Illuminant Mapping: An Approach Towards Image Splicing Detection

In recent years,computers have become quite powerful and with its prowess and clandestine image editing software,image editing has come to a new level.Even though image editing comes easy with all these tools,so does image forgery.With the right time and place image forgery can change perceptions of how an event is viewed and its effects on people's life.Thus,image forgery detection is becoming important.Among several image forgeries,combining two or more images to create a whole new different image is a tricky one.It can be said as image composition.In image composition the trickiest part might be matching the color combination of different images.In this thesis an analysis is done on image color inconsistencies and an approach about how to detect it on human face and other splicing based on illuminant mapping.After that two approaches are proposed based on illuminant mapping,one is for regional splicing and the other is for human face forgery.The work done on this thesis is listed below in terms of novelty:1.An image forgery detection method is proposed consisting two modules:a.Regional splicing detection module b.Face splicing detection module.2.Regional splicing detection is done via segmentation on an image as vertical and horizontal bands.Using five GGE algorithms illuminant colors are calculated for these bands.The results are then encoded as feature vectors.Difference between illuminant colors and reference illuminant colors are then classified and later a score is given as fake or original for an image from the output detection map.3.Face splicing detection is done using IIC and GGE extraction methods from an input image to extract illuminant maps.Faces are then detected in these maps without human interaction and paired and features are extracted from faces as telltales.Splicing is detected based on these extracted features.With the methods shown in this thesis and later evaluated,promising results are found for both regional splicing and face splicing detection module.Using a blind approach and SVM models trained beforehand 61%of image pixels have been correctly classified for regional splicing detection module.Face forgery detection module also showed an impressive result using kNN classifiers and color descriptors.In a cross dataset scenario 67%accuracy is achieved for face splicing detection module.Future development is planned for both modules with more advanced options.