| Image segmentation,as one of the most important research topics in the field of artificial intelligence,is the critical pre-processing of image recognition and computer vision.Semantic segmentation,as one of the sub-problems of image segmentation,is the main goal of this paper.Semantic segmentation is widely used in high-level visual inference such as image analysis,object detection and autonomous driving,as well as low-level image synthesis and image enhancement.With the development of deep learning,the research of semantic segmentation algorithms based on neural networks has attracted widespread attention.But deep learning requires enormous computational resources which is not suitable for the situation where computational resources are limited,and there is still no rigorous mathematical proof of interpretability.The traditional method based on conditional random field(CRF),which solves the segmentation problem within the framework of the probability graph model,not only requires less computational resources,but also has rigorous interpretability.In the current CRF study,they usually use a fixed compatibility matrix in potential energy,which leads to a situation where abnormal object appears in the segmentation result.This paper proposes an adaptive compatibility matrix with different compatibility penalties for different labels of different data sets.In order to reduce the high computational complexity of the fully connected CRF,we extend the adaptive compatibility matrix to the superpixel CRF,and combine the label cost as a direct punishment for abnormal objects,which not only reduces the computational complexity,but also obtains higher segmentation performance.The main contributions in this article are as follows:1.This paper proposes a CRF algorithm combined with adaptive compatibility matrix.Aiming at the problem of existing CRF segmentation algorithms,the concept of adaptive compatibility matrix is proposed,so that the compatibility penalty between any two labels is not a fixed 0,1 value,but from the dataset The adaptive values are learned,and the solution algorithm of the adaptive compatibility matrix based on the variational method is proposed under the assumption of local independence.Then the learned adaptive compatibility matrix is combined with the fully connected CRF,which greatly alleviates the occurrence of abnormal objects in the original CRF segmentation algorithm.2.This paper proposes a superpixel CRF algorithm combined with adaptive compatibility matrix.Due to the high complexity of the solution of fully connected CRF,the segmentation results cannot be obtained in near real-time,so this paper combines the adaptive compatibility matrix and the superpixel CRF.Starting from the situation of two categories,the corresponding potential energy conversion and solution algorithm is proposed,and then extended to multiple categories.While reducing the calculation time,it alleviates the problem of abnormal objects that also exist in the original superpixel CRF segmentation.3.This paper proposes a superpixel CRF algorithm based on label cost.Since the adaptive effect in the superpixel CRF is acting on the superpixel rather than the pixel in the fully connected CRF,there are still abnormal objects in the superpixel CRF combined with the adaptive compatibility matrix So we propose a new label cost,which is added to the overall energy as a direct punishment for abnormal objects,and proposes a downgrading process for the label cost so that it can be converted into potential energy and paired potential energy.The problem of abnormal objects in superpixel CRFs is further solved with the addition of very little calculation time.The effectiveness of our algorithms have been verified on PASCAL VOC 2012 and MSRC-21,two multi-class semantic segmentation datasets,and the full-CRF combined with the adaptive compatibility matrix compared with the fixed compatibility matrix achieved significant improvements in both mIoU and pixel accuracy,and greatly alleviated the problem of abnormal objects cooccurrence.The superpixel-CRF combining the adaptive compatibility matrix and the label cost greatly reduces the calculation time of the full-CRF,alleviates the problem of abnormal objects co-occurrence in the original superpixel-CRF,and achieves significant improvements in both mIoU and pixel accuracy. |
