Research On Image Semantic Segmentation Based On DCNN

In recent years,image semantic segmentation has become one of the most active tasks in the field of computer vision.Its goal is to divide image pixels into semantically meaningful regions.With the development of computer hardware,machine learning algorithms are gradually replaced by deep learning methods,and deep learning,especially the method of using deep convolutional neural networks,has achieved great success in semantic segmentation tasks.Image semantic segmentation is often affected by color,texture,and shape,resulting in a compromised segmentation result.The traditional method uses artificial construction features to solve the pixel classification problem,and the effect is very poor.Deep learning uses data training instead of artificial construction,which greatly improves the accuracy.In this context,this paper uses the deep learning framework MatConvNet and TensorFlow to perform pixel classification and semantic segmentation tasks.This paper is based on deep learning.The main work is as follows:1.The first method proposed in this paper first replaces VGN-16 of FCN with ResNet-101 in order to obtain higher precision;then optimizes the upsampling result by using jump connection structure in conv3 and conv4;Final segmentation results.2.The second method proposed in this paper first introduces the DB module of the multi-scale dense network to the WDB module,so that the network can increase the network depth without increasing the parameters;then,the TD and TU operations are added to the encoder and decoder respectively.To improve the WDB and DB modules;finally,combining FC-CRF before output makes the boundaries of semantic segmentation change.3.The setting of the parameters determines whether the algorithm is efficient.In this paper,the setting of each parameter is described in detail,so that the algorithm is most efficient.4.In order to verify the feasibility and effectiveness of the proposed method,training,testing and verification are carried out on two different data sets respectively,and compared with other classical algorithms to show the superiority of the method.