Research On Road Scene Semantic Segmentation Algorithm Based On Fully Convolutional Neural Network

With the continuous improvement of the performance of computers,cameras and other equipment,the development of machine learning has reached a new stage.At the same time,due to the continuous rise of all kinds of artificial intelligence,the emergence of new products such as face recognition and autonomous driving,the application of target recognition,semantic segmentation,scene understanding and other technologies also constantly improve the technical requirements.The task of image semantic segmentation plays an important role in the field of image processing.By marking each pixel in a given image,the target class of each pixel is determined,then we can obtain the required pixel level semantic segmentation graph.The quality of the segmentation will directly affect the quality of the subsequent scene understanding.Therefore,the study of image semantic segmentation task has great significance and broad application prospect in the field of image processing.Traditional image segmentation algorithm is to extract relevant features according to the characteristics of the target in the image,which is not suitable for the complex scene,and the segmentation efficiency and accuracy are far from meeting the requirements of various tasks.Image segmentation algorithm based on the deep learning starts from Full Convolution Network(FCN),many excellent semantic segmentation network models have emerged continuously,such as ResNet,PSPNet,SegNet,ENet and so on.These network models can achieve good segmentation effect or high segmentation speed,but they do not achieve the balance between segmentation accuracy and prediction speed.It shows they can not meet the requirements of accuracy and real-time application products.Aiming at these problems,thesis proposes an improved semantic segmentation network,and the specific research work are as follows:(1)The basic feature extraction and sampling module of the network is the improved ResNet-50 module.The basic feature extraction and sampling module samples the original image at different scales with different resolutions through three branches.The first and second branches share the parameters.The low-resolution feature map can obtain global information by the deep network.Higher-resolution and high-resolution feature map can obtain more small target features and local details through shallow network.Then the features of the three branches are fused.It makes the final feature map have more scale image information.This model structure can effectively reduce the number of parameters and improve the speed of network operation.(2)The network adopts the dense connection spatial pyramid pooling structure based on coprime factor,which can reduce the grid effect caused by the introduction of atrous convolution,and at the same time make the network obtain more scale feature information and improve the network segmentation effect.We use the global average pooling to collect the feature map of each atrous convolution layer in the spatial pyramid model.It is more convenient to classify the feature map.This pooling layer has no super parameters,which will not increase the computation of the network and avoid overfitting.(3)The network uses the module named cascade feature fusion unit to fuse the feature graph of the three branches to obtain a more detailed segmentation graph.Besides,we use the tag guidance strategy in the module to improve the accuracy of network segmentation.We adopt the cross-entropy-loss function with weighted to make the network segmentation not too dependent on a branch.(4)We test the algorithm model on Cityscapes data set,and compare it with the experimental results of PSPNet,SegNet and other networks.It shows the effectiveness of the network.