Τα τελευταία χρόνια, η δημοτικότητα του διαδικτύου αυξάνεται ολοένα και περισσότερο και σε συνδυασμό με την κινηματογραφική βιομηχανία που ανθίζει με γρήγορους ρυθμούς , έχει σαν αποτέλεσμα έναν τεράστιο αριθμό βίντεο κοινής χρήσης στο διαδίκτυο και μια πληθώρα κινηματογραφικών ταινιών, στα οποία έχει άμεση πρόσβαση μεγάλη μερίδα του πληθυσμού, συμπεριλαμβανομένων και διάφορων ευαίσθητων κοινωνικών ομάδων, παραδείγματος χάρη παιδιά και εφήβους. Η προστασία τέτοιων ατόμων αλλά και η επιθυμία γνώσης του περιεχομένου ενός βίντεο δημιούργησε την αναγκαιότητα ανάπτυξης αποτελεσματικών, αυτόματων ανιχνευτών βίας.Στην παρούσα διπλωματική παρουσιάζονται οι μέθοδοι που έχουν προταθεί στο συγκεκριμένο πεδίο. Στην συνέχεια, υιοθετείται μια εκ των μεθόδων και αναπτύσσεται αλγόριθμος, με σκοπό τη μελέτη της απόδοσης του. / In recent years, the popularity of the internet growing more and more.This results a huge number of video sharing on the internet and a plethora of films. A large portion of population has direct access in such videos,including sensitive and different social groups , for example children and adolescents . The protection of such persons and the desire knowing the content of a video, created the necessity to develop efficient , automated violence detectors.In this dissertation we present methods that have been proposed in this field . Then , we have adopted one of the methods and we have developed an algorithm in order to study its accuracy.

The need to quantify distance between two groups of objects is prevalent throughout the signal processing world. The difference of group means computed using the Euclidean, or L2 distance, is one of the predominant distance measures used to compare feature vectors and groups of vectors, but many problems arise with it when high data dimensionality is present. Maximum mean discrepancy (MMD) is a recent unsupervised kernel-based pattern recognition method which may improve differentiation between two distinct populations over many commonly used methods such as the difference of means, when paired with the proper feature representations and kernels. MMD-based distance computation combines many powerful concepts from the machine learning literature, such as data distribution-leveraging similarity measures and kernel methods for machine learning. Due to this heritage, we posit that dissimilarity-based classification and changepoint detection using MMD can lead to enhanced separation between different populations. To test this hypothesis, we conduct studies comparing MMD and the difference of means in two subareas of image analysis and understanding: first, to detect scene changes in video in an unsupervised manner, and secondly, in the biomedical imaging field, using clinical ultrasound to assess tumor response to treatment. We leverage effective computer vision data descriptors, such as the bag-of-visual-words and sparse combinations of SIFT descriptors, and choose from an assessment of several similarity kernels (e.g. Histogram Intersection, Radial Basis Function) in order to engineer useful systems using MMD. Promising improvements over the difference of means, measured primarily using precision/recall for scene change detection, and k-nearest neighbour classification accuracy for tumor response assessment, are obtained in both applications.

This thesis describes the theoretical, experimental, and practical aspects of a noise-limited method for scene-change detection in images. The research is divided into three sections: noise analysis and modelling, dual illumination scene-change modelling, and integration of noise into the scene-change model. The sources of noise within commercially available digital cameras are described, with a new model for image noise derived for charge-coupled device (CCD) cameras. The model is validated experimentally through the development of techniques that allow the individual noise components to be measured from the analysis of output images alone. A generic model for complementary metal-oxide-semiconductor (CMOS) cameras is also derived. Methods for the analysis of spatial (inter-pixel) and temporal (intra-pixel) noise are developed. These are used subsequently to investigate the effects of environmental temperature on camera noise. Based on the cameras tested, the results show that the CCD camera noise response to variation in environmental temperature is complex whereas the CMOS camera response simply increases monotonically. A new concept for scene-change detection is proposed based upon a dual illumination concept where both direct and ambient illumination sources are present in an environment, such as that which occurs in natural outdoor scenes with direct sunlight and ambient skylight. The transition of pixel colour from the combined direct and ambient illuminants to the ambient illuminant only is modelled. A method for shadow-free scene-change is then developed that predicts a pixel's colour when the area in the scene is subjected to ambient illumination only, allowing pixel change to be distinguished as either being due to a cast shadow or due to a genuine change in the scene. Experiments on images captured in controlled lighting demonstrate 91% of scene-change and 83% of cast shadows are correctly determined from analysis of pixel colour change alone. A statistical method for detecting shadow-free scene-change is developed. This is achieved by bounding the dual illumination model by the confidence interval associated with the pixel's noise. Three benefits arise from the integration of noise into the scene-change detection method: - The necessity for pre-filtering images for noise is removed; - All empirical thresholds are removed; and - Performance is improved. The noise-limited scene-change detection algorithm correctly classifies 93% of scene-change and 87% of cast shadows from pixel colour change alone. When simple post-analysis size-filtering is applied both these figures increase to 95%.