Research On Image Segmentation Method Based On Self-training And Contrast Learning

Semantic segmentation is an integral part of the computer vision application domain,and it has been widely used in various scenarios,such as smart driving,saliency detection,and artificial intelligence monitoring.Due to the rapid development of deep learning techniques,its application in these areas has become more and more widespread.Meanwhile,supervised training methods have become increasingly complex.Many efforts have shifted to semi-supervised and weakly supervised tasks.The key to semi-supervised semantic segmentation is how to effectively use large amounts of unlabeled data.A common practice is to use labeled data to generate pseudo-labels for unlabeled data,or to use some data augmentation methods.However,the pseudolabels generated by these operations are of low quality and seriously interfere with the subsequent segmentation task.Most of the current weakly supervised semantic segmentation methods are based on category activation graphs to generate pseudolabels,which can lead to problems such as imprecise edge segmentation and inaccurate localization of pseudolabels.To address the above difficulties,two different image segmentation algorithms are proposed in this paper,which are semi-supervised and weakly supervised semantic segmentation based algorithms.The specific work in this paper is as follows:1.semi-supervised semantic segmentation approach based on self-training and iterative training strategies:In this work,this paper finds that the fusion of self-training and iterative training strategies can significantly improve the quality of the generated pseudo-labels.High-quality pseudo-labels can minimize the errors in the model training process and thus have a crucial impact on the semantic segmentation task.In this paper,we further classify the unlabeled data into two confidence types: reliable and unreliable images.The approach in this paper does not discard unreliable images because they tend to occupy a portion of the dataset,which is crucial for the network to learn contextual information about the entire dataset.Based on this idea,a semi-supervised semantic segmentation framework is proposed in this paper.In this paper,we explore the solution to the related problem through experiments on two public benchmark datasets,Pascal VOC 2012 and Cityscapes.The experimental results show that the approach in this paper is ahead of some state-of-the-art methods.We demonstrate the practicality and efficiency of our proposed technique after various combinations of ablation experiments.2.Weakly supervised semantic segmentation methods based on contrast learning:Weakly supervised semantic segmentation is a complex computer vision task that requires the classification of images and requires the provision of customized features.Traditional(Class Activation Mapping,CAM)methods can only simulate the approximate position of objects in an image and cannot provide accurate recognition.Therefore,developing more accurate weakly supervised semantic segmentation techniques is an important challenge for current researchers.In this paper,we find that there are obvious differences between the semantic information of foreground objects and background in feature space representation: when foreground objects have a similar appearance,they are closer to each other;and when background objects have similar color or texture,they are also closer to each other.Based on the above relationships,this paper proposes a novel deep learning framework with contrast learning and simulation learning mechanisms for weakly supervised semantic segmentation.The approach in this paper uses a new contrast loss,which forces the network to preferentially learn category-independent foreground and background separation activation maps.In addition,this paper proposes an attention modulation module(Channel and Spatial Block(CSB))that focuses on rearranging the distribution of feature importance in terms of channel and spatial order.CSB not only helps to make full use of the captured location information so that regions of interest can be captured accurately,but it also captures inter-channel relationships efficiently.The approach in this paper achieves a new state-of-the-art performance on the Pascal VOC 2012 public benchmark dataset,surpassing not only current methods trained with image-level supervision but also some methods that rely on stronger supervision.The experiments also show that the scheme in this paper is plug-and-play and can be used with other methods to improve their performance.