Franck Thollard

From Text Detection in Videos to Person Identification

We present in this article a video OCR system that detects and recognizes overlaid texts in video as well as its application to person identification in video documents. We proceed in several steps. First, text detection and temporal tracking are performed. After adaptation of images to a standard OCR system, a final post-processing combines multiple transcriptions of the same text box. The semi-supervised adaptation of this system to a particular video type (video broadcast from a French TV) is proposed and evaluated. The system is efficient as it runs 3 times faster than real time (including the OCR step) on a desktop Linux box. Both text detection and recognition are evaluated individually and through a person recognition task where it is shown that the combination of OCR and audio (speaker) information can greatly improve the performances of a state of the art audio based person identification system.

A Tabular Pruning Rule in Tree-Based Fast Nearest Neighbor Search Algorithms

Some fast nearest neighbor search (NNS) algorithms using metric properties have appeared in the last years for reducing computational cost. Depending on the structure used to store the training set, different strategies to speed up the search have been defined. For instance, pruning rules avoid the search of some branches of a tree in a tree-based search algorithm. In this paper, we propose a new and simple pruning rule that can be used in most of the tree-based search algorithms. All the information needed by the rule can be stored in a table (at preprocessing time). Moreover, the rule can be computed in constant time. This approach is evaluated through real and artificial data experiments. In order to test its performance, the rule is compared to and combined with other previously defined rules.

Content-Based re-ranking of text-based image search results

This article presents a method for re-ranking images retrieved by classical search engine using key words for entering queries. This method uses the visual content of the images and it is based on the idea that the relevant images should be similar to each other while the non-relevant images should be different from each other and from relevant images. This idea has been implemented by ranking the images according to their average distances to their nearest neighbors. This query-dependent re-ranking is completed by a query-independent re-ranking taking into account the fact that some types of images are non-relevant for almost all queries. This idea is implemented by training a classifier on results from all queries in the training set. The re-ranking is successfully evaluated on classical datasets built with ExaleadTM and Google ImagesTM search engines.

Text detection and recognition for person identification in videos

This article presents a demo of person search in audiovisual broadcast using only the text available in a video and in resources external to the video. We also present the different steps used to recognize characters in video for multi-modal person recognition systems. Text detection is realized using the text features (texture, color, contrast, geometry, temporal information). The text recognition itself is performed by the Google Tesseract free software. The method was successfully evaluated on a broadcast news corpus that contains 59 videos from the France 2 French TV channel.