Research On Semantic Segmentation Based On Multi-scale Information Fusion

Computer vision has received increasing attention from scholars in deep learning region because of its far ranging applications.Semantic segmentation,as one of the most important areas,has also been flourishing.Many emerging industrys have put forward higher demands for semantic segmentation,the not only for accuracy but also for efficiency.Traditional image segmentation can only distinguish background from target as set,but cannot cope with more complex vision tasks.However,these requirements fit with the domain and application goals of deep learning in computer vision,including semantic segmentation and scene parsing.Therefore,semantic segmentation based on CNN is developing rapidly,achieve pixel-level end-to-end segmentation tasks,that is to say that labeling pixels in an image with different colors to divide them into different regions.There are diverse targets in images,differences in size and lighting conditions of the same object,and similar features among different objects,which increase the difficulty of semantic segmentation and therefore require a wider range of feature information to complete the segmentation.Convolutional neural networks have strong local perception ability,but are poor in acquiring global features.At the same time,due to network layers more and more deeper,the resolution of the feature map gradually decreases,and the extracted features change from specific features such as tread and tinct to increasingly feature information,thus lacking detailed information The loss of detail information leads to rough segmentation results and incomplete object contours.To address these problems,this thesis use CNN on Deeplab,raise a novel method,named Residual Blocks and Multi-granular Feature Extraction Networks(RMF-Net),based on DeeplabV3.The network uses ResNet as the base network,and designs the class residual structure to extract detailed information and global information at the same time.The main contributions are:(1)adding Fine-grained Feature Extract Module(FFE)in the middle of the underlying network,increasing the feature expression through the residual structure,and merging the residual structure to increase the perceptual field of the network,while maintaining a similar computational consumption as the conventional convolution.(2)Global Feature Extract Module(GFE)is added after the Fine-grained Feature Module to obtain more extensive feature information by calculating the index of similarity on all point in the feature map,which solves the problem that the convolution operation only obtains local The problem of obtaining only local features by convolution operation is solved.The problem of detail loss and inter-class confusion is solved by obtaining granularity information in both detail and global dimensions through two modules.The method is verified on public data,and comparing other methods of semantic segmentation to prove the advantages of this thesis's method.The method in the above improves the results,but there are still some shortcomings,so it is further improved on the proposed RMF-Net.When fusing feature maps of different stages,the use of dimensional splicing is too rigid and does not take into account the consistency of feature maps of different stages,so the Feature Fusion Moudle(FFM)is proposed.The feature maps are fused with each other by stitching,convolution and global pooling operations,and the obtained feature maps have consistent information,which makes up for the defect of rigid fusion.In addition,the computational consumption of conventional convolution is too large for model deployment to mobile or embedded devices,so the conventional convolution is improved to reduce the computational consumption of the network while obtaining an effective feature map.Though the experiment in public datasets,modified RMF-Net can further improve network results.