Weakly Supervised Learning Of A Generative Adversarial Nets For Semantic Segmentation

Image semantic segmentation because of its wide application,has very important practical value and theoretical significance,and has been focus of industrial personnel and research scholars.With the rise of big data and cloud computing,the ability of computers to deal with problems has been greatly enhanced,and data collection has become easier.These advances promote the deep learning-based image semantic segmentation algorithm.Since the creation of the deep learning network,it has performed impressively in the field of computer vision.However,fully supervised models represented by convolutional neural networks are datadriven and rely on massive data training.The Pixel-level tag,which requires the labelers to spend a lot of time annotating,needs a high degree of professionalism.Therefore,it takes a lot of manpower and material resources.Weakly supervised image semantic segmentation,which has attracted more and more attention,effectively solves this problem.In view of the limitations of the current weakly supervised image semantic segmentation method,the algorithm in this paper simulates the visual selective attention mechanism and obtains the prediction label used to guide the segmentation network.In addition,the algorithm obtains more comprehensive and accurate feature information by optimizing the structure of the antagonistic network.Finally,the algorithm can get accurate segmentation results of category and fine edge.The main contents and contributions of this article are as follows:In this paper,an algorithm based on generative adversarial networks for the weakly supervised image semantic segmentation is proposed.Firstly,this paper proposes a semantic feature extraction module based on the features of human vision,which includes focused attention and significant attention.In this module,the edge information and position information of the target is obtained by super-pixel segmentation and class activation map fusion,and the prediction label is obtained by conditional random field refinement.Then the algorithm introduces the prediction tag obtained from the semantic extraction module into the generative adversarial network to do supervision information.The algorithm first trains the discriminator with fixed generator parameters,then trains the generator with fixed discriminator parameters,and then iterates until reaching the Nash equilibrium,thus,the segmentation result is obtained.Experiments show that this method effectively obtains pixellevel prediction supervision information from image-level weak tags,reduces the dependence on a large number of data tags and obtains more accurate segmentation results.This paper proposes a weakly supervised image semantic segmentation algorithm based on Improved Generative Adversarial Networks.To deal with the problem of losing part of information through multi-layer network processing,the model starts with the introduction of jump structure,variability convolution and multi-scale,and constructs a generator based on variability convolution and a multi-scale discriminator,which greatly retained global information and local information.Experiments show that this method can effectively slove the problem of imperfect edges caused by the loss of feature information,and the segmentation details are more refined.Experiments show that this method can effectively improve the problem of imperfect edges caused by the loss of feature information,and the local segmentation results are more refined.In summary,the weakly supervised image semantic segmentation method proposed in this paper can establish the relationship between image-level class labels and pixels,capture comprehensive and accurate feature information,and improve the segmentation ability of the network.