Research On Real-Time Semantic Segmentation Algorithm Based On Deep Learning

As a key research task in the field of computer vision,image semantic segmentation is the foundation of many applications.By predicting the category of each pixel in an image,the image is divided into several regions with specific semantic information.Semantic segmentation can generate denser predictions,provide richer information,and facilitate scene understanding.With the continuous development of convolutional neural networks,semantic segmentation networks have also achieved relatively accurate segmentation accuracy.However,due to the high complexity of the network,i.e.,the large number of parameters and high computational complexity,the running speed is slow and cannot be applied in some realtime scenarios.To reduce model complexity,it is necessary to balance between high accuracy and high speed.Therefore,the research on effective real-time semantic segmentation algorithms is of great significance.This article focuses on this problem and the specific research contents are as follows:1)A real-time semantic segmentation network based on global feature guidance combined with traditional image boundary operators is proposed.Most dual branch real-time semantic segmentation networks use the original image as the input of low-level feature extraction branches.In this paper,the boundary image extracted by traditional image boundary operators is used to replace the original input image,which can enhance the detailed feature representation through feature fusion modules.In addition,in the backbone part of feature extraction,the global feature guidance module is used to extract the global information of hierarchical features,and at the end of the network,the two extracted features are fused and predicted.Experiments show that the network has a resolution of 960 in input images × At720,the Tensor RT inference test achieved a running speed of 210 FPS,and achieved 68.6%MIo U on the Cam Vid dataset.2)A real-time semantic segmentation network based on multiscale feature fusion and improved down sampling is proposed.Semantic segmentation networks enhance the receptive field of feature maps through continuous downsampling,reducing resolution while retaining semantic information,reducing computational complexity and memory consumption,but causing partial loss of detail information.Therefore,this paper treats the feature map as an image block,and expands the feature channel by using equidistant sampling.In the feature fusion section,the CBAM convolutional attention module,which combines spatial and channel attention,is combined with the pyramid multiscale feature fusion module to more fully utilize the features extracted from the network and fuse global context information.At the same time,boundary assist loss is introduced during training to optimize the prediction effect of the model.The experiment shows that the fourth chapter network has an accuracy improvement of 0.8% MIo U on the Cam Vid dataset compared to the third chapter network.In addition,this network achieved a testing accuracy of 69.66% MIo U on the Cityscapes dataset.3)The real-time semantic segmentation evaluation system is realized,and the proposed real-time semantic segmentation algorithm is evaluated.The system provides image prediction and video stream prediction functions,you can select pictures or video files locally,or open the camera to capture video in real time,and visualize the prediction graph obtained by model inference,which more intuitively displays the model prediction effect.Tests show that this paper proposes that the network performs well in real life scene prediction,and the contours of trees,vehicles,buildings,sky and other objects are clear,which has certain practical value.