Research On Image Semantic Segmentation Based On Neural Network

With the deepening of deep learning in image research,image semantic segmentation based on deep learning has been widely used in geographical and natural images,medical images,automatic driving and other fields.Image semantic segmentation plays an important role in computer vision.The segmentation results will directly affect the accuracy of subsequent classification or recognition.In recent years,image semantic segmentation has made re-markable achievements in both fully supervised learning and weakly supervised learning.In this paper,we fully investigate the previous image semantic segmentation research,and find that there are two main shortcomings:(1)The segmentation results of small objects in the image are rough.(2)The results of image object edge segmentation are rough.In order to solve the above two shortcomings,this paper proposes to increase the attention in the Encoder-Decoder structure to enhance the correlation of feature channel,increase the cor-relation of feature channel and spatial position in the feature extraction module,adopt the multi-level filter as the upsampling to make the results of small objects in the image more accurate,and add the dense conditional random field to make the objects'boundary more accurate smooth.The main work and contributions of this paper are as follows:(1)In view of the fact that the backbone network does not consider enhancing the correlation of feature channels,the attention is introduced into the backbone network,and the weights of the cross position of the feature graph is enhanced by asymmetric convolution.This method mainly uses the Full Connection Layers and Sigmoid,Re LU activation functions,and does not add extra parameters and too much computation.The results show that Attention Xcep-tion as the backbone network,the MIo U of PASCAL VOC 2012,Cityscapes and zhongwei datasets are increased by 1.60%,1.53%and 1.68%.(2)The feature extraction module and dual attention module are combined as a new feature extraction module.The location attention module reuses the image context information as local features to enhance the correlation between local features and context information;the channel attention module reuses the image channel information to enhance the correlation between channels.The results show that Attention-ASPP is used as feature extraction mod-ule,and the MIo U of PASCAL VOC 2012,Cityscapes and zhongwei datasets are increased by 1.71%,1.64%and 2.03%.(3)Aiming at the rough segmentation result of small objects in the image,a multi-level fil-ter algorithm is proposed as the unsampling method of the Decoder.The multi-level filter divides the whole feature map into equal N~2sub parts for unsampling and superposition fusion to recover the object's contours.This method is easier to capture and recover the lost details in the convolution process than bilinear interpolation and cubic convolution interpo-lation.The results show that the MIo U of PASCAL VOC 2012,Cityscapes and zhongwei datasets are increased by 1.75%,1.68%and 2.45%.(4)In order to solve the problem of unsmooth edge segmentation of image's objects,the Fully Connected Conditional Random Field is used as the post-processing method.In order to achieve the goal of semantic optimization for each pixel,local and global pixel constraints are added to the Fully Connected Conditional Random Field,and the boundary is segmented accurately.The results show that the MIo U of PASCAL VOC 2012,Cityscapes and zhong-wei datasets are increased by 1.70%,1.63%and 2.17%.In order to verify the effectiveness of the improved image semantic segmentation method designed in this paper,the improved algorithm and the current advanced image semantic segmentation algorithm model are tested on the PASCAL VOC 2012,Cityscapes and zhong-wei Datasets.Compared with Deep Lab V3+network,the MIo U of the improved algorithm is improved by 2.3%,3.2%and 3.5%respectively.Experimental data show that,compared with other segmentation methods,the improved algorithm proposed in this paper has im-proved the accuracy of pixel recognition and object boundary segmentation,which strongly proves that the improved algorithm can not only capture image multi-scale information,but also effectively capture information between image channels.