Semantic Segmentation Based On Convolutional Neural Networks

The image semantic segmentation algorithm of full convolutional neural network(FCN)can not only directly process the image of any size,save the tedious intermediate steps,achieve end-to-end pixel segmentation,but also improve the accuracy of segmentation.However,when the pooling layer in the full convolutional neural network reduces the complexity of the network model and improves the operation speed,the resolution of the feature map will be reduced by many times due to the use of the pooling layer.To handle the problem of segmenting objects at multi-scale,this paper proposed a method for image semantic segmentation based on multi-scale feature fusion.The specific research is as follows:Firstly,the new model used a jump connection method.It used the feature maps obtained from the pooling layers pool3 and pool4 with the feature maps extracted by the atrous spatial pyramid pooling module.It can combine rough deep information with fine low-level information to further improve the accuracy of semantic segmentation.Secondly,the new model improved the atrous spatial pyramid pooling(ASPP)module.The new model changed the four kinds of atrous convolution with different atrous rates parallel Fc6 in the convolution layer of ASPP module into six kinds of atrous convolution with different atrous rates parallel,which enhanced the network's ability to capture feature information from different scales and further improved the accuracy of image segmentation.Thirdly,the new model used a fully connected conditional random field(CRF)method.It uses CRF at the outer end to improve the performance of the pixel-level classifier,which can not only capture the details of the boundary,but also adapt to the long-distance dependence and improve the segmentation effect.The training and validation on PASCAL VOC 2012 data sets showed that the experimental results achieve 82.0% of the mean pixel accuracy and 71.9% of the mean intersection over union.The results show that the model can achieve good image semantics segmentation effect by improving the ASPP method and adopting the fully connected CRF method.