I am new to sci-kit learn. I have viewed the online tutorials but they all seem to leverage existing data (e.g., digits, iris, etc). I need the information on how to process images so that they can be used by scikit learn.
Details of my Study: I have a webcam set up outside my office. It captures all of the traffic on my street that passes in the field of view. I have cropped several hundred images of sedans, trucks and SUV's. The goal is to predict whether a vehicle is one of these categories. I have applied Histogram Oriented Gradients (HOG) to these images which I have attached for your review to see the differences in the categories. This blog will not allow me to post any images but you can see them here https://stats.stackexchange.com/questions/149421/obtaining-a-hog-feature-vector-for-implementation-in-svm-in-python. I posted the same question at this site but no response. This post is the closest answer I have found. Resize HOG feature for Scikit-Learn classifier
I wish to train an SVM classifier based on these images. I understand that there are algorithms that exist in scikit-image that prepares the HOG images for use in scikit-learn. Can someone help me understand this process. I am also grateful for any thoughts based on your experience as to the probability of success of this classification study. I also understand that I need to train the model using a negative images ( ones with no vehicles. How is this done?
I know I am asking a lot but I am surprised no one that I am aware of has done a tutorial on these early steps. It seems like a fairly elementary study.
I want to build a web application that lets users upload documents, videos, images, music, and then give them an ability to search them. Think of it as Dropbox + Semantic Search.
When user uploads a new file, e.g. Document1.docx, how could I automatically generate tags based on the content of the file? In other words no user input is needed to determine what the file is about. If suppose that Document1.docx is a research paper on data mining, then when user searches for data mining, or research paper, or document1, that file should be returned in search results, since data mining and research paper will most likely be potential auto-generated tags for that given document.
1. Which algorithms would you recommend for this problem?
2. Is there an natural language library that could do this for me?
3. Which machine learning techniques should I look into to improve tagging precision?
4. How could I extend this to video and image automatic tagging?
Thanks in advance!
The most common unsupervised machine learning model for this type of task is Latent Dirichlet Allocation (LDA). This model automatically infers a collection of topics over a corpus of documents based on the words in those documents. Running LDA on your set of documents would assign words with probability to certain topics when you search for them, and then you could retrieve the documents with the highest probabilities to be relevant to that word.
There have been some extensions to images and music as well, see http://cseweb.ucsd.edu/~dhu/docs/research_exam09.pdf.
LDA has several efficient implementations in several languages:
many implementations from the original researchers
http://mallet.cs.umass.edu/, written in Java and recommended by others on SO
PLDA: a fast, parallelized C++ implementation
These guys propose an alternative to LDA.
Automatic Tag Recommendation Algorithms for
Social Recommender Systems
http://research.microsoft.com/pubs/79896/tagging.pdf
Haven't read thru the whole paper but they have two algorithms:
Supervised learning version. This isn't that bad. You can use Wikipedia to train the algorithm
"Prototype" version. Haven't had a chance to go thru this but this is what they recommend
UPDATE: I've researched this some more and I've found another approach. Basically, it's a two-stage approach that's very simple to understand and implement. While too slow for 100,000s of documents, it (probably) has good performance for 1000s of docs (so it's perfect for tagging a single user's documents). I'm going to try this approach and will report back on performance/usability.
In the mean time, here's the approach:
Use TextRank as per http://qr.ae/36RAP to generate a tag list for a single document. This generates a tag list for a single document independent of other documents.
Use the algorithm from "Using Machine Learning to Support Continuous
Ontology Development" (https://www.researchgate.net/publication/221630712_Using_Machine_Learning_to_Support_Continuous_Ontology_Development) to integrate the tag list (from step 1) into the existing tag list.
Text documents can be tagged using this keyphrase extraction algorithm/package.
http://www.nzdl.org/Kea/
Currently it supports limited type of documents (Agricultural and medical I guess) but you can train it according to your requirements.
I'm not sure how would the image/video part work out, unless you're doing very accurate object detection (which has it's own shortcomings). How are you planning to do it ?
You want Doc-Tags (https://www.Doc-Tags.com) which is a commercial product that automatically and Unsupervised - generates Contextually Accurate Document Tags. The built-in Reporting functionality makes the product a light-weight document management system.
For Developers wanting to customize their own approach - the source code is available (very cheap) and the back-end service xAIgent (https://xAIgent.com) is very inexpensive to use.
I posted a blog article today to answer your question.
http://scottge.net/2015/06/30/automatic-image-and-video-tagging/
There are basically two approaches to automatically extract keywords from images and videos.
Multiple Instance Learning (MIL)
Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), and the variants
In the above blog article, I list the latest research papers to illustrate the solutions. Some of them even include demo site and source code.
Thanks, Scott
Would OCR Software be able to reliably translate an image such as the following into a list of values?
UPDATE:
In more detail the task is as follows:
We have a client application, where the user can open a report. This report contains a table of values.
But not every report looks the same - different fonts, different spacing, different colors, maybe the report contains many tables with different number of rows/columns...
The user selects an area of the report which contains a table. Using the mouse.
Now we want to convert the selected table into values - using our OCR tool.
At the time when the user selects the rectangular area I can ask for extra information
to help with the OCR process, and ask for confirmation that the values have been correct recognised.
It will initially be an experimental project, and therefore most likely with an OpenSource OCR tool - or at least one that does not cost any money for experimental purposes.
Simple answer is YES, you should just choose right tools.
I don't know if open source can ever get close to 100% accuracy on those images, but based on the answers here probably yes, if you spend some time on training and solve table analisys problem and stuff like that.
When we talk about commertial OCR like ABBYY or other, it will provide you 99%+ accuracy out of the box and it will detect tables automatically. No training, no anything, just works. Drawback is that you have to pay for it $$. Some would object that for open source you pay your time to set it up and mantain - but everyone decides for himself here.
However if we talk about commertial tools, there is more choice actually. And it depends on what you want. Boxed products like FineReader are actually targeting on converting input documents into editable documents like Word or Excell. Since you want actually to get data, not the Word document, you may need to look into different product category - Data Capture, which is essentially OCR plus some additional logic to find necessary data on the page. In case of invoice it could be Company name, Total amount, Due Date, Line items in the table, etc.
Data Capture is complicated subject and requires some learning, but being properly used can give quaranteed accuracy when capturing data from the documents. It is using different rules for data cross-check, database lookups, etc. When necessary it may send datafor manual verification. Enterprises are widely usind Data Capture applicaitons to enter millions of documents every month and heavily rely on data extracted in their every day workflow.
And there are also OCR SDK ofcourse, that will give you API access to recognition results and you will be able to program what to do with the data.
If you describe your task in more detail I can provide you with advice what direction is easier to go.
UPDATE
So what you do is basically Data Capture application, but not fully automated, using so-called "click to index" approach. There is number of applications like that on the market: you scan images and operator clicks on the text on the image (or draws rectangle around it) and then populates fields to database. It is good approach when number of images to process is relatively small, and manual workload is not big enough to justify cost of fully automated application (yes, there are fully automated systems that can do images with different font, spacing, layout, number of rows in the tables and so on).
If you decided to develop stuff and instead of buying, then all you need here is to chose OCR SDK. All UI you are going to write yoursself, right? The big choice is to decide: open source or commercial.
Best Open source is tesseract OCR, as far as I know. It is free, but may have real problems with table analysis, but with manual zoning approach this should not be the problem. As to OCR accuracty - people are often train OCR for font to increase accuracy, but this should not be the case for you, since fonts could be different. So you can just try tesseract out and see what accuracy you will get - this will influence amount of manual work to correct it.
Commertial OCR will give higher accuracy but will cost you money. I think you should anyway take a look to see if it worth it, or tesserack is good enough for you. I think the simplest way would be to download trial version of some box OCR prouct like FineReader. You will get good idea what accuracy would be in OCR SDK then.
If you always have solid borders in your table, you can try this solution:
Locate the horizontal and vertical lines on each page (long runs of
black pixels)
Segment the image into cells using the line coordinates
Clean up each cell (remove borders, threshold to black and white)
Perform OCR on each cell
Assemble results into a 2D array
Else your document have a borderless table, you can try to follow this line:
Optical Character Recognition is pretty amazing stuff, but it isn’t
always perfect. To get the best possible results, it helps to use the
cleanest input you can. In my initial experiments, I found that
performing OCR on the entire document actually worked pretty well as
long as I removed the cell borders (long horizontal and vertical
lines). However, the software compressed all whitespace into a single
empty space. Since my input documents had multiple columns with
several words in each column, the cell boundaries were getting lost.
Retaining the relationship between cells was very important, so one
possible solution was to draw a unique character, like “^” on each
cell boundary – something the OCR would still recognize and that I
could use later to split the resulting strings.
I found all this information in this link, asking Google "OCR to table". The author published a full algorithm using Python and Tesseract, both opensource solutions!
If you want to try the Tesseract power, maybe you should try this site:
http://www.free-ocr.com/
Which OCR you are talking about?
Will you be developing codes based on that OCR or you will be using something off the shelves?
FYI:
Tesseract OCR
it has implemented the document reading executable, so you can feed the whole page in, and it will extract characters for you. It recognizes blank spaces pretty well, it might be able to help with tab-spacing.
I've been OCR'ing scanned documents since '98. This is a recurring problem for scanned docs, specially for those that include rotated and/or skewed pages.
Yes, there are several good commercial systems and some could provide, once well configured, terrific automatic data-mining rate, asking for the operator's help only for those very degraded fields. If I were you, I'd rely on some of them.
If commercial choices threat your budget, OSS can lend a hand. But, "there's no free lunch". So, you'll have to rely on a bunch of tailor-made scripts to scaffold an affordable solution to process your bunch of docs. Fortunately, you are not alone. In fact, past last decades, many people have been dealing with this. So, IMHO, the best and concise answer for this question is provided by this article:
https://datascience.blog.wzb.eu/2017/02/16/data-mining-ocr-pdfs-using-pdftabextract-to-liberate-tabular-data-from-scanned-documents/
Its reading is worth! The author offers useful tools of his own, but the article's conclusion is very important to give you a good mindset about how to solve this kind of problem.
"There is no silver bullet."
(Fred Brooks, The Mitical Man-Month)
It really depends on implementation.
There are a few parameters that affect the OCR's ability to recognize:
1. How well the OCR is trained - the size and quality of the examples database
2. How well it is trained to detect "garbage" (besides knowing what's a letter, you need to know what is NOT a letter).
3. The OCR's design and type
4. If it's a Nerural Network, the Nerural Network structure affects its ability to learn and "decide".
So, if you're not making one of your own, it's just a matter of testing different kinds until you find one that fits.
You could try other approach. With tesseract (or other OCRS) you can get coordinates for each word. Then you can try to group those words by vercital and horizontal coordinates to get rows/columns. For example to tell a difference between a white space and tab space. It takes some practice to get good results but it is possible. With this method you can detect tables even if the tables use invisible separators - no lines. The word coordinates are solid base for table recog
We also have struggled with the issue of recognizing text within tables. There are two solutions which do it out of the box, ABBYY Recognition Server and ABBYY FlexiCapture. Rec Server is a server-based, high volume OCR tool designed for conversion of large volumes of documents to a searchable format. Although it is available with an API for those types of uses we recommend FlexiCapture. FlexiCapture gives low level control over extraction of data from within table formats including automatic detection of table items on a page. It is available in a full API version without a front end, or the off the shelf version that we market. Reach out to me if you want to know more.
Here are the basic steps that have worked for me. Tools needed include Tesseract, Python, OpenCV, and ImageMagick if you need to do any rotation of images to correct skew.
Use Tesseract to detect rotation and ImageMagick mogrify to fix it.
Use OpenCV to find and extract tables.
Use OpenCV to find and extract each cell from the table.
Use OpenCV to crop and clean up each cell so that there is no noise that will confuse OCR software.
Use Tesseract to OCR each cell.
Combine the extracted text of each cell into the format you need.
The code for each of these steps is extensive, but if you want to use a python package, it's as simple as the following.
pip3 install table_ocr
python3 -m table_ocr.demo https://raw.githubusercontent.com/eihli/image-table-ocr/master/resources/test_data/simple.png
That package and demo module will turn the following table into CSV output.
Cell,Format,Formula
B4,Percentage,None
C4,General,None
D4,Accounting,None
E4,Currency,"=PMT(B4/12,C4,D4)"
F4,Currency,=E4*C4
If you need to make any changes to get the code to work for table borders with different widths, there are extensive notes at https://eihli.github.io/image-table-ocr/pdf_table_extraction_and_ocr.html
I have a database of images. When I take a new picture, I want to compare it against the images in this database and receive a similarity score (using OpenCV). This way I want to detect, if I have an image, which is very similar to the fresh picture.
Is it possible to create a fingerprint/hash of my database images and match new ones against it?
I'm searching for a alogrithm code snippet or technical demo and not for a commercial solution.
Best,
Stefan
As Pual R has commented, this "fingerprint/hash" is usually a set of feature vectors or a set of feature descriptors. But most of feature vectors used in computer vision are usually too computationally expensive for searching against a database. So this task need a special kind of feature descriptors because such descriptors as SURF and SIFT will take too much time for searching even with various optimizations.
The only thing that OpenCV has for your task (object categorization) is implementation of Bag of visual Words (BOW).
It can compute special kind of image features and train visual words vocabulary. Next you can use this vocabulary to find similar images in your database and compute similarity score.
Here is OpenCV documentation for bag of words. Also OpenCV has a sample named bagofwords_classification.cpp. It is really big but might be helpful.
Content-based image retrieval systems are still a field of active research: http://citeseerx.ist.psu.edu/search?q=content-based+image+retrieval
First you have to be clear, what constitutes similar in your context:
Similar color distribution: Use something like color descriptors for subdivisions of the image, you should get some fairly satisfying results.
Similar objects: Since the computer does not know, what an object is, you will not get very far, unless you have some extensive domain knowledge about the object (or few object classes). A good overview about the current state of research can be seen here (results) and soon here.
There is no "serve all needs"-algorithm for the problem you described. The more you can share about the specifics of your problem, the better answers you might get. Posting some representative images (if possible) and describing the desired outcome is also very helpful.
This would be a good question for computer-vision.stackexchange.com, if it already existed.
You can use pHash Algorithm and store phash value in Database, then use this code:
double const mismatch = algo->compare(image1Hash, image2Hash);
Here 'mismatch' value can easly tell you the similarity ratio between two images.
pHash function:
AverageHash
PHASH
MarrHildrethHash
RadialVarianceHash
BlockMeanHash
BlockMeanHash
ColorMomentHash
These function are well Enough to evaluate Image Similarities in Every Aspects.
I would like to know what algorithm is used to obtain an image and get the objects present in the image and process (give information about) it. And also, how is this done?
I agree with Sid Farkus, there is no simple answer to this question.
Maybe you can get started by checking out the Open Computer Vision Library. There is a Wiki page on object detection with links to a How-To and to papers.
You may find other examples and approaches (i.e. algorithms); it's likely that the algorithms differ by application (i.e. depending on what you actually want to detect).
There are many ways to do Object Detection and it still an open problem.
You can start with template matching, It is probably the simplest way to solve, which consists of making a convolution with the known image (IA) on the new image (IB). It is a fairly simple idea because it is like applying a filter on the signal, the filter will generate a maximum point in the image when it finds the object, as shown in the video. But that technique has several cons, does not handle variants in scale or rotation so it has no real application.
Also you can find another option more robust feature matching, which consist on create a dataset with features such as SIFT, SURF or ORB of different objects with this you can train a SVM to recognize objects
You can also check deformable part models. However, The state of the art object detection is based on deep-learning such as Faster R-CNN, Alexnet, which learn the features that will be used to detect/recognize the objects
Well this is hardly an answerable question, but for most computer vision applications a good starting point is the Hough Transform