The polarimetric imaging consists in acquiring images containing information relating to the polarization of the scattered light from a scene. The objective of this thesis is to use the properties of this type of imaging to enhance the contrast between several objects of interest.Considering the optimization of the contrast between two objects of interest, we demonstrate that, if the time for the measurement is fixed, it is the acquisition of a single image with optimized states in illumination and analysis that achieves the best performance. That is why we have developed an imager that can generate and analyze any polarization state on the Poincaré sphere, using liquid crystal cells. These states can be controlled to modify the contrast in the images and we show that the optimal states" maximizing the contrast depend on the measurement conditions. Specifically, the value of the polarization states maximizing the contrast between two objects interest depends on the measurement noise (noise detector, Poisson noise, Speckle) and also of spatial fluctuations of polarimetric properties in the scene. Improper estimate of the noise source may therefore lead to a significant loss of contrast.We then consider a more complex imaging scenario where the scene can be illuminated non-uniformly. We propose a method of acquisition using all the degrees of freedom provided by our imaging and show that this method can significantly increase the contrast compared to results obtained with other types of polarimetric imaging such as OSC imaging (Orthogonal State Contrast).We then extend our studies to amulti-target case" where more than two objects must be distinguished. In particular, we show that increasing the number of images can degrade the contrast and that there is an optimum number of images to be acquired if one works with a fixed acquisition time.Finally, we propose a method to automate our imaging to optimize contrast by combining iteratively the acquisition of polarimetric images and optimized segmentation algorithm using statistical active contours. The first experimental results demonstrate the advantage of this integration of digital processing algorithms in the core of the image acquisition process.
