Nafaa NACEREDDINE

In this paper, we address the problem of estimatingthe parameters of Asymmetric Generalized Gaussian Distribution(AGGD) using three estimation mehods, namely, Maximum LikelihoodEstimation (MLE), Moment Matching Estimation (MME)and Entropy Matching Estimaion (EME). For this purpose, thesemethods are applied on an unimodal histogram fitting of animage corrupted with AGGD noise. Experiments show that theeffectiveness of each method comparatively to the other onedepends on the variation range of the shape factor.

Abstract—In histogram-based image segmentation, the AsymmetricGeneralized Mixture Model (AGGMM) is a powerful toolto fit accurately the real images histograms by handling, amongothers, any asymmetry of the modes. However, the ExpectationMaximization (EM) algorithm, used for the estimation of themixture model parameters, is known to be very sensitive tostarting conditions and can lead to erroneous segmentationresults when the initialization is not adequate. In this paper, wepropose a new method to initialize the AGGMM. This methodis based on geometrical aspects of the histogram. First experimentationsimplying synthetic images generated by AsymmetricGeneralized Mixture Distribution (AGGD) model, reveal a goodrecovering of the input mixture parameters when applying theproposed method. Second experimentations involving real-worldimages have shown, how the initial parameters computed bythe proposed method permit to achieve better histogram fittingwith less EM algorithm running time in comparison to otherinitialization methods.

Quality control by artificial vision with applications in industrial manufacturingis a challenging task due to the significant variability of surface defects.In this work, we propose to use content based image retrieval CBIR to manage the large data produced by surface inspection systems. The performance of the CBIR system was evaluated using textural features extracted from NEU database that collects six kinds of surface defects of the hot-rolled steel strip. Different similarity measurements were used to retrieve the most similar images to the query image. The effectiveness of Wavelet based local binary patterns WLBP features was shown in the experimental results for the retrieval of surface defects. WLBP features using Chi square distance achieved the highest retrieval values compared to LBP, GLCM, and EHD features.