Research On Lightweight Semantic Segmentation For Road Scenes

With the rise of concepts such as autonomous driving and smart cities,artificial intelligence technology has entered people’s production and life with various terminal products.Allowing smart terminals to recognize and understand the environment around users is the basis for many applications and is also a current research hotspot.direction.Environmental perception is an important module in the intelligent driving system.Its main task is to perceive pedestrians,vehicles,traffic lights and other objects around the vehicle,so that the system can refer to and make correct path planning.Image semantic segmentation is a basic and important task in the field of computer vision.Through the classification and recognition of each pixel in the image,various objects in the image are labeled,and it can be applied to assisted driving,automatic driving,scene analysis,drones and other scenes.in.With the development of semantic segmentation models,deep convolutional neural networks can have high recognition accuracy and can achieve good segmentation results.However,the model has a large volume and many parameters,requires strong computing power,memory consumption and time-consuming,and cannot be applied.To the real-time scene.Based on the lightweight technology of deep convolutional neural networks,this paper proposes a real-time lightweight network.In order to be applicable to real-time scenes and for the characteristics of visual images,this paper proposes a lightweight model MS-Pspnet that can be used for real-time semantic segmentation based on the research of the existing classic semantic segmentation models and lightweight semantic segmentation models..In the design of the feature extraction network residual module,the depth separable convolution and grouped convolution are used to reduce the parameter amount of the residual module,and then the module structure and stacking method are designed according to the memory usage,and a pyramid pooling module is designed to strengthen the feature extraction module Extract contextual information at different scales.In order to improve the accuracy of the model,a lightweight attention module is added to the convolution module,low-dimensional feature information is obtained in the auxiliary branch of the main feature extraction network,and the dilated convolution is added to increase the receptive field,and adjust the expansion rate and activation of the cavity convolution function.After construction,the model is trained and tested on the Cityscapes dataset.The model in this paper is experimentally verified on the Cityscapes data set.Compared with other deep convolutional network models,the parameters of the model in this paper are relatively small.The segmentation accuracy of the model in this paper can reach 66%-70% on NVIDIA1080 ti.The reasoning speed exceeds 30,which basically achieves the balance between segmentation accuracy and speed,and becomes an effective method to solve semantic segmentation tasks.