can make this part of the process more efficient?
So you want to do the sorting outside of the loop:
sorted_dist_idcs = dist.argsort(kind='mergesort', axis=None)
Then using some variables from the original code, this is what I could come up with, though it still feels like a major round-trip..
loc_to_incl_sorted = locations_to_include.take(sorted_dist_idcs)
sorted_dist_idcs_to_incl = sorted_dist_idcs[loc_to_incl_sorted]
required_idcs = sorted_dist_idcs_to_incl[:number_required]
selected_data = testdata.take(required_idcs)
selected_alternates = alternatedata.take(required_idcs)
Note the required_idcs refer to locations in the testdata and not available_data as in the original code. And this snippet I used take for the purpose of conveniently indexing the flattened array.
#moarningsun - thanks for the comment and answer. These got me on the right track, but don't quite work for me when the gap is < radius from the edge of the data: in this case I use a window around the gap cell which is "trimmed" to the data bounds. In this situation the indices reflect the "full" window and thus can't be used to select cells from the bounded window.
Unfortunately I edited that part of my code out when I clarified the original question but it's turned out to be relevant.
I've realised now that if you use argsort again on the output of argsort then you get ranks; i.e. the position that each item would have when the overall array was sorted. We can safely mask these and then take the smallest number_required of them (and do this on a structured array to get the corresponding data at the same time).
This implies another sort within the loop, but in fact we can use partitioning rather than a full sort, because all we need is the smallest num_required items. If num_required is substantially less than the number of data items then this is much faster than doing the argsort.
For example with num_required = 80 and num_available = 15000 the full argsort takes ~900µs whereas argpartition followed by index and slice to get the first 80 takes ~110µs. We still need to do the argsort to get the ranks at the outset (rather than just partitioning based on distance) in order to get the stability of the mergesort, and thus get the "right one" when distance is not unique.
My code as shown below now runs in ~610uS on real data, including the actual calculations that aren't shown here. I'm happy with that now, but there seem to be several other apparently minor factors that can have an influence on the runtime that's hard to understand.
For example putting the circle_template in the structured array along with dist, ranks, and another field not shown here, doubles the runtime of the overall function (even if we don't access circle_template in the loop!). Even worse, using np.partition on the structured array with order=['ranks'] increases the overall function runtime by almost two orders of magnitude vs using np.argpartition as shown below!
# radius will in reality be ~100
radius = 2
y,x = np.ogrid[-radius:radius+1, -radius:radius+1]
dist = np.sqrt(x**2 + y**2)
circle_template = dist > radius
ranks = dist.argsort(axis=None,kind='mergesort').argsort().reshape(dist.shape)
diam = radius * 2 + 1
# putting circle_template in this array too doubles overall function runtime!
fullWindowArray = np.zeros((diam,diam),dtype=[('ranks',ranks.dtype.str),
('thisdata',dayDataStack.dtype.str),
('alternatedata',dayDataStack.dtype.str),
('dist',spatialDist.dtype.str)])
fullWindowArray['ranks'] = ranks
fullWindowArray['dist'] = dist
# this will in reality be a very large 3 dimensional array
# representing daily images with some gaps, indicated by 0s
dataStack = np.zeros((2,5,5))
dataStack[1] = (np.random.random(25) * 100).reshape(dist.shape)
dataStack[0] = (np.random.random(25) * 100).reshape(dist.shape)
testdata = dataStack[1]
alternatedata = dataStack[0]
random_gap_locations = (np.random.random(25) * 30).reshape(dist.shape) > testdata
testdata[random_gap_locations] = 0
testdata[radius, radius] = 0
# in reality we will loop here to go through every gap (zero) location in the data
# for each image
gapz, gapy, gapx = 1, radius, radius
desLeft, desRight = gapx - radius, gapx + radius+1
desTop, desBottom = gapy - radius, gapy + radius+1
extentB, extentR = dataStack.shape[1:]
# handle the case where the gap is < search radius from the edge of
# the data. If this is the case, we can't use the full
# diam * diam window
dataL = max(0, desLeft)
maskL = 0 if desLeft >= 0 else abs(dataL - desLeft)
dataT = max(0, desTop)
maskT = 0 if desTop >= 0 else abs(dataT - desTop)
dataR = min(desRight, extentR)
maskR = diam if desRight <= extentR else diam - (desRight - extentR)
dataB = min(desBottom,extentB)
maskB = diam if desBottom <= extentB else diam - (desBottom - extentB)
# get the slice that we will be working within
# ranks, dist and circle are already populated
boundedWindowArray = fullWindowArray[maskT:maskB,maskL:maskR]
boundedWindowArray['alternatedata'] = alternatedata[dataT:dataB, dataL:dataR]
boundedWindowArray['thisdata'] = testdata[dataT:dataB, dataL:dataR]
locations_to_exclude = np.logical_or(boundedWindowArray['circle_template'],
np.logical_or
(boundedWindowArray['thisdata']==0,
boundedWindowArray['alternatedata']==0))
# the places that are inside the circular mask and where both images
# have data
locations_to_include = ~locations_to_exclude
number_available = np.count_nonzero(locations_to_include)
# we only want to do the interpolation calculations from the nearest n
# locations that have data available, n will be ~100 in reality
number_required = 3
data_to_use = boundedWindowArray[locations_to_include]
if number_available > number_required:
# argpartition seems to be v fast when number_required is
# substantially < data_to_use.size
# But partition on the structured array itself with order=['ranks']
# is almost 2 orders of magnitude slower!
reqIndices = np.argpartition(data_to_use['ranks'],number_required)[:number_required]
data_to_use = np.take(data_to_use,reqIndices)
else:
# we just use available_data and available_alternates as they are...
pass
# now do stuff with the selected data to calculate a value for the gap cell
I am still a beginner but I want to write a character-recognition-program. This program isn't ready yet. And I edited a lot, therefor the comments may not match exactly. I will use the 8-connectivity for the connected component labeling.
from PIL import Image
import numpy as np
im = Image.open("D:\\Python26\\PYTHON-PROGRAMME\\bild_schrift.jpg")
w,h = im.size
w = int(w)
h = int(h)
#2D-Array for area
area = []
for x in range(w):
area.append([])
for y in range(h):
area[x].append(2) #number 0 is white, number 1 is black
#2D-Array for letter
letter = []
for x in range(50):
letter.append([])
for y in range(50):
letter[x].append(0)
#2D-Array for label
label = []
for x in range(50):
label.append([])
for y in range(50):
label[x].append(0)
#image to number conversion
pix = im.load()
threshold = 200
for x in range(w):
for y in range(h):
aaa = pix[x, y]
bbb = aaa[0] + aaa[1] + aaa[2] #total value
if bbb<=threshold:
area[x][y] = 1
if bbb>threshold:
area[x][y] = 0
np.set_printoptions(threshold='nan', linewidth=10)
#matrix transponation
ccc = np.array(area)
area = ccc.T #better solution?
#find all black pixel and set temporary label numbers
i=1
for x in range(40): # width (later)
for y in range(40): # heigth (later)
if area[x][y]==1:
letter[x][y]=1
label[x][y]=i
i += 1
#connected components labeling
for x in range(40): # width (later)
for y in range(40): # heigth (later)
if area[x][y]==1:
label[x][y]=i
#if pixel has neighbour:
if area[x][y+1]==1:
#pixel and neighbour get the lowest label
pass # tomorrows work
if area[x+1][y]==1:
#pixel and neighbour get the lowest label
pass # tomorrows work
#should i also compare pixel and left neighbour?
#find width of the letter
#find height of the letter
#find the middle of the letter
#middle = [width/2][height/2] #?
#divide letter into 30 parts --> 5 x 6 array
#model letter
#letter A-Z, a-z, 0-9 (maybe more)
#compare each of the 30 parts of the letter with all model letters
#make a weighting
#print(letter)
im.save("D:\\Python26\\PYTHON-PROGRAMME\\bild2.jpg")
print('done')
OCR is not an easy task indeed. That's why text CAPTCHAs still work :)
To talk only about the letter extraction and not the pattern recognition, the technique you are using to separate the letters is called Connected Component Labeling. Since you are asking for a more efficient way to do this, try to implement the two-pass algorithm that's described in this article. Another description can be found in the article Blob extraction.
EDIT: Here's the implementation for the algorithm that I have suggested:
import sys
from PIL import Image, ImageDraw
class Region():
def __init__(self, x, y):
self._pixels = [(x, y)]
self._min_x = x
self._max_x = x
self._min_y = y
self._max_y = y
def add(self, x, y):
self._pixels.append((x, y))
self._min_x = min(self._min_x, x)
self._max_x = max(self._max_x, x)
self._min_y = min(self._min_y, y)
self._max_y = max(self._max_y, y)
def box(self):
return [(self._min_x, self._min_y), (self._max_x, self._max_y)]
def find_regions(im):
width, height = im.size
regions = {}
pixel_region = [[0 for y in range(height)] for x in range(width)]
equivalences = {}
n_regions = 0
#first pass. find regions.
for x in xrange(width):
for y in xrange(height):
#look for a black pixel
if im.getpixel((x, y)) == (0, 0, 0, 255): #BLACK
# get the region number from north or west
# or create new region
region_n = pixel_region[x-1][y] if x > 0 else 0
region_w = pixel_region[x][y-1] if y > 0 else 0
max_region = max(region_n, region_w)
if max_region > 0:
#a neighbour already has a region
#new region is the smallest > 0
new_region = min(filter(lambda i: i > 0, (region_n, region_w)))
#update equivalences
if max_region > new_region:
if max_region in equivalences:
equivalences[max_region].add(new_region)
else:
equivalences[max_region] = set((new_region, ))
else:
n_regions += 1
new_region = n_regions
pixel_region[x][y] = new_region
#Scan image again, assigning all equivalent regions the same region value.
for x in xrange(width):
for y in xrange(height):
r = pixel_region[x][y]
if r > 0:
while r in equivalences:
r = min(equivalences[r])
if not r in regions:
regions[r] = Region(x, y)
else:
regions[r].add(x, y)
return list(regions.itervalues())
def main():
im = Image.open(r"c:\users\personal\py\ocr\test.png")
regions = find_regions(im)
draw = ImageDraw.Draw(im)
for r in regions:
draw.rectangle(r.box(), outline=(255, 0, 0))
del draw
#im.show()
output = file("output.png", "wb")
im.save(output)
output.close()
if __name__ == "__main__":
main()
It's not 100% perfect, but since you are doing this only for learning purposes, it may be a good starting point. With the bounding box of each character you can now use a neural network as others have suggested here.
OCR is very, very hard. Even with computer-generated characters, it's quite challenging if you don't know the font and font size in advance. Even if you're matching characters exactly, I would not call it a "beginning" programming project; it's quite subtle.
If you want to recognize scanned, or handwritten characters, that's even harder - you'll need to use advanced math, algorithms, and machine learning. There are quite a few books and thousands of articles written about this topic, so you don't need to reinvent the wheel.
I admire your effort, but I don't think you've gotten far enough to hit any of the actual difficulties yet. So far you're just randomly exploring pixels and copying them from one array to another. You haven't actually done any comparison yet, and I'm not sure the purpose of your "random walk".
Why random? Writing correct randomized algorithms is quite difficult. I would recommend starting with a deterministic algorithm first.
Why are you copying from one array to another? Why not just compare directly?
When you get the comparison, you'll have to deal with the fact that the image is not exactly the same as the "prototype", and it's not clear how you'll deal with that.
Based on the code you've written so far, though, I have an idea for you: try writing a program that finds its way through a "maze" in an image. The input would be the image, plus the start pixel and the goal pixel. The output is a path through the maze from the start to the goal. This is a much easier problem than OCR - solving mazes is something that computers are great for - but it's still fun and challenging.
Most OCR algorithms these days are based on neural network algorithms. Hopfield networks are a good place to start. Based on the Hopfield Model available here in C, I built a very basic image recognition algorithm in python similar to what you describe. I've posted the full source here. It's a toy project and not suitable for real OCR, but can get you started in the right direction.
The Hopfield model is used as an autoassociative memory to store and recall a set of bitmap images. Images are stored by calculating a corresponding weight matrix. Thereafter, starting from an arbitrary configuration, the memory will settle on exactly that stored image, which is nearest to the starting configuration in terms of Hamming distance. Thus given an incomplete or corrupted version of a stored image, the network is able to recall the corresponding original image.
A Java applet to toy with an example can be found here; the network is trained with example inputs for the digits 0-9. Draw in the box on the right, click test and see the results from the network.
Don't let the mathematical notation intimidate you, the algorithms are straightforward once you get to source code.
OCR is very, very difficult! What approach to use to attempt OCR will be based on what you are trying to accomplish (hand writing recongnition, computer generated text reading, etc.)
However, to get you started, read up on Neural Networks and OCR. Here are a few jump-right-in articles on the subject:
http://www.codeproject.com/KB/cs/neural_network_ocr.aspx
http://www.codeproject.com/KB/dotnet/simple_ocr.aspx
Use your favorite search engine to find information.
Have fun!