Deep Learning And Semi-Supervised Learning Based Image Semantic Segmentation Tech Nology Research

With the rapid development of intelligence and automation,image semantic segmentation,as a basic computer vision technology,plays a key role in many application scenarios such as automatic driving obstacle avoidance,satellite image recognition and medical imaging image recognition.It provides rich positioning information and category information,which is the basis of many subsequent tasks.With the rise of deep learning,many segmentation models based on neural networks have emerged in the field of image semantic segmentation.The semantic segmentation task of image needs to assign labels to each pixel of the image,so it has very high requirements for accurate positioning.Therefore,it is necessary to design a method for accurate identification and refined positioning to go hand in hand.At the same time,semantic segmentation of image labels has a high labeling cost.Usually,real data sets contain a large amount of data without semantic labels and a small amount of data with semantic labels.It is very important to take advantage of the large amount of information in the data without labels and weak labels.In this paper,based on the neural network of full convolution structure,the refined semantic segmentation module is designed,and multiple groups of experiments and in-depth studies are conducted on the above two problems by combining multiple methods such as weak,semi-supervised,transfer learning and data amplification.In this paper,a hollow convolution combination module named d-link is designed to realize the sensor field amplification and multi-scale feature fusion of the network,and d-linknet and d-unet with d-link structure as the core are designed.The experimental results on a variety of data sets prove that this structure has a significant effect on improving the network segmentation accuracy.On the basis of d-linknet,this paper further designs d-linkbranch containing attention module,and USES global pooling information to carry out semantic information of different levels and channels.In this paper,the effects of weak and semi-supervised learning and data amplification methods on the performance of neural networks are explored.Using the branch structure of d-linkbranch to realize the joint training of weak supervised learning and strong supervised learning.Then,the two methods are combined to realize the weak and semi-supervised joint network training method and obtain better segmentation results.This paper summarizes the universal image morphological and color amplification and testing amplification methods to make full use of semantic label data.