Research On Road Semantic Segmentation Method Based On Memory And Attention Mechanism

With the development of deep convolutional neural networks,many computer vision tasks have been greatly improved.As one of the most important tasks in computer vision,image semantic segmentation has also made good progress by relying on deep learning related technologies.As we all know,image semantic segmentation plays an important role in various tasks in real life,such as autonomous driving and augmented reality.Image semantic segmentation is a difficult task in computer vision.The difficulty lies in the fact that the foreground and background objects in reality are very complex.Usually these objects have different shapes,sizes,and colors.In many semantic segmentation tasks,street view segmentation is even more difficult to achieve the ideal state.Although deep convolutional neural networks have improved the effect of traditional semantic segmentation tasks a lot,there are still several problems that need to be solved for street scene segmentation that is actually applied in real life.These problems can be summarized as the following two points: 1)Street scene segmentation contains complex foreground and background objects.These objects are not only different in shape and size,but also easy to overlap with each other,making it more difficult to train deep convolutional nerves than ordinary semantic segmentation tasks.Network;2)Street view segmentation is usually used in the preprocessing of autonomous driving images,but the speed of traditional methods cannot meet the requirements of autonomous driving.To solve the above problems,this paper proposes a semantic segmentation network based on memory discrimination enhancement and multi-mode sampling,which can segment target objects in complex street scenes.The network uses the image pyramid model to obtain multi-scale information and global information,and introduces a memory module to reconstruct the input features to improve the expression ability of the features,thereby enhancing the network model's ability to segment complex scenes.At the same time,based on the requirements of real-time applications for segmentation speed,this paper proposes a lightweight attention segmentation network.The network can obtain a great speed increase while ensuring the original segmentation effect,making real-time applications possible.In summary,this article improves the performance of the original fully convolutional neural network from the perspective of segmentation quality and segmentation speed,so that it can perform better when performing street scene segmentation tasks.