Detection of abnormal nuclei in cervical smear images based on visual attention model

A novel idea for detecting abnormal nuclei in cervical smear images is presented. The suspect cells often appear with different externals from surrounding normal ones. Therefore, we are able to find them and focus on processing them instead of segmenting all nuclei in the images. This method combines the bottom-up attention mechanism and the top-down target-driven detection method. We extract both direction and brightness features of the image and construct a saliency map which is linearly combined with the high response area detected using annular template matching method. Then, we use an inhibition-of-return as well as winner-take-all mechanism to detect the regions of interest one by one. This process will be followed by segmentation and recognition of the found nuclei. The satisfactory results on extraction and computing speed show that this model can extract the abnormal nucleus regions without processing other part of the image.