Research On Semi-supervised Image Semantic Segmentation Based On Convolutional Neural Network

Image semantic segmentation is one of the basic visual tasks,because its goal is to classify images at the pixel level,which greatly assists and improves other visual tasks.The semantic segmentation task usually requires pixel-level annotation data,which take considerable time and labor costs.This thesis implements semi-supervised semantic segmentation based on convolutional neural networks by introducing generative adversarial structures.The main work of this article is as follows:First of all,in order to solve the problems of multi-scale feature fusion and the low accuracy of the existing semi-supervised generative adversarial network when solving segmentation tasks,this thesis proposes an improved semi-supervised network based on codec adversarial structure.The Deeplab segmentation network is used as the generator's reference network,and a decoding module is designed on this basis to form a codec generator.At the same time,a global average pooling layer is introduced in the fully convolutional discriminator to fuse the global information of the image.Secondly,in view of the many factors that the network is affected by during the training process,this thesis verifies the performance of the semi-supervised network.The performance of the generator,discriminator and overall network are verified separately,and the final network training scheme is summarized based on the experimental results and the segmentation results are compared and analyzed.The experimental results on the PASCAL VOC dataset show that the improved semi-supervised network in this thesis has achieved certain improvements in segmentation accuracy and effect.In short,semi-supervised learning is an important research direction to solve the task of semantic segmentation in the future,and has far-reaching significance in theoretical research and practice.This thesis proves the effectiveness of the semi-supervised segmentation network based on encoder-decoder adversarial structure through experiments.