Research On Image Semantic Segmentation Algorithm Based On Manifold Regularization

According to the image semantic segmentation methods based deep learning,they usually design the loss function only considering the cross-entropy of a single-pixel between the predicted value and the true value,ignoring the influence of neighboring pixels on the classification result.In response to this problem,we propose a manifold regularized image semantic segmentation algorithm.First,the original image and the segmented image are divided into several sub-image patches of the same size through the sub-image patch division algorithm.Second,through the sub-image patches of the original image and the segmented image,calculating the potential geometric constraint relationship between the input data and the prediction result on the manifold surface.Third,we use the results of manifold constraints to optimize the parameters in the semantic segmentation network.Finally,the image semantic segmentation algorithm of manifold regularization are deployed on three different tasks:fully-supervised semantic segmentation,semi-supervised semantic segmentation and weakly-supervised semantic segmentation.In order to verify the effectiveness of the proposed method,we have done multiple sets of comparative experiments for different segmentation tasks.First of all,for image semantic segmentation of fully-supervise,the experiment has an accuracy value of 78.0%,which is 0.5% higher than the original network,on the Cityscapes dataset.The accuracy value is 69.5%,which is an increase of 2.1% on the PASCAL VOC 2012 dataset.Secondly,for semi-supervised and weakly-supervised image semantic segmentation,the experimental accuracy on the PASCAL VOC 2012 data set is 48.40% and 50.0%,which are3.7% and 1.1% higher than the original network.Experimental results prove that the proposed algorithm optimizes the effect of semantic segmentation.Image semantic segmentation algorithm of manifold regularization captures the context information in the image without increasing the computational complexity of the inference network,which reduces the loss of the intrinsic structure during the network forward calculation process and achieves better segmentation accuracy.This algorithm is suitable for a variety of different segmentation tasks and has certain practicability.The paper has 23 pictures,9 tables,and 64 references.