بررسي و بهبود در روش هاي رده بندي در تصاوير ويدئويي به منظور تشخيص مفاهيم معنايي

چكيده انگليسي :

Improving the Classification Methods for Semantic Video Concept Detection Marjan Momtazpour momtazpour@ec iut ac ir April 14 2010 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M sc Language Persian Supervisor Mohammad Hossein Saraee Saraee@cc iut ac ir Advisor Maziar Palhang Palhang@cc iut ac ir Abstract With the explosive growth of data including text sound and video the needs for automatic indexing searching organizing and categorizing of these data have been increased as well The majority of these data are generated in multimedia environment which include sounds and videos Recently video concept detection has received lots of attention The main goal of video concept detection is to determine the presence of certain semantic concepts such as Outdoor Airplane Car and Person in a video frame Nowadays great developments in video concept detection have been reported in some domains Classification is one of the most accepted methods applicable for video concept detection While different classifiers could be employed for this purpose SVM is the most popular one SVM classifiers behave well in complex classification problems However the learning phase can be a time consuming process especially when a large number of features should be considered Each extra feature can increase the cost space and time of the classification process Therefore the researchers within the image processing area have great desires to design and implement classifiers with small feature sets In addition there is a need to have a sufficient set of features to achieve high recognition rates Consequently this led us to search for an optimum subset of features from the available ones In video concept detection through machine learning approaches a large number of features have been already considered in the literature However classification based on large number of features not only increase the classification time but also can decrease the performance of the classification system Therefore feature selection can be used as an appropriate solution to decrease the number of features The ambition of this thesis is to improve the behavior of the classification in terms of speed and performance In this thesis we propose a novel framework to select the most important features from the traditional low level ones used in the existing classifiers This work will decrease the time of learning and classification by selecting the most relevant features In addition to the improvement of classification time the performance of the classifier remains almost unchanged For this purpose Genetic Algorithm GA is employed for feature selection After feature selection base classifiers are trained in order to detect the existence of semantic concepts in videos Furthermore fusion is employed to improve the performance of based classifiers In the fusion step the selected features by GA have been used and compared with traditional fusion method The results show that by pruning the non relevant features significant improvements can be achieved in terms of the classifier s metrics especially in Average Precision AP F Score Accuracy Recall Precision and Specificity are the other performance metrics that are measured in simulations Keywords Semantic concept detection Classification Support Vector Machine Feature Selection Genetic Algorithm

ممتازپور، مرجان

بررسي و بهبود در روش هاي رده بندي در تصاوير ويدئويي به منظور تشخيص مفاهيم معنايي

ماشين بردار پشتيبان , انتخاب خصيصه , الگوريتم ژنتيك