Research On Real-Time Semantic Segmentation Methods Based On Convolutional Neural Network

Image Semantic Segmentation is an extremely popular research direction in the field of computer vision at present.Its purpose is to divide the input image into several groups of pixel regions with a certain semantic meaning,and automatically mark the category of each region,and finally output Images with semantic annotations.The convolutional neural network(Convolutional Neural Network,CNN)has raised the requirements of image semantic segmentation tasks to a new level.Although the traditional convolutional neural network has greatly improved the accuracy of segmentation,aside from accuracy,the primary consideration for the CNN network should be the computational complexity of the model.Because the network running speed is very low when the complexity is too high,some real-time application scenarios that need to meet the requirements of low delay,such as automatic driving scenarios,have extremely strict requirements on the running efficiency of the model.In addition,a fast and good lightweight model is also the best choice for the deployment of mobile devices.At present,the requirements for model practice and landing are strict,so it is particularly important to achieve a balance between high precision and high efficiency in image semantic segmentation.Therefore,it is of great research significance to balance the accuracy and speed of the model in the real-time semantic segmentation task.This article focuses on this task.The specific research content is as follows:Firstly,a lightweight convolutional neural network for real-time semantic segmentation is designed.Compared with the development of lightweight network in recent years,it is more inclined to adopt shallow structure.This network is dedicated to designing deeper network structure to obtain better feature expression,while maintaining faster reasoning speed and higher segmentation accuracy.The new extremely efficient residual module uses depth separable expansion convolution to learn the feature representation of receptive fields of different scales,and in addition,it also adds multiple jump connection branches to collect context information from the intermediate convolution layer.The model size is only 0.8M,and the running speed of 60 FPS is achieved with the input image resolution of 1024×512.It achieves the best balance of testing speed and accuracy on urban landscape data sets.Secondly,a real-time semantic segmentation network based on attention guidance mechanism is designed.Taking the encoder-decoder structure as the backbone of the network,with the extremely efficient residual module feature extraction unit,the newly proposed adaptive attention module is used to capture the correlation information between each pixel,and the low-level and the high-level features are connected to improve the feature expression ability of the network.Experiments show that the model size of the network is only 0.81 M,and the accuracy on the urban landscape dataset has reached 72.4%,achieving a 60 FPS operating speed,which exceeds the real-time standard of mobile devices.Thirdly,a lightweight semantic segmentation network based on parallel branching is designed.A deep category semantic branch with rich semantic information,a shallow spatial detail branch with rich detail information and a global attention module are designed to establish the connection between the whole network.On the basis of two branches,one deep and one shallow,the global attention module is introduced again,and the information of the category semantic branch is used to guide the spatial detail branch to further locate the details,and achieve a balance between speed and accuracy.Experiments show that the accuracy of the network on the urban landscape data set reaches 73.3%,and the running speed reaches 54 FPS,which can be applied to more complex real-time scene understanding tasks.