Improving semantic image segmentation with a probabilistic superpixel-based fully connected CRF

Deep convolutional neural networks (DCNNs) have been driving significant advances in image semantic segmentation tasks due to their power-ful feature representation for recognition. However, their performance in pre-serving object boundaries is still not satisfactory. Visual mechanism theory in-dicates that image segmentation tasks require not only recognition, like DCNNs, but also local visual attention capability. Considering that superpixel is good at grasping detailed local structure, we proposed a Probabilistic Superpix-el-based fully connected CRF model (PSP-CRF) to refine label assignments as a post-processing optimization method. Firstly, we employ the SLIC segmented algorithm to obtain superpixels and employ the well-known DCNNs to produce coarse prediction probabilistic maps at each pixel. We present an effective strategy based on entropy to convert the pixel-level color, position and texture features to the normalized probabilistic superpixels. Secondly, we construct a fully connected CRF model with the probabilistic superpixels as vertexes in-stead of pixels. Thirdly, We optimize the PSP-CRF model to obtain the final la-bel assignment results by employing a highly efficient mean field inference al-gorithm. The experimental results on the PASCAL VOC segmentation dataset demonstrate that the proposed method can improve the segmentation perfor-mance to 82% mIoU while increasing the computational efficiency.