Research On Street-oriented Lightweight Semantic Segmentation Model Based On Residual Network

One of the key points for realizing autonomous driving is the understanding of street scenes.Image semantic segmentation can provide rich pixel-level information for the understanding of street scenes.Using this information,each target in the street scene can be finely classified.With the rapid development of neural networks,the performance of image semantic segmentation models based on neural networks has been greatly improved.The current semantic segmentation model has reached a bottleneck in the improvement of segmentation accuracy.The research on the semantic segmentation model has begun to develop in the direction of lightweighting the model.Existing semantic segmentation models based on neural network have too many parameters and huge model volume.In order to improve these problems,this paper deeply studies convolutional neural networks and image semantic segmentation models.First of all,this paper studies the composition of convolutional neural networks,including convolutional layers,pooling layers,activation functions,batch normalization layers,etc.On this basis,this paper studies a typical semantic segmentation model based on convolutional neural networks and some Neural network lightweight method.Finally,this paper select a lightweight semantic segmentation model Link Net based on the residual network as the initial model.In order to achieve a smaller and better semantic segmentation model S-Link Net,the model Link Net is improved to achieve efficient segmentation of street scene images.In terms of model improvement,first use Res Net-34 residual network to improve model performance,then use deep separation convolution to reduce the model parameter amount,finally use channel pruning compression technology to compress the model.The improved semantic segmentation model is trained and tested on the City Scapes data set.The experimental results show that the deep residual network can significantly improve the model segmentation performance,and the deep separation convolution can greatly reduce the model parameters,the model volume can be compressed by using channel pruning technology.The semantic segmentation model S-Link Net improves the accuracy of the initial model Link Net segmentation m IOU by 4.6%,the model parameter amount is reduced by16.8%,and the model volume is reduced by 21.9%.Compared with other semanticsegmentation models,the semantic segmentation model proposed in this paper achieves a better balance between accuracy and volume.