Research On Semi-supervised Image Semantic Segmentation Method Based On Full Convolution Network

Image semantic segmentation is used in the task of visual scene understanding in life.With the rapid development of society,the computer performance and hardware performance are also improving day by day,and people’s requirements for the accuracy of image recognition algorithm are also increasing.How to automatically obtain high-quality labeled images from labeled images has become a research hotspot.Because the scale of unlabeled image is much larger than that of labeled image,it is difficult to improve the classification performance of deep neural network when the labeled information is limited.Therefore,the semi supervised image semantic segmentation algorithm is selected.Because of the characteristics and advantages of the semi supervised network,some pixel level label images and a large number of coarse level images can be combined to form the input data set.Therefore,this paper studies the semi supervised semantic segmentation problem by using the generation confrontation network.The specific research contents are as follows:Firstly,the research of generating countermeasure network based on codec.For the semi supervised segmentation problem,the operation methods using generative network have a common disadvantage,that is,there are insufficient multi-scale feature fusion and insufficient discrimination accuracy.In this paper,the improved codec structure is integrated into the generative countermeasure neural network.The main network is based on Deep lab,and the decoding module is added to the generative network structure to form the codec structure,which is used to ensure the semantic information of small objects.At the same time,a global average pooling layer is added in the discrimination network,which combines the deep information with the global average pooling layer to retain the semantic information in the image.Secondly,because the system will be disturbed by many factors in the process of image processing and output,experiments are carried out to compare the overall performance difference of the network,and the experimental parameters with the best effect are retained to ensure the best segmentation effect of the training scheme.The result comparison on Pascal VOC 2012 data set shows that the segmentation accuracy brought by the network structure adopted in this paper is more accurate.In short,in the semantic segmentation task,the semi supervised network model makes the semi supervised learning an important part of the segmentation task with its unique advantages,which is of far-reaching significance in practical application and theoretical research.Through comparative experiments,this paper verifies the effectiveness of the semi supervised segmentation network with codec.