Research On Fast Image Segmentation Algorithm Based On Full Convolutional Network

The natural gap between human vision and computer understanding of images makes the semantic segmentation a huge challenge.In many areas such as autonomous driving,medical imaging,and image processing,sophisticated semantic segmentation results are very important.Therefore,whether it is academic or industrial,research on semantic segmentation algorithms has always been a hot topic.With the innovation of deep learning in recent years,CNN has demonstrated its powerful feature extraction ability and achieved success in many fields.How to use the convolutional network for image semantic segmentation has become the focus of research.The current convolutional network-based semantic segmentation algorithm is based on variants of FCN,U-net,and Deeplab algorithms.Regardless of the U-net barrel structure or the hollow convolution of the Deeplab algorithm,it only considers the accuracy of the algorithm,and does not consider the speed.A semantic segmentation algorithm with higher hardware requirements and slower speed is difficult to fall in practice application.Therefore,this paper analyzes the U-net algorithm of the full convolution network,summarizes its advantages and disadvantages,redesigns and trains a new end-to-end image semantic segmentation algorithm model,main works as follows:First of all,the full-convolution-based fast semantic image segmentation algorithm designed in this paper can input images of any size without any requirement for size.In order to speed up the algorithm network,this paper replaces most of the convolution operation in the algorithm with separable convolution,which greatly reduces the parameter amount of the algorithm.The fast upsampling module is fast with no training parameters.By using the fusion of channel information and spatial information,the network enhances the feature learning of channel information during the learning process and enhances the representation ability of the model.This paper analyzes the bottleneck of U-net network in the lack of context information collection.Therefore,this paper designs a parallel context information acquisition module with Bottleneck structure,which effectively improves the network context information collection capability and further improves the output accuracy.Secondly,the training and testing methods of the fast image segmentation algorithm model based on full convolution are designed.The new network model is trained and tested on the KITTI road segmentation dataset and the hyperparameters such as learning rate and L2 regularization parameters are carried out optimization.Finally,by comparing with the top-ranking algorithms on other KITTI road segmentation datasets,the fast image semantic segmentation algorithm designed in this paper greatly improves the computational speed while ensuring accuracy.Under the input of the original resolution,the computation time is other networks.1/20,0.05s,the maximum F index(MaxF)reached 95.82%...