Release Notes: Input image features are now extracted from rectangular, fixed-location regions in TSL colorspace, additional global image features are calculated by pooling stats from local regions, a new default training dataset is provided, DCT transform is fixed, and a few convenience functions are added for tweaking the dataset and observing the learning rate.
Release Notes: This version adds recognition of close-up and index images, and relies more heavily on global image properties. New feature extractors include Dominant Color features, Color Layout descriptor, ad-hoc Fourier statistics, and local 1st and 2nd order derivative statistics. General DCT and 2D DCT transforms were added. In addition, miscellaneous small fixes, features, and heuristics were implemented.
Release Notes: This is a complete rewrite in the R language for statistical computing. Java and matlab are no longer required. Images are now described by a mix of local and global image properties with some spatial geometric information included. Random Forests are used as the classification engine. The current image categories attempted to be recognized are "nonpron", "pron", "latex/fetish", "japanese cg", "manga", and "bwpron". Detectors for other classes can now be easily built by supplying new training images in separate directories and running a few functions.
Release Notes: This version adds LogitBoosted REPTrees and a Support Vector Machine as user-selectable classifier options. It now morphologically smoothes the visualized predictions, predicts a new class of 'black-and-white content', and uses a more carefully chosen training set for the 'drawn' class.
Release Notes: Images are now evaluated in multiple scales. It attempts to detect blond and brunette hair in addition to the classes present in 0.5.0. The feature extraction routines were tweaked and more training data added. This version requires Matlab, gcc, and Java.
Release Notes: This version is able to predict new types of offensiveness: latex, net stockings, and skin-like patches, in addition to the previously recognized classes "drawn" and "default". This version is a total rewrite and requires Matlab. The predictions are now made on small local image windows. Local Binary Patterns (LBPs) and HSV histograms are used for feature extraction and Random Forests are applied for classification.
Release Notes: The classifier was retrained with a training set augmented with several hundreds of previously misclassified images. The WEKA package is no longer required for operation.
Release Notes: This version is a large overhaul. Previous ad-hoc feature extraction routines were changed to ones based on Minkowski integrals and Haralick texture statistics. Training image set was augmented with hundreds of previously problematic images, totaling over 8000 images. A new model was built in a cost-sensitive fashion using depth 7 decision trees and 95 iterations of LogitBoosting. This experimental branch is straightforwardly usable only on Linux style systems.
Release Notes: This version is coded entirely in Java, making it usable on any platform having Java2 v1.4. In addition to source code, a binary .JAR is now available. The classification engine has been retrained with a more challenging training set.