Font Size: a A A

Research On Image Semantic Segmentation Algorithm Based On Deep Learning

Posted on:2021-03-10Degree:MasterType:Thesis
Country:ChinaCandidate:Y WangFull Text:PDF
GTID:2392330611489665Subject:Intelligent Building
Abstract/Summary:PDF Full Text Request
In recent years,more and more intelligent buildings use security,access control and other intelligent systems.And image semantic segmentation plays an important role in these intelligent systems.For example,it is used in target recognition in security system and face recognition in access control system.Image semantic segmentation is a technology,which combines image segmentation and image recognition to achieve the purpose of classifying each pixel accurately.It can also be seen in aerospace,military applications,industrial manufacturing,medical diagnosis,unmanned driving,image retrieval and clothing matching.Therefore,it is necessary to continue to study the subject of image semantic segmentation.Through reading a large number of domestic and foreign scholars' literature,this paper studied the algorithm of image semantic segmentation.In the traditional algorithm and deep learning algorithm,selected the deep learning algorithm for in-depth study,which mainly uses two kinds of network models: FCN and GAN.The main research work of this paper is as follows:(1)This paper proposed a new image semantic segmentation algorithm called SCFCN by combining traditional image segmentation with deep learning method.This paper improved the vgg19 in the FCN,added the super-pixel segmentation and edge detection segmentation in the traditional image segmentation to the segmentation process of vgg19,and introduced the efficient dilated pyramid module,adjusted the skipping structure of FCN,so that the low-level features of the image obtained by super-pixel segmentation and edge detection segmentation were fully fused with the high-level features of the network output to extract the multi-scale features of the image,finally,the result of pixel classification was obtained by softmax classifier layer.TakingMIoU as the evaluation standard,which was 6.9% higher than the original FCN,and the regional accuracy also increased by 2.4%.(2)Based on SCFCN,this paper proposed a new GAN model SCAGAN for image semantic segmentation.In SCFCN,the pyramid pool of dilated space in deeplabv2 was introduced and transformed into the encoder-decoder structure as the generation model of GAN.In order to improve the segmentation accuracy of the model,this paper used the antagonistic learning between the generated network and the discriminative network and verified on the data set Pascal voc 2012,at last,it found that the MIoU reached 70.10%,which was improved compared with FCN and SCFCN.
Keywords/Search Tags:Deep learning, Image semantic segmentation, SLIC, Canny, Convolution neural network, FCN, GAN
PDF Full Text Request
Related items