Design And Research Of Semantic Segmentation Algorithm For Weakly Supervised Image Based On Semantic Affinity Among Pixels

In the task of deep learning to solve image semantic segmentation,the lack of segmentation tags seriously affects the development and progress of image semantic segmentation technology.In order to reduce the requirements for labeling information,the image semantic segmentation model based on weak supervision information has been fully affirmed by the industry.In the current mainstream weakly supervised semantic segmentation algorithms,most of them use the transfer learning deep convolutional neural network model to obtain the salient regions of the classification network's response to the object category.The pseudo-label mask is further obtained by optimizing the saliency region,which is used as the supervised information training segmentation network model to achieve the goal of semantic segmentation tasks.In this paper,a brand-new image semantic segmentation algorithm is designed.This method only uses image-level tags that are easier to obtain as supervision information without any additional data or annotations.This method learns the semantic intimacy relationship between pixels by mining the semantic category information between neighboring pixels.First,on the basis of the calculated class activation map,the local pixel labels are divided into two types of semantically similar and dissimilar sets.By comparing the training guidance of the loss function,the semantic features of similar pixels are pulled in,and the semantic features of dissimilar pixels are excluded,and a predictive mapping function of the semantic relationship in the latent space between adjacent image coordinates is obtained.Realize the conversion of the intimacy between pixels into learning a globally consistent invariant mapping relationship,which is only related to the potential spatial distribution of neighboring pixels.With the help of this mapping function,accurate prediction of unmarked pixels in saliency regions can be achieved,thereby obtaining a more high-quality pseudo-label mask,and then completing the optimization of image segmentation prediction effects.At the same time,in order to simplify the multi-stage weakly supervised segmentation algorithm,this paper further designs a single-stage segmentation model based on the semantic intimacy relationship between pixels.This method realizes the collaborative optimization of classification,positioning,and segmentation tasks under a backbone network,and can realize the prediction and generation of semantic segmentation results in a single-stage training process.This end-to-end model is more advantageous in solving real-world scenarios.The methods proposed in this paper are all trained and evaluated on the PASCAL VOC 2012 semantic segmentation dataset.In the experimental results obtained by the sufficient comparison experiment,the methods in this paper have achieved certain improvements,which further verify the effectiveness of the text method.Experimental results show that this method has significant advantages over other methods in solving detailed segmentation in complex scenes.