Summary
Semantic analysis of multimedia content is an ongoing research area that has gained a lot of attention over the last few years. Additionally, machine learning techniques are widely used for multimedia analysis with great success. This work presents a combined approach aiming at the semantic adaptation of neural network classifiers in a multimedia framework. Our proposal is based on a fuzzy reasoning engine which is able to evaluate the outputs and the confidence levels of the neural network classifier, using a domain specific knowledge base. The results obtained by the fuzzy reasoning engine are used as input for the adaptation of the network classifier, further increasing its ability to provide accurate classification of the specific content. The improved performance of the adapted neural network is used by a semantic segmentation algorithm that merges neighbouring regions satisfying certain criteria. In that way, fine image segmentation and classification are established.
See the full content of this document
Extract
Semantic Adaptation of Neural Network Classifiers in Image Segmentation
(ProQuest: ... denotes formulae omitted.)
1. IntroductionAutomatic image segmentation has been one of the major problems in the area of image processing and computer vision. For that reason, a plethora of techniques has been proposed in the literature, including feature clustering [1, 2], mathematical morphology [3], graph-based techniques [4, 5]. Furthermore, in many cases machine learning techniques that handle specific aspects related to learning classification or adaptation were used [6, 7, 8, 9]. Lately, the usage of semantic analysis in multimedia applications has gained great attention [10] that is also reflected on recent R&D activities of European IST projects, such as Acemedia, Muscle, K-Space, X-Media and Mesh. More specifically, research effort has been focused on combining certain knowledge about the domain in which an image belongs to, towards a semantically meaningful segmentation. To this aim, Borenstein et al. in [11] proposed the combination of top-down (model-driven) and bottom-up segmentation, where information in the image level can solve ambiguities during the steps of a region-based segmentation process. In [12] a Bayesian network was used to include low- and mid-level features for the classification of indoor or outdoor images, while unsupervised fuzzy classification of regions was used for segmentation purposes in [13]. Finally, spatial information on the regions in an image has been used to reduce the size of the possible solutions, increasing in that way the accuracy in segmentation and object recognition [14].In this paper, we propose an architecture that combines semantic image analysis technologies together with machine learning techniques and adaptation in order to provide improved segmentation performance of static or moving images. The architecture of the proposal is illustrated in Fig. 1.An image or a video frame is initially processed by a hierarchical segmentation algorithm [15] which partitions it in a number of regions that may have a symbolic interp...See the full content of this document
Sponsored links
