Research Of Semantic Segmentation Based On Fully Convolutional Networks And System Implementation

Semantic segmentation is an essential topic in computer vision widely used in different applications such as autonomous driving and medical image analysis.It labels each pixel of an image with a semantic class,perforaming pixelwise classification.This thesis proposes two improved semantic segmentation networks,applied respectively to 2D scene images and 3D brain MRI images,namely,scene parsing and MRI segmentation.This thesis mainly contains:(1)The multi-stage convolution and pooling in the semantic segmentation network will reduce image size and lead to losing location information.Moreover,it will be less accurate for semantic segmentation when there are too many classes.This thesis proposes a novel network by combining ResNeSt and PPM.The former is based on grouped convolution and attention mechanism,while the latter based on multi-scale feature fusion.This network extracts and fuses relatively essential features in different convolution groups using the attention mechanism,reducing the loss of image features during downsampling and obtaining more accurate location information.Experiments on the ADE20K dataset show that the MIoU achieves 41.54%,and the PA is 80.21%,which is 0.81%higher on MIoU and 0.44%higher on PA than the PSPNet.(2)Because of the large size of 3D brain MRI images,each GPU can only take one or two samples,which leads to small batch size.That may easily cause inaccurate estimation of data distribution during batch normalization and result in internal covariate shift.This thesis replaces batch normalization(BN)with group normalization(GN)to improve the 3D U-Net.GN separates channels into groups and normalizes within each group along the channel dimension,which helps with the above problems caused by the small batch size.Experiments on the ADNI dataset show that the average Dice coefficient of the improved 3D U-Net is 0.8261,higher than 0.8151 of the original 3D U-Net.(3)With the two networks proposed in this thesis,a semantic segmentation system is designed and developed,performing semantic segmentation tasks on single or batch 2D scene images and 3D brain MRI images.Also,online 2D and 3D overlay viewers are implemented to simultaneously display the original image and the segmentation result image with varying blend ratios.In addition,the system also contains several management modules which give administrative users a better way to manage all users,tasks,containers,and logs.This thesis designs,implements,and evaluates the proposed 2D and 3D semantic segmentation networks,verifying the effectiveness of proposed networks.Subsequently,both segmentation runners are wrapped as HTTP services and packaged into Docker images with their trained models,making distribution and deployment easier.Finally,the semantic segmentation system makes it fast and convenient for users to invoke 2D scene parsing and 3D MRI segmentation services.