Search

International audience

Indexing video by the concept is one of the most appropriate solutions for such problem. It's based on an association between a concept and its corresponding visual, sound or textual features. This kind of association is not a trivial task. It requires knowledge about the concept and its context. In this paper, we investigate a new concept detection approach to improve the performance of content-based multimedia documents retrieval systems. To achieve this goal, we tackle the problem from different plans and make four contributions at various stages of the indexing process. We first propose a new weakly supervised semi-automatic method based on the genetic algorithm to extract and annotate the video plans for training set. Subsequently, we develop a method to detect the basic concepts. We also deal with the issue of noise reduction when generating visual dictionary (BoVS). The different contributions are tested and evaluated on a big dataset (TRECVID 2015).

Following the technological advances carried out recently, there was an explosion in the number of available videos and their accessibility. This was largely justified by the fall of the acquisition prices of the memory supports and the increase their capacity, which facilitated the storage of a large number of video document databases in the computer systems. To make the exploitation of these collections effective, it is necessary to install tools that facilitate access and handle them. In this paper, we propose a new approach for semantic browsing in a large video collection. Our approach is based, essentially, on a structuring of the video concepts. Indeed, a weight is calculated to measure the membership degree of a concept at different levels (context / concept / video). We developed the Video Navigation System (VINAS) to test our approach. A visual metaphor inspired by tag cloud concept is also set up in our VINAS system to help with the comprehensive semantic video browsing. A user study made on our VINAS system showed that the proposed approach can help the user with the semantic browsing in video collection.

International audience

Semi-automatic or automatic image indexation emerged because manual image indexation is slow and tedious. Generally, this first indexation is used as part of a content-based image retrieval system (CBIR). To have a powerful CBIR system, it is necessary to be concerned with three main facets: 1) the choice of the descriptors (based on shape, colour and texture and/or a combination between them); 2) the process of indexation and finally; 3) the retrieval process. In this work, we focus mainly on an indexing based on genetic algorithm and particle swarm optimisation (PSO) algorithm. We chose an optimal combination of colour, shape and texture (PCM: powerful combination method) descriptors. The fruit of our research work is implemented in a system called image search engine (ISE) which showed a very promising performance. In fact, the performance evaluation of the PCM method of our descriptors combination showed upgrades of the average precision metric from 66.6% to 89.30% for the 'food' category colour histogram, from 77.7% to 100% concerning CCV for the 'flower' category, and from 44.4% to 87.65% concerning the co-occurrence matrix for the 'building' category using the Corel dataset. Likewise, our ISE system showed much more interesting performance compared to what was shown in previous works.

Concept detection has been considered a difficult problem and has attracted the interest of the content-based multimedia retrieval community. This detection implies an association between the concept and the visual content. In other words, the visual characteristics extracted from the video. This includes taking knowledge about the concept itself and its context. This work focuses on the problem of concept detection. For that, several stages are elaborated: first, a method of extraction and semi-automatic annotation of the video plans for the training set is proposed. This new method is based on the genetic algorithm. Then, a preliminary concept detection is carried out to generate the visual dictionary (BoVS). This second step is improved thanks to a noise reduction mechanism. This article's contribution has proven its effectiveness by testing it on a large dataset.

International audience