Python implementation of FALCON: Feedback Adaptive Loop for Content-Based Retrieval
halcon is a python implementation of the Feedback Adaptive Loop for Content-Based Retrieval (FALCON) algorithm as described in
- Leejay Wu, Christos Faloutsos, Katia P. Sycara, and Terry R. Payne. 2000. FALCON: Feedback Adaptive Loop for Content-Based Retrieval. In Proceedings of the 26th International Conference on Very Large Data Bases (VLDB '00), Amr El Abbadi, Michael L. Brodie, Sharma Chakravarthy, Umeshwar Dayal, Nabil Kamel, Gunter Schlageter, and Kyu-Young Whang (Eds.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 297-306.
FALCON is, as described in the article abstract, "a novel method that is designed to handle disjunctive queries within metric spaces. The user provides weights for positive examples; our system 'learns' the implied concept and returns similar objects."
pip3 install --user halcon
There is only one method that you need to know about
halcon.search.query(good_set, candidates, alpha=-5, metric='euclidean', normalization='zscore', debug=False)
Here is a brief description of each of the input arguments
candidatesare two lists of lists where each member of both lists has the same shape.
record = [ <identifier>, <initial_score>, <feature_vector>]
For example in
wine.py, I download a CSV file where the first
feature_vectorlooks like this
and then I modify it like this
good_set =  identifier = 'wine00' initial_score = 1 feature_vector = [1,14.23,1.71,2.43,15.6,127,2.8,3.06,.28,2.29,5.64,1.04,3.92,1065] good_set.append([identifier, initial_score, feature_vector])
For more information about the definition of the initial score, please refer to the article. In all my examples I use a initial score of 1, that is, all images have the same weight. The identifier should be unique (though not enforced), so you can tell images apart. This package assumes every object is represented by a feature vector. Feature calculation and feature selection is beyond the scope of this package. There are many feature calculation/machine learning packages out there that you might find useful, like
alpha. For more information about alpha, please refer to the article. The recommended value by the paper is -5, which is the default value used in this package.
metric. In the research article, a measure of distance
dis used to calculate the distance between two feature vectors. The default value is
euclidean(Euclidean distance) and other supported metrics are 1)
cityblock(Manhattan distance) and 2)
normalization. Feature normalization option. Default is
zscore. Alternative option is
debug. If debug flag is on, then it should print more information about the calculation as they happen.
$ python examples/iris.py This example uses the iris dataset from Machine Learning Repository Center for Machine Learning and Intelligent Systems http://archive.ics.uci.edu/ml/datasets/Iris I will use the first feature vector as my query image [[0, 1, array([ 5.1, 3.5, 1.4, 0.2, 1. ])]] And I will use the rest of the feature vectors to find the most similar images Now notice that feature vector with iid1 has the same values iid0 [1, 1, array([ 5.1, 3.5, 1.4, 0.2, 1. ])] So I expect that if FALCON is working correctly, then iid1 should be the top hit! Elapsed time: 0.0221660137177 seconds Ranking Identifier Class Score --------- ------------ --------------- ----------- 0 1 Iris-setosa 0 1 28 Iris-setosa 1.27788e-43 2 5 Iris-setosa 2.40121e-40 3 29 Iris-setosa 2.40121e-40 4 40 Iris-setosa 5.83391e-40 5 8 Iris-setosa 7.04398e-39 6 18 Iris-setosa 1.1259e-35 7 41 Iris-setosa 1.51906e-34 8 50 Iris-versicolor 6.99696e-34 9 37 Iris-setosa 1.09221e-32 10 12 Iris-setosa 1.22203e-32 11 49 Iris-setosa 2.05046e-32 12 11 Iris-setosa 4.25801e-31 13 21 Iris-setosa 6.55842e-31 14 47 Iris-setosa 5.54098e-29 15 36 Iris-setosa 7.93943e-29 16 7 Iris-setosa 2.16985e-28 17 20 Iris-setosa 4.23544e-28 18 25 Iris-setosa 1.67453e-27 19 3 Iris-setosa 2.40919e-27 Do the top results in the list above belong to the same class as the query image? If so, then SCORE! It seems to work.
$ python examples/wine.py This example uses the wine dataset from Machine Learning Repository Center for Machine Learning and Intelligent Systems http://archive.ics.uci.edu/ml/datasets/Wine I will use the first three feature vectors as my query wine set And I will use the rest of the feature vectors to find the most similar images Elapsed time: 0.0280928611755 seconds Ranking Identifier Score --------- ------------ ----------- 0 wine1 0 1 wine2 0 2 wine3 0 3 wine21 2.77663e-05 4 wine30 0.000629879 5 wine23 0.00252617 6 wine49 0.00318536 7 wine57 0.00456123 8 wine36 0.0152067 9 wine39 0.0197516 10 wine58 0.0243848 11 wine9 0.024467 12 wine55 0.045762 13 wine24 0.046893 14 wine7 0.113906 15 wine45 0.188355 16 wine27 0.201802 17 wine41 0.206469 18 wine31 0.288536 19 wine56 0.291853
Bugs and Questions
To submit bugs about the source code visit
To submit bugs about the documentation visit
For any other inquiries visit those links as well.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.