Due to the massive use of Internet together with the proliferation of media devices, contentbased image retrieval has become an active discipline in computer science. A common contentbased image retrieval approach requires that the user gives a regular image (e.g, a photo) asa query. However, having a regular image as query may be a serious problem. Indeed, peoplecommonly use an image retrieval system because they do not count on the desired image.An easy alternative way to express what the user is looking for is by giving a line-basedhand-drawing, a sketch, leading to the sketch-based image retrieval. This kind of query isalso supported by the fact that emerging touch screen based technology is becoming morepopular, allowing the user to make an sketch directly on the screen.In this work, we propose methods aiming to deal with the sketch based image retrievalproblem. The rst method is a global method that computes a histogram of local orientationsbased on the gradient squared technique.This proposal method exhibits outperformingresults in comparison with current global methods.To the best of our knowledge, any ofthe current sketch based image retrieval approaches have not exploited yet the structuralcomposition of an image (or sketch) represented by a set of strokes. Therefore, the secondproposed method is a local approach that exploits the structural information given a sketchrepresentations. This method is based on representing an image by a set of simple shapescalledkeyshapes.Each keyshape belongs to one of ve keyshape classes. The set of keyshapescomputed over an image de nes the image structure representing the image in a highersemantic level than interest points do.To this end, we propose a method for detectingkeyshapes that processes stroke pieces to classify it as one of the keyshape classes.We experimentally compare our results with current methods, showing an increase in theretrieval e ectiveness. We also demonstrate that our approach may successfully be combinedwith other leading methods.The reason for that stems from the fact that our method isbased on a novel feature, di erent from those used by the other current methods. We showthat a combination of our method with one of the state of the art signi cantly outperformscurrent SBIR methods, achieving an increase in e ectiveness of almost22%.Finally, in this work we show two applications of our keyshape-based approach. In the rstcase, we adapt our method to be applied in the context of sketch-based 3D model retrieval,achieving competitive results w.r.t.current methods.In the second case, we extend ourmethod for the hand segmentation problem in semi-controlled environments. In particularour keyshape-based approach permits locating a hand in an image in order to improve thesubsequent segmentation stage.PFCHA-BecasDoctorado en Informática147p.PFCHA-BecasTERMINADA