Research On Image Semantic Segmentation Algorithm Based On Convolutional Neural Network

Image semantic segmentation has always been an important research hotspot in the field of computer vision,which plays an important role in image understanding.This paper makes in-depth research on the image segmentation method based on convolutional neural network,aiming at the problems of partial occlusion of objects,uneven illumination intensity and background interference that affect the segmentation results,and proposes three types of image semantic segmentation networks.(1)To solve the problem of incomplete segmentation of large size objects contained in the image by fully convolution network,this thesis proposes the image semantic segmentation model based on contextual structure.The model uses atrous convolution layers to replace part of the pooling layers in fully convolution network,which can keep the increase of the receptive field of the convolution network without losing more details.In addition,the contextual structure module is introduced to extract the global features of the image by large-scale convolution.And uses the multi-scale hole convolution based on residual module to obtain the target information of different scales.Compared with the FCN segmentation results,the accuracy of the image semantic segmentation model based on contextual structure is improved by 4.37% on Pascal VOC2012 dataset.(2)Due to the complexity of the image environment in the actual scene,it is often affected by the illumination intensity,resulting in the loss of the texture and color information of the objects,so it is difficult to obtain the ideal segmentation results.To solve this problem,this paper proposes the image semantic segmentation model based on multi-scale feature fusion.The model adopts a atrous spatial pyramid pooling module based on multi-scale feature fusion.The features of different branches are closely connected so that the features of each branch can receive semantic information of different scales,and obtain richer high-level semantic features.Then,the low-level features extracted by deep convolution neural network are added to the decode.By fully fusing high-level features and low-level features,image segmentation accuracy can be effectively improved.Compared with the segmentation results of Deep Lab v3,image semantic segmentation model based on multi-scale feature fusion achieves a2.82% accuracy improvement on the PASCAL VOC2012 dataset,and a 1.12%accuracy improvement on the ADE20 K dataset.(3)In the process of image segmentation,duo to the complex image environment,the large number of object categories contained in each image and the different size of objects,it is easy to cause different degrees of over segmentation and under segmentation.This thesis studies image semantic segmentation methods based on feature segmentation and attention mechanism,the feature segmentation module is introduced into the low-level features,and uses different size convolution kernels to extract features of different scales.Then,different scale features are fused by channel splicing to obtain features with different scale information.In addition,the model proposes a decoder module based on attention mechanism to fully integrates high-level features and low-level features,highlight the important feature information and suppress other useless information.Compared with the segmentation results of Deep Lab v3+,the segmentation results of image semantic segmentation model based on feature segmentation and attention mechanism on PASCAL VOC2012 and ADE20 K are improved by 1.43% and 0.78% respectively.