New general features based on superpixels for image segmentation learning

Abstract

Segmenting an image is usually one of the major and most challenging steps in the pipeline of biomedical image analysis. One classical and promising approach is to consider segmentation as a classification task, where the aim is to assign to each pixel the label of the objects it belongs to. Pixels are therefore described by a vector of features, where each feature is calculated on the pixel itself or, more frequently, on a sliding window centered on the pixel. In this work, we propose to replace the sliding window by superpixels, i.e. regions which adapt to the image content. We call the resulting features SAF (Superpixel Adaptive Feature). Their contribution is highlighted on a biomedical database of melanocytes images. Qualitative and quantitative analyses show that they are better suited for segmentation purposes than the sliding window approach.