| Semantic segmentation is a segmentation task for image pixels,that is,each pixel in the image is assigned a specific semantic label.Convolutional neural networks are used in image segmentation due to their translation invariance.The semantic segmentation model uses this as the basic network,which greatly improves the accuracy of segmentation.Semantic segmentation,as one of the core objects of computer vision tasks,has been applied to a large extent in the practical development fields such as video surveillance,unmanned driving,geological detection,and robotic sensing.However,in the mobile field or devices with limited computer resources,the classical semantic segmentation model limits its practical application in terms of model size,inference speed,calculation and memory cost.Semantic segmentation models need to take into account segmentation accuracy and reasoning speed.Therefore,it is essential to construct a lightweight,efficient,and real-time semantic segmentation method.In view of the fact that classical semantic segmentation algorithms cannot achieve efficient inference speed and there is a natural contradiction between semantic segmentation accuracy and inference speed.This paper studies the characteristics of the shallow network and the deep network to extract features,changes the convolution method and the number of channels to capture the features reasonably,and introduces the singular matrix decomposition and approximate extraction of typical features.Specifically,this paper proposes a semantic segmentation method based on a lightweight convolutional neural network,called LCNNet.This method builds a lightweight backbone network based on the powerful encoder and streamlined decoder.By reasonably changing the convolution method and the number of channels,multi-scale feature fusion is performed on the abstract semantic information in the deep network layer and the spatial detail information in the shallow network layer.What’s more,it is designed as the main component of the powerful encoder,in order to extract context information in an efficient and concise way.In the simplify decoder,drawing on the ideas of singular value decomposition and low-rank matrix approximation,the singular value attention module is designed to extract robust and representative typical features from the feature map to ensure that the segmentation accuracy is further improved without affecting the real-time inference speed of the model.A large number of experiments have proved the superiority of the LCNNet method.Firstly,the proposed multi-scale feature fusion module and singular value attention module are verified through ablation experiments;In the Cam Vid data set and Cityscapes data set,this paper’s method and other excellent semantic segmentation methods are further compared and analyzed experimentally in terms of accuracy and speed.A large number of experiments were proved that the LCNNet method achieves efficient segmentation efficiency while ensuring segmentation accuracy. |
