Résumé:
In this thesis, we propose two different contributions in the field of computer vision. The first contribution concerns the feature description and matching part, where we proposed an orientation invariant feature descriptor without an additional step dedicated to this task. Wؤe exploited the information provided by two representations of the image (intensity and gradient) for a better understanding and representation of the feature point distribution.
The provided information is summarized in two cumulative histograms and used in the feature description and matching process. In the context of object detection, we introduced an unsupervised learning method based on k-means clustering. Which we used as an outlier pre-elimination phase after the matching process to improve our descriptor precision. In the second contribution, we proposed an edge detector based on a statistical modelization of the image surface. We used two classical and widely used metrics to constitute our detector properties, which are the mean and standard deviation. Our approach was to take advantage from the intensities fluctuation for a better understanding of the image surface. Moreover, our detector is able to highlight pertinent regions in the image. This property has been exploited in the present work, to identify important image contours. Be sides its novelty and efficiency, the main advantage of our detector is its simplicity, which makes it easy to implement in terminals with low processing capacity, it’s also little memory consumer and doesn’t need a training phase which makes it independent from the availability of labeled datasets. Experiments shown the robustness of our descriptor to image changes and a clear increase in terms of precision of the tested descriptors after the pre-elimination phase. Our edge detector shows good performences compared to state-of-the-art detectors.
