(Springer, 2018) Amrane, Abdesalam; Meziane, Abdelkrim; Boulkrinat, Nour El Houda
Image segmentation for object detection is one of the most fundamental problems in computer vision, especially in object-region extraction task. Most popular approaches in the segmentation/object detection tasks use sliding-window or super-pixel labeling methods. The first method suffers from the number of window proposals, whereas the second suffers from the over-segmentation problem. To overcome these limitations, we present two strategies: the first one is a fast algorithm based on the region growing method for segmenting images into homogeneous regions. In the second one, we present a new technique for similar region merging, based on a three similarity measures, and computed using the region adjacency matrix. All of these methods are evaluated and compared to other state-of-the-art approaches that were applied on the Berkeley image database. The experimentations yielded promising results and would be used for future directions in our work.
