Research On Image Semantic Segmentation Algorithm Based On Feature Enhancement

As an important pre task of scene understanding,semantic segmentation has extensive and active research in the field of computer vision.It can be seen in humancomputer interaction,medical diagnosis and object behavior analysis,which provides bottom support for complex applications such as satellite remote sensing system and automatic driving.Image semantic segmentation is an intensive prediction task.Its essence is the extension of image classification task.It is a high-level task to classify each pixel on the image.Since the method based on deep convolutional neural networks(DCNN)won the championship in Image Net competition,researchers in the image field have focused on it.Due to the excellent accuracy and the development of hardware equipment,a large number of image semantic segmentation methods based on DCNN have become the mainstream research idea.The two main factors affecting the performance of semantic segmentation algorithm are spatial information and semantic information.Spatial information contains the detailed features such as texture,edge and color of the target,while semantic information corresponds to different local features of the target.Semantic information is a higher-level abstract expression.High resolution feature maps correspond to rich spatial detail information.Feature maps with strong semantic information generally have small resolution and lack sufficient spatial detail information,which is not conducive to the accurate positioning of targets.The two restrict each other.How to make full use of them to capture the characteristic information suitable for segmentation task is very key.The ability of network model to capture and use this information will directly affect the performance.Therefore,from the perspective of enhancing the features obtained by the model,this paper makes a lot of investigation and Research on the image semantic segmentation algorithm based on DCNN and studies the two-dimensional image semantic segmentation.The specific work is as follows:(1)For common algorithms,there is no well-utilization characteristics,this paper proposes a Dual Branch Information Supplement Network(DBISNet).DBISNet adopts a double branch structure,which is used to obtain low-level detail information and high-level semantic abstraction respectively.These two kinds of information can strengthen the segmentation features from different spatial receptive fields and produce a high-precision segmentation score map.After fusing the features with different information obtained by two branches,the network uses the channel attention mechanism for weighted calculation,so that the network can obtain better features.There is a high correlation between pixels.The features after detail supplement and semantic supplement have stronger expressiveness and contain a large amount of information conducive to segmentation.It is verified on PASCAL VOC2012 dataset and Cityscapes dataset,and the experimental results show the effectiveness of the network.(2)Based on DBISNet,this paper improves it,adds context information branches,and puts forward multi-level feature fusion network(MFFNet).The spatial detail branch of the main branch is used to retain a large amount of spatial detail information;Semantic supplement branch is used to add more high-level semantic abstract information;The context information branch is mainly responsible for extracting multi-scale global information.The network proposes a feature fusion guidance module,which is used to model the corresponding relationship of pixels between different feature maps and effectively fuse features from different branches.In order to get a clear target boundary,a self enhancement feature module is proposed.In the context information branch,pyramid pooling is used to obtain multi-scale context information to solve the pixel misclassification problem caused by the multiscale of the target.On the mainstream semantic segmentation datasets PASCAL VOC2012 and Cityscapes,MFFNet obtains 81.12% m Io U and 74.56% m Io U,which is obviously better than the experimental comparison algorithm.In summary,this paper aims to enhance the features extracted by deep convolution neural network and make it suitable for semantic segmentation tasks by using the strategies of void convolution,context information modeling,multi-scale fusion and attention mechanism.Two methods are proposed in this paper.The experimental results on standard semantic segmentation data sets show that the method can achieve good segmentation accuracy.