Pascal Bertolino

Multiresolution segmentation using the irregular pyramid

Segmentation in regions may be achieved using pyramidal methods that provide hierarchical partitions of the original image. The irregular pyramid is a stack of adjacency graphs of the regions in which the shape of the regions is content based. Using the hierarchical decomposition of an image given by this pyramid, a multiresolution oriented process is proposed. Many segmentations at different resolutions are obtained by changing the value of a scale parameter. Thus, each entity may be extracted from the pyramid at its best resolution.

Similarity-based and perception-based image segmentation

In this paper, we present a segmentation method that provides perceptually relevant partitions without any a priori knowledge of the image content: first a local homogeneity analysis detects the image areas that have to be segmented. Then segmentation using a similarity criterion is locally performed. At last, segmented regions are grouped using Gestalt criteria. The whole method is presented in a hierarchical framework.

Sensarea: An authoring tool to create accurate clickable videos

We present a user-friendly software application that can be used in a post-production environment and that allows to automatically or semi-automatically perform spatiotemporal segmentation of video objects. The central element in the application is a local pyramid-based segmentation algorithm. Several tools are provided to interactively guide or correct the automatic process when necessary. The masks of the extracted objects can be exported in a Flash or XML vectorized format and can be synchronized to the original video for many applications such as clickable videos.

Moving object extraction with a localized pyramid

We present a tracking initialization method that combines a rough extraction of moving objects and a refined segmentation of their contours. The extraction of moving objects is obtained with a classical global motion compensation. To obtain accurate contours of the objects, a spatial segmentation is performed with an original localized graph pyramid that focuses the segmentation process either on the object areas or on their borders.

Voronoi Pyramids Controlled by Hopfield Neural Networks

We present a novel method for irregular pyramid construction. The contributions of this paper are twofold: (1) Instead of starting with the original regular pixel grid, we first apply an adaptive Voronoi tessellation to the image. This provides the advantage that the number of cells at the bottom level of the pyramid is already reduced as compared to the number of pixels of the original image. (2) For the construction of the irregular pyramid, we present a Hopfield neural network which controls the decimation process. The decimation by Hopfield networks includes naturally the concept of the adaptive pyramid. The validity of our approach is demonstrated by several examples in image segmentation.

Mathematical modeling of the LCD response time

— Techniques to reduce LCD motion blur are extensively used in industry and they depend on an inherent LCD parameter: response time. However, normative response time is not a sufficient reference to improve LCD performance and all the gray-to-gray response-time quantities are required to obtain good improvement. However, measuring and gathering all the gray-to-gray transitions takes an excessive amount of time. Consequently, we propose a novel LCD model to simulate as well as compute gray-to-gray transitions (response time and behavior) from a reduced measurement set in order to decrease the response-time measurement.

Detecting people in videos for their immersion in a virtual space

This paper presents an application developed to extract people in indoor or outdoor environments using a fixed camera. People are then inserted, in real time, into a dynamic comics environment in which they interact as actors or/and witnesses. The originality of our method relies on two main points. First, the change detection method combines two kinds of masks: region and contour masks. Then, in order to detect people who are moving or not, a reference image is introduced. It is built and updated, taking into account not only local changes in the sequence, but also global, either slow or fast, illumination changes that often occur when the sequence lasts several hours. Results show that people are well and quickly extracted.

Label prediction and local segmentation for accurate video object tracking

This paper presents an approach dedicated to accurately track one or several semantic objects in a video sequence. The accurate tracking of the partition object boundary is obtained by a label prediction. This prediction is performed thanks to motion vectors obtained with two different block-matching uses. In the predicted partition, a local segmentation is necessary only where matching failed and close to the predicted boundaries, in order to get the most accurate boundaries. This local segmentation is then followed by a classification step. During the classification a backward projection is used to assign or not a region to a given object.

Foreground object segmentation for moving camera sequences based on foreground-background probabilistic models and prior probability maps

This paper deals with foreground object segmentation in the context of moving camera sequences. The method that we propose computes a foreground object segmentation in a MAP-MRF framework between foreground and background classes. We use region-based models to model the foreground object and the background region that surrounds the object. Moreover, the global background of the sequence is also included in the classification process by using pixel-wise color GMM. We compute the foreground segregation for each one of the frames by using a Bayesian classification and a graphcut regularization between the classes, where the prior probability maps for both, foreground and background, are included in the formulation, thus using the cumulative knowledge of the object from the segmentation obtained in the previous frames. The results presented in the paper show how the false positive and false negative detections are reduced, meanwhile the robustness of the system is improved thanks to the use of the prior probability maps in the classification process.

Edge detection for biomedical images: A self-adaptive and randomized operator

Most of the time segmentation methods involving a contour approach need several thresholds. This arises many problems in automating the process for biomedical imagery (such as morphometry). We give some examples of well known operators in section II. In section III, we present our algorithm which is a randomized operator without any threshold and we describe how to implement it and how to choose criteria in order to obtain the contour image. We finally give some improvements in section IV, and we conclude by the perspectives on this new approach.