I am currently working on OCR software and my idea is to use templates to try to recognize data inside invoices.
However scanned invoices can have several 'flaws' with them:
Not all invoices, based on a single template, are correctly aligned under the scanner.
People can write on invoices
etc.
Example of invoice: (Have to google it, sadly cannot add a more concrete version as client data is confidential obviously)
I find my data in the invoices based on the x-values of the text.
However I need to know the scale of the invoice and the offset from left/right, before I can do any real calculations with all data that I have retrieved.
What have I tried so far?
1) Making the image monochrome and use the left and right bounds of the first appearance of a black pixel. This fails due to the fact that people can write on invoices.
2) Divide the invoice up in vertical sections, use the sections that have the highest amount of black pixels. Fails due to the fact that the distribution is not always uniform amongst similar templates.
I could really use your help on (1) how to identify important points in invoices and (2) on what I should focus as the important points.
I hope the question is clear enough as it is quite hard to explain.
Detecting rotation
I would suggest you start by detecting straight lines.
Look (perhaps randomly) for small areas with high contrast, i.e. mostly white but a fair amount of very black pixels as well. Then try to fit a line to these black pixels, e.g. using least squares method. Drop the outliers, and fit another line to the remaining points. Iterate this as required. Evaluate how good that fit is, i.e. how many of the pixels in the observed area are really close to the line, and how far that line extends beyond the observed area. Do this process for a number of regions, and you should get a weighted list of lines.
For each line, you can compute the direction of the line itself and the direction orthogonal to that. One of these numbers can be chosen from an interval [0°, 90°), the other will be 90° plus that value, so storing one is enough. Take all these directions, and find one angle which best matches all of them. You can do that using a sliding window of e.g. 5°: slide accross that (cyclic) region and find a value where the maximal number of lines are within the window, then compute the average or median of the angles within that window. All of this computation can be done taking the weights of the lines into account.
Once you have found the direction of lines, you can rotate your image so that the lines are perfectly aligned to the coordinate axes.
Detecting translation
Assuming the image wasn't scaled at any point, you can then try to use a FFT-based correlation of the image to match it to the template. Convert both images to gray, pad them with zeros till the originals take up at most 1/2 the edge length of the padded image, which preferrably should be a power of two. FFT both images in both directions, multiply them element-wise and iFFT back. The resulting image will encode how much the two images would agree for a given shift relative to one another. Simply find the maximum, and you know how to make them match.
Added text will cause no problems at all. This method will work best for large areas, like the company logo and gray background boxes. Thin lines will provide a poorer match, so in those cases you might have to blur the picture before doing the correlation, to broaden the features. You don't have to use the blurred image for further processing; once you know the offset you can return to the rotated but unblurred version.
Now you know both rotation and translation, and assumed no scaling or shearing, so you know exactly which portion of the template corresponds to which portion of the scan. Proceed.
If rotation is solved already, I'd just sum up all pixel color values horizontally and vertically to a single horizontal / vertical "line". This should provide clear spikes where you have horizontal and vertical lines in the form.
p.s. Generated a corresponding horizontal image with Gimp's scaling capabilities, attached below (it's a bit hard to see because it's only one pixel high and may get scaled down because it's > 700 px wide; the url is http://i.stack.imgur.com/Zy8zO.png ).
Not sure if this may or may not be valid here on SO, but I was hoping someone can advise of the correct algorithm to use.
I have the following RAW data.
In the image you can see "steps". Essentially I wish to get these steps, but then get a moving average of all the data between. In the following image, you can see the moving average:
However you will notice that at the "steps", the moving average decreases the gradient where I wish to keep the high vertical gradient.
Is there any smoothing technique that will take into account a large vertical "offset", but smooth the other data?
Yup, I had to do something similar with images from a spacecraft.
Simple technique #1: use a median filter with a modest width - say about 5 samples, or 7. This provides an output value that is the median of the corresponding input value and several of its immediate neighbors on either side. It will get rid of those spikes, and do a good job preserving the step edges.
The median filter is provided in all number-crunching toolkits that I know of such as Matlab, Python/Numpy, IDL etc., and libraries for compiled languages such as C++, Java (though specific names don't come to mind right now...)
Technique #2, perhaps not quite as good: Use a Savitzky-Golay smoothing filter. This works by effectively making least-square polynomial fits to the data, at each output sample, using the corresponding input sample and a neighborhood of points (much like the median filter). The SG smoother is known for being fairly good at preserving peaks and sharp transistions.
The SG filter is usually provided by most signal processing and number crunching packages, but might not be as common as the median filter.
Technique #3, the most work and requiring the most experience and judgement: Go ahead and use a smoother - moving box average, Gaussian, whatever - but then create an output that blends between the original with the smoothed data. The blend, controlled by a new data series you create, varies from all-original (blending in 0% of the smoothed) to all-smoothed (100%).
To control the blending, start with an edge detector to detect the jumps. You may want to first median-filter the data to get rid of the spikes. Then broaden (dilation in image processing jargon) or smooth and renormalize the the edge detector's output, and flip it around so it gives 0.0 at and near the jumps, and 1.0 everywhere else. Perhaps you want a smooth transition joining them. It is an art to get this right, which depends on how the data will be used - for me, it's usually images to be viewed by Humans. An automated embedded control system might work best if tweaked differently.
The main advantage of this technique is you can plug in whatever kind of smoothing filter you like. It won't have any effect where the blend control value is zero. The main disadvantage is that the jumps, the small neighborhood defined by the manipulated edge detector output, will contain noise.
I recommend first detecting the steps and then smoothing each step individually.
You know how to do the smoothing, and edge/step detection is pretty easy also (see here, for example). A typical edge detection scheme is to smooth your data and then multiply/convolute/cross-corelate it with some filter (for example the array [-1,1] that will show you where the steps are). In a mathematical context this can be viewed as studying the derivative of your plot to find inflection points (for some of the filters).
An alternative "hackish" solution would be to do a moving average but exclude outliers from the smoothing. You can decide what an outlier is by using some threshold t. In other words, for each point p with value v, take x points surrounding it and find the subset of those points which are between v - t and v + t, and take the average of these points as the new value of p.
I have some map files consisting of 'polylines' (each line is just a list of vertices) representing tunnels, and I want to try and find the tunnel 'center line' (shown, roughly, in red below).
I've had some success in the past using Delaunay triangulation but I'd like to avoid that method as it does not (in general) allow for easy/frequent modification of my map data.
Any ideas on how I might be able to do this?
An "algorithm" that works well with localized data changes.
The critic's view
The Good
The nice part is that it uses a mixture of image processing and graph operations available in most libraries, may be parallelized easily, is reasonable fast, may be tuned to use a relatively small memory footprint and doesn't have to be recalculated outside the modified area if you store the intermediate results.
The Bad
I wrote "algorithm", in quotes, just because I developed it and surely is not robust enough to cope with pathological cases. If your graph has a lot of cycles you may end up with some phantom lines. More on this and examples later.
And The Ugly
The ugly part is that you need to be able to flood fill the map, which is not always possible. I posted a comment a few days ago asking if your graphs can be flood filled, but didn't receive an answer. So I decided to post it anyway.
The Sketch
The idea is:
Use image processing to get a fine line of pixels representing the center path
Partition the image in chunks commensurated to the tunnel thinnest passages
At each partition, represent a point at the "center of mass" of the contained pixels
Use those pixels to represent the Vertices of a Graph
Add Edges to the Graph based on a "near neighbour" policy
Remove spurious small cycles in the induced Graph
End- The remaining Edges represent your desired path
The parallelization opportunity arises from the fact that the partitions may be computed in standalone processes, and the resulting graph may be partitioned to find the small cycles that need to be removed. These factors also allow to reduce the memory needed by serializing instead of doing calcs in parallel, but I didn't go trough this.
The Plot
I'll no provide pseudocode, as the difficult part is just that not covered by your libraries. Instead of pseudocode I'll post the images resulting from the successive steps.
I wrote the program in Mathematica, and I can post it if is of some service to you.
A- Start with a nice flood filled tunnel image
B- Apply a Distance Transformation
The Distance Transformation gives the distance transform of image, where the value of each pixel is replaced by its distance to the nearest background pixel.
You can see that our desired path is the Local Maxima within the tunnel
C- Convolve the image with an appropriate kernel
The selected kernel is a Laplacian-of-Gaussian kernel of pixel radius 2. It has the magic property of enhancing the gray level edges, as you can see below.
D- Cutoff gray levels and Binarize the image
To get a nice view of the center line!
Comment
Perhaps that is enough for you, as you ay know how to transform a thin line to an approximate piecewise segments sequence. As that is not the case for me, I continued this path to get the desired segments.
E- Image Partition
Here is when some advantages of the algorithm show up: you may start using parallel processing or decide to process each segment at a time. You may also compare the resulting segments with the previous run and re-use the previous results
F- Center of Mass detection
All the white points in each sub-image are replaced by only one point at the center of mass
XCM = (Σ i∈Points Xi)/NumPoints
YCM = (Σ i∈Points Yi)/NumPoints
The white pixels are difficult to see (asymptotically difficult with param "a" age), but there they are.
G- Graph setup from Vertices
Form a Graph using the selected points as Vertex. Still no Edges.
H- select Candidate Edges
Using the Euclidean Distance between points, select candidate edges. A cutoff is used to select an appropriate set of Edges. Here we are using 1.5 the subimagesize.
As you can see the resulting Graph have a few small cycles that we are going to remove in the next step.
H- Remove Small Cycles
Using a Cycle detection routine we remove the small cycles up to a certain length. The cutoff length depends on a few parms and you should figure it empirically for your graphs family
I- That's it!
You can see that the resulting center line is shifted a little bit upwards. The reason is that I'm superimposing images of different type in Mathematica ... and I gave up trying to convince the program to do what I want :)
A Few Shots
As I did the testing, I collected a few images. They are probably the most un-tunnelish things in the world, but my Tunnels-101 went astray.
Anyway, here they are. Remember that I have a displacement of a few pixels upwards ...
HTH !
.
Update
Just in case you have access to Mathematica 8 (I got it today) there is a new function Thinning. Just look:
This is a pretty classic skeletonization problem; there are lots of algorithms available. Some algorithms work in principle on outline contours, but since almost everyone uses them on images, I'm not sure how available such things will be. Anyway, if you can just plot and fill the sewer outlines and then use a skeletonization algorithm, you could get something close to the midline (within pixel resolution).
Then you could walk along those lines and do a binary search with circles until you hit at least two separate line segments (three if you're at a branch point). The midpoint of the two spots you first hit, or the center of a circle touching the three points you first hit, is a good estimate of the center.
Well in Python using package skimage it is an easy task as follows.
import pylab as pl
from skimage import morphology as mp
tun = 1-pl.imread('tunnel.png')[...,0] #your tunnel image
skl = mp.medial_axis(tun) #skeleton
pl.subplot(121)
pl.imshow(tun,cmap=pl.cm.gray)
pl.subplot(122)
pl.imshow(skl,cmap=pl.cm.gray)
pl.show()
I'm working on a game, and I've come up with a rather interesting problem: clever ways to draw starfields.
It's a 2D game, so the action can scroll in the X and Y directions. In addition, we can adjust the scale to show more or less of the play area. I'd also like the starfield to have fake parallax to give an impression of depth.
Right now I'm doing this in the traditional way, by having a big array of stars, each of which is tagged by a 'depth' factor. To draw, I translate each star according to the camera position multiplied by the 'depth', so some stars move a lot, and some move a little. This all works fine, but of course since I have a finite number of stars in my array I have issues when the camera moves too far or we zoom out too much. This is will all work, but is involving lots of code and special cases.
This offends my sense of elegance. There has got be a better way of achieving this.
I've considered procedurally generating my stars, which allows me to have an unlimited number: e.g. by using a fixed seed and PRNG to determine the coordinates. I would need to divide the sky up into tiles, generate the seed by hashing the tile coordinates, and then draw, say, 100 stars per tile. This allows me to extend my starfield indefinitely in all directions while still only needing to consider the tiles that are visible --- but this doesn't work with the 'depth' factor, as this allows stars to stray outside their tile. I could simply use multiple layered non-parallax starfields using this algorithm but this strikes me as cheating.
And, of course, I need to do all this every frame, so it's got to be fast.
What do you all reckon?
Have a few layers of stars.
For each layer, use a seeded random number generator (or just an array) to generate the amount of blank space between a star and the next one (a poisson distibution, if you want to be picky about it). You want the stars pretty sparse, so the blank space will often be more than whole row. The back layers will be more dense than the front ones, obviously.
Use this to give yourself several tiles each (say) two screens wide. Scroll the starfield by keeping track of where that "first" star is for each layer.
The player won't notice the tiling, because you scroll the tiles at different rates for each layer, especially if you use a few layers that are each fairly sparse.
As stars in the background don't move as fast as those in the foreground, you could maybe make multi-layer tiles for the background and replace them with one-layer-ones when you've got time to do that. Oh, and how about repeating patterns in the background layers? This would maybe allow you to pregenerate all background tiles - you could still shift them in height and overlay multiple ones with random offsets or so to make it look random.
Is there anything wrong with wrapping the star field around in X and Y? Because of your depth, the wraparound distance should depend on the depth, but you can do that. Each recorded star at (x,y,depth) should appear at all points
[x + j * S * depth, y + k * S * depth]
for all integers j and k. S is a wraparound parameter. If S is 1 then wraparound happens immediately and all stars are always shown somewhere. If S is higher wraparound doesn't happen immediately and some stars are shown off screen. You'll probably want S big enough to ensure no repeats at maximum zoom out.
Each frame, render the stars on one single bitmap/layer. They are only dots, and so it will be faster than using any algorithm with multiple layers.
Now you need an infinite 2D-grid of 3D-boxes filled with a finite number of stars. For each box, you can define an individual RANDOM_SEED value, using its grid-coordinates. The stars in each box can be generated on-the-fly.
Remember to correct the perspective when you zoom: Each 3D-box has a near-rectangle (front-face) and a far-rectangle. You will see more stars of neighbouring boxes, whenever the far-rectangle or near-rectangle shrinks in your view.
Your far-rectangles should never be smaller than half the width of the near-rectangles, otherwise it might be troublesome: You might have to scan huge lists of stars where most of them are out of bounds. You can realize stars behind the far-rectangles via additional 2D-grids of 3D-boxes with other sizes and depths.
Why not combine the coordinates of the starfield 3D boxes to form the random number seed? Use a global "adjustment" if you want to produce different universes. That way you don't need to track the boxes you can't see because the contents are fixed by their location.
I'm looking to create a base table of images and then compare any new images against that to determine if the new image is an exact (or close) duplicate of the base.
For example: if you want to reduce storage of the same image 100's of times, you could store one copy of it and provide reference links to it. When a new image is entered you want to compare to an existing image to make sure it's not a duplicate ... ideas?
One idea of mine was to reduce to a small thumbnail and then randomly pick 100 pixel locations and compare.
Below are three approaches to solving this problem (and there are many others).
The first is a standard approach in computer vision, keypoint matching. This may require some background knowledge to implement, and can be slow.
The second method uses only elementary image processing, and is potentially faster than the first approach, and is straightforward to implement. However, what it gains in understandability, it lacks in robustness -- matching fails on scaled, rotated, or discolored images.
The third method is both fast and robust, but is potentially the hardest to implement.
Keypoint Matching
Better than picking 100 random points is picking 100 important points. Certain parts of an image have more information than others (particularly at edges and corners), and these are the ones you'll want to use for smart image matching. Google "keypoint extraction" and "keypoint matching" and you'll find quite a few academic papers on the subject. These days, SIFT keypoints are arguably the most popular, since they can match images under different scales, rotations, and lighting. Some SIFT implementations can be found here.
One downside to keypoint matching is the running time of a naive implementation: O(n^2m), where n is the number of keypoints in each image, and m is the number of images in the database. Some clever algorithms might find the closest match faster, like quadtrees or binary space partitioning.
Alternative solution: Histogram method
Another less robust but potentially faster solution is to build feature histograms for each image, and choose the image with the histogram closest to the input image's histogram. I implemented this as an undergrad, and we used 3 color histograms (red, green, and blue), and two texture histograms, direction and scale. I'll give the details below, but I should note that this only worked well for matching images VERY similar to the database images. Re-scaled, rotated, or discolored images can fail with this method, but small changes like cropping won't break the algorithm
Computing the color histograms is straightforward -- just pick the range for your histogram buckets, and for each range, tally the number of pixels with a color in that range. For example, consider the "green" histogram, and suppose we choose 4 buckets for our histogram: 0-63, 64-127, 128-191, and 192-255. Then for each pixel, we look at the green value, and add a tally to the appropriate bucket. When we're done tallying, we divide each bucket total by the number of pixels in the entire image to get a normalized histogram for the green channel.
For the texture direction histogram, we started by performing edge detection on the image. Each edge point has a normal vector pointing in the direction perpendicular to the edge. We quantized the normal vector's angle into one of 6 buckets between 0 and PI (since edges have 180-degree symmetry, we converted angles between -PI and 0 to be between 0 and PI). After tallying up the number of edge points in each direction, we have an un-normalized histogram representing texture direction, which we normalized by dividing each bucket by the total number of edge points in the image.
To compute the texture scale histogram, for each edge point, we measured the distance to the next-closest edge point with the same direction. For example, if edge point A has a direction of 45 degrees, the algorithm walks in that direction until it finds another edge point with a direction of 45 degrees (or within a reasonable deviation). After computing this distance for each edge point, we dump those values into a histogram and normalize it by dividing by the total number of edge points.
Now you have 5 histograms for each image. To compare two images, you take the absolute value of the difference between each histogram bucket, and then sum these values. For example, to compare images A and B, we would compute
|A.green_histogram.bucket_1 - B.green_histogram.bucket_1|
for each bucket in the green histogram, and repeat for the other histograms, and then sum up all the results. The smaller the result, the better the match. Repeat for all images in the database, and the match with the smallest result wins. You'd probably want to have a threshold, above which the algorithm concludes that no match was found.
Third Choice - Keypoints + Decision Trees
A third approach that is probably much faster than the other two is using semantic texton forests (PDF). This involves extracting simple keypoints and using a collection decision trees to classify the image. This is faster than simple SIFT keypoint matching, because it avoids the costly matching process, and keypoints are much simpler than SIFT, so keypoint extraction is much faster. However, it preserves the SIFT method's invariance to rotation, scale, and lighting, an important feature that the histogram method lacked.
Update:
My mistake -- the Semantic Texton Forests paper isn't specifically about image matching, but rather region labeling. The original paper that does matching is this one: Keypoint Recognition using Randomized Trees. Also, the papers below continue to develop the ideas and represent the state of the art (c. 2010):
Fast Keypoint Recognition using Random Ferns - faster and more scalable than Lepetit 06
BRIEF: Binary Robust Independent Elementary Features - less robust but very fast -- I think the goal here is real-time matching on smart phones and other handhelds
The best method I know of is to use a Perceptual Hash. There appears to be a good open source implementation of such a hash available at:
http://phash.org/
The main idea is that each image is reduced down to a small hash code or 'fingerprint' by identifying salient features in the original image file and hashing a compact representation of those features (rather than hashing the image data directly). This means that the false positives rate is much reduced over a simplistic approach such as reducing images down to a tiny thumbprint sized image and comparing thumbprints.
phash offers several types of hash and can be used for images, audio or video.
This post was the starting point of my solution, lots of good ideas here so I though I would share my results. The main insight is that I've found a way to get around the slowness of keypoint-based image matching by exploiting the speed of phash.
For the general solution, it's best to employ several strategies. Each algorithm is best suited for certain types of image transformations and you can take advantage of that.
At the top, the fastest algorithms; at the bottom the slowest (though more accurate). You might skip the slow ones if a good match is found at the faster level.
file-hash based (md5,sha1,etc) for exact duplicates
perceptual hashing (phash) for rescaled images
feature-based (SIFT) for modified images
I am having very good results with phash. The accuracy is good for rescaled images. It is not good for (perceptually) modified images (cropped, rotated, mirrored, etc). To deal with the hashing speed we must employ a disk cache/database to maintain the hashes for the haystack.
The really nice thing about phash is that once you build your hash database (which for me is about 1000 images/sec), the searches can be very, very fast, in particular when you can hold the entire hash database in memory. This is fairly practical since a hash is only 8 bytes.
For example, if you have 1 million images it would require an array of 1 million 64-bit hash values (8 MB). On some CPUs this fits in the L2/L3 cache! In practical usage I have seen a corei7 compare at over 1 Giga-hamm/sec, it is only a question of memory bandwidth to the CPU. A 1 Billion-image database is practical on a 64-bit CPU (8GB RAM needed) and searches will not exceed 1 second!
For modified/cropped images it would seem a transform-invariant feature/keypoint detector like SIFT is the way to go. SIFT will produce good keypoints that will detect crop/rotate/mirror etc. However the descriptor compare is very slow compared to hamming distance used by phash. This is a major limitation. There are a lot of compares to do, since there are maximum IxJxK descriptor compares to lookup one image (I=num haystack images, J=target keypoints per haystack image, K=target keypoints per needle image).
To get around the speed issue, I tried using phash around each found keypoint, using the feature size/radius to determine the sub-rectangle. The trick to making this work well, is to grow/shrink the radius to generate different sub-rect levels (on the needle image). Typically the first level (unscaled) will match however often it takes a few more. I'm not 100% sure why this works, but I can imagine it enables features that are too small for phash to work (phash scales images down to 32x32).
Another issue is that SIFT will not distribute the keypoints optimally. If there is a section of the image with a lot of edges the keypoints will cluster there and you won't get any in another area. I am using the GridAdaptedFeatureDetector in OpenCV to improve the distribution. Not sure what grid size is best, I am using a small grid (1x3 or 3x1 depending on image orientation).
You probably want to scale all the haystack images (and needle) to a smaller size prior to feature detection (I use 210px along maximum dimension). This will reduce noise in the image (always a problem for computer vision algorithms), also will focus detector on more prominent features.
For images of people, you might try face detection and use it to determine the image size to scale to and the grid size (for example largest face scaled to be 100px). The feature detector accounts for multiple scale levels (using pyramids) but there is a limitation to how many levels it will use (this is tunable of course).
The keypoint detector is probably working best when it returns less than the number of features you wanted. For example, if you ask for 400 and get 300 back, that's good. If you get 400 back every time, probably some good features had to be left out.
The needle image can have less keypoints than the haystack images and still get good results. Adding more doesn't necessarily get you huge gains, for example with J=400 and K=40 my hit rate is about 92%. With J=400 and K=400 the hit rate only goes up to 96%.
We can take advantage of the extreme speed of the hamming function to solve scaling, rotation, mirroring etc. A multiple-pass technique can be used. On each iteration, transform the sub-rectangles, re-hash, and run the search function again.
My company has about 24million images come in from manufacturers every month. I was looking for a fast solution to ensure that the images we upload to our catalog are new images.
I want to say that I have searched the internet far and wide to attempt to find an ideal solution. I even developed my own edge detection algorithm.
I have evaluated speed and accuracy of multiple models.
My images, which have white backgrounds, work extremely well with phashing. Like redcalx said, I recommend phash or ahash. DO NOT use MD5 Hashing or anyother cryptographic hashes. Unless, you want only EXACT image matches. Any resizing or manipulation that occurs between images will yield a different hash.
For phash/ahash, Check this out: imagehash
I wanted to extend *redcalx'*s post by posting my code and my accuracy.
What I do:
from PIL import Image
from PIL import ImageFilter
import imagehash
img1=Image.open(r"C:\yourlocation")
img2=Image.open(r"C:\yourlocation")
if img1.width<img2.width:
img2=img2.resize((img1.width,img1.height))
else:
img1=img1.resize((img2.width,img2.height))
img1=img1.filter(ImageFilter.BoxBlur(radius=3))
img2=img2.filter(ImageFilter.BoxBlur(radius=3))
phashvalue=imagehash.phash(img1)-imagehash.phash(img2)
ahashvalue=imagehash.average_hash(img1)-imagehash.average_hash(img2)
totalaccuracy=phashvalue+ahashvalue
Here are some of my results:
item1 item2 totalsimilarity
desk1 desk1 3
desk1 phone1 22
chair1 desk1 17
phone1 chair1 34
Hope this helps!
As cartman pointed out, you can use any kind of hash value for finding exact duplicates.
One starting point for finding close images could be here. This is a tool used by CG companies to check if revamped images are still showing essentially the same scene.
I have an idea, which can work and it most likely to be very fast.
You can sub-sample an image to say 80x60 resolution or comparable,
and convert it to grey scale (after subsampling it will be faster).
Process both images you want to compare.
Then run normalised sum of squared differences between two images (the query image and each from the db),
or even better Normalised Cross Correlation, which gives response closer to 1, if
both images are similar.
Then if images are similar you can proceed to more sophisticated techniques
to verify that it is the same images.
Obviously this algorithm is linear in terms of number of images in your database
so even though it is going to be very fast up to 10000 images per second on the modern hardware.
If you need invariance to rotation, then a dominant gradient can be computed
for this small image, and then the whole coordinate system can be rotated to canonical
orientation, this though, will be slower. And no, there is no invariance to scale here.
If you want something more general or using big databases (million of images), then
you need to look into image retrieval theory (loads of papers appeared in the last 5 years).
There are some pointers in other answers. But It might be overkill, and the suggest histogram approach will do the job. Though I would think combination of many different
fast approaches will be even better.
I believe that dropping the size of the image down to an almost icon size, say 48x48, then converting to greyscale, then taking the difference between pixels, or Delta, should work well. Because we're comparing the change in pixel color, rather than the actual pixel color, it won't matter if the image is slightly lighter or darker. Large changes will matter since pixels getting too light/dark will be lost. You can apply this across one row, or as many as you like to increase the accuracy. At most you'd have 47x47=2,209 subtractions to make in order to form a comparable Key.
Picking 100 random points could mean that similar (or occasionally even dissimilar) images would be marked as the same, which I assume is not what you want. MD5 hashes wouldn't work if the images were different formats (png, jpeg, etc), had different sizes, or had different metadata. Reducing all images to a smaller size is a good bet, doing a pixel-for- pixel comparison shouldn't take too long as long as you're using a good image library / fast language, and the size is small enough.
You could try making them tiny, then if they are the same perform another comparison on a larger size - could be a good combination of speed and accuracy...
What we loosely refer to as duplicates can be difficult for algorithms to discern.
Your duplicates can be either:
Exact Duplicates
Near-exact Duplicates. (minor edits of image etc)
perceptual Duplicates (same content, but different view, camera etc)
No1 & 2 are easier to solve. No 3. is very subjective and still a research topic.
I can offer a solution for No1 & 2.
Both solutions use the excellent image hash- hashing library: https://github.com/JohannesBuchner/imagehash
Exact duplicates
Exact duplicates can be found using a perceptual hashing measure.
The phash library is quite good at this. I routinely use it to clean
training data.
Usage (from github site) is as simple as:
from PIL import Image
import imagehash
# image_fns : List of training image files
img_hashes = {}
for img_fn in sorted(image_fns):
hash = imagehash.average_hash(Image.open(image_fn))
if hash in img_hashes:
print( '{} duplicate of {}'.format(image_fn, img_hashes[hash]) )
else:
img_hashes[hash] = image_fn
Near-Exact Duplicates
In this case you will have to set a threshold and compare the hash values for their distance from each
other. This has to be done by trial-and-error for your image content.
from PIL import Image
import imagehash
# image_fns : List of training image files
img_hashes = {}
epsilon = 50
for img_fn1, img_fn2 in zip(image_fns, image_fns[::-1]):
if image_fn1 == image_fn2:
continue
hash1 = imagehash.average_hash(Image.open(image_fn1))
hash2 = imagehash.average_hash(Image.open(image_fn2))
if hash1 - hash2 < epsilon:
print( '{} is near duplicate of {}'.format(image_fn1, image_fn2) )
If you have a large number of images, look into a Bloom filter, which uses multiple hashes for a probablistic but efficient result. If the number of images is not huge, then a cryptographic hash like md5 should be sufficient.
I think it's worth adding to this a phash solution I built that we've been using for a while now: Image::PHash. It is a Perl module, but the main parts are in C. It is several times faster than phash.org and has a few extra features for DCT-based phashes.
We had dozens of millions of images already indexed on a MySQL database, so I wanted something fast and also a way to use MySQL indices (which don't work with hamming distance), which led me to use "reduced" hashes for direct matches, the module doc discusses this.
It's quite simple to use:
use Image::PHash;
my $iph1 = Image::PHash->new('file1.jpg');
my $p1 = $iph1->pHash();
my $iph2 = Image::PHash->new('file2.jpg');
my $p2 = $iph2->pHash();
my $diff = Image::PHash::diff($p1, $p2);
I made a very simple solution in PHP for comparing images several years ago. It calculates a simple hash for each image, and then finds the difference. It works very nice for cropped or cropped with translation versions of the same image.
First I resize the image to a small size, like 24x24 or 36x36. Then I take each column of pixels and find average R,G,B values for this column.
After each column has its own three numbers, I do two passes: first on odd columns and second on even ones. The first pass sums all the processed cols and then divides by their number ( [1] + [2] + [5] + [N-1] / (N/2) ). The second pass works in another manner: ( [3] - [4] + [6] - [8] ... / (N/2) ).
So now I have two numbers. As I found out experimenting, the first one is a major one: if it's far from the values of another image, they are not similar from the human point of view at all.
So, the first one represents the average brightness of the image (again, you can pay most attention to green channel, then the red one, etc, but the default R->G->B order works just fine). The second number can be compared if the first two are very close, and it in fact represents the overall contrast of the image: if we have some black/white pattern or any contrast scene (lighted buildings in the city at night, for example) and if we are lucky, we will get huge numbers here if out positive members of sum are mostly bright, and negative ones are mostly dark, or vice versa. As I want my values to be always positive, I divide by 2 and shift by 127 here.
I wrote the code in PHP in 2017, and seems I lost the code. But I still have the screenshots:
The same image:
Black & White version:
Cropped version:
Another image, ranslated version:
Same color gamut as 4th, but another scene:
I tuned the difference thresholds so that the results are really nice. But as you can see, this simple algorithm cannot do anything good with simple scene translations.
On a side note I can notice that a modification can be written to make cropped copies from each of two images at 75-80 percent, 4 at the corners or 8 at the corners and middles of the edges, and then by comparing the cropped variants with another whole image just the same way; and if one of them gets a significantly better similarity score, then use its value instead of the default one).