Image Semantic Segmentation Method Based On Superpixel Classification

Image segmentation is widely used in automobile driverless,medical image diagnosis,security video monitoring and other fields.The accuracy of complex scene image segmentation based on the clustering of pixel distance,color,or texture,is limited;Image semantic segmentation is difficult to realize the accurate annotation of image pixel semantic labels.This master dissertation focuses on image semantic segmentation method based on super-pixel classification based on FCN-8s model.Aiming at the problem of under segmentation and poor real-time performance of SLIC algorithm,a fast SLIC superpixel image segmentation algorithm,namely "F-SLIC algorithm",is proposed.The K-means clustering method based on texture features is used to modify under segmentation,improve pixel sampling and reduce computational complexity.The experimental results show that "F-SLIC algorithm" has 22% faster computing speed,lower under segmentation error rate and higher boundary recall rate,compared with SLIC algorithm.In order to solve that FCN-8s model has low accuracy of image semantic segmentation,an improved FCN-8s model,namely "C-FCN model" is proposed.Replacing part of the convolutional layer of the FCN-8s model with a combined atrous convolution kernels can obtain a larger receptive field.It also achievies the purpose of improving the accuracy of the semantic segmentation results without increasing model parameters.In order to further improve the segmentation accuracy to the target boundary of "C-FCN model",it is combined with "F-SLIC algorithm",which is used to get the target boundary of the training image set,and "C-FCN model" is trained to improve the sensitivity of the model to the target boundary.The full connection conditional random field is introduced to optimize the semantic segmentation results of "C-FCN model",and "FC-FCN model" is established.The experimental results show that the population of sample mean intersection over union on the Pascal VOC 2012 data set of the "C-FCN model" achievies 66.3% and its pixel accuracy achievies 87.8%,which increase by 4.1% and 11.7% respectively,compared with FCN-8s model.For "FC-FCN model",the population of sample mean intersection over union and its pixel accuracy achieve 73.8% and 93.5% respectively,both of which are higher than other models.The visual effect of superpixel image semantic segmentation has been significantly improved.Figure [39] Table [4] Reference [52]...