Self-Similarity Descriptor and Local Descriptor-Based Composite Sketch Matching

Abstract

Composite sketching belongs to the forensic science where the sketches are drawn using freely available composite sketch generator tools. Compared to pencil sketches, composite sketches are more effective because it consumes less time. It can be easily adopted by people across different regions; moreover, it does not require any skilled artist for drawing the suspects faces. Software tool used to generate the faces provides more features which can be used by the eyewitness to provide better description, which increases the clarity of the sketches. Even the minute details of the eyewitness description can be captured with great accuracy, which is mostly impossible in pencil sketches. Now that a composite sketch is provided, it has to be identified effectively. In this paper we have analyzed two state-of-the-art techniques for composite sketch image recognition: Self-similarity descriptor (SSD)-based composite sketch recognition and local descriptors (LD)-based composite sketch recognition. SSD is mainly used for developing a SSD dictionary-based feature extraction and Gentle Boost KO classifier-based composite sketch to digital face image matching algorithm. LD is mainly used for multiscale patch-based feature extraction and boosting approach for matching composites with digital images. These two techniques are validated on FACES and IdentiKit databases. From our analysis we have found that SSD descriptor works better than LD. Using SSD method we obtained the results for FACES (ca) as 51.9 which is greater when compared to LD which gives a result of 45.8. Similarly, using SSD, values of 42.6 and 45.3 for FACES (As) and IdentiKit (As), respectively, are obtained which are much better than the values of 20.2 and 33.7 for FACES (As) and IdentiKit (As), respectively, using LD method.