| Semantic segmentation classifies each pixel in a realistic image to produce a semantic image with different categories of segmentation.Semantic segmentation is mainly used in applications such as cancer cell or organ segmentation for medical imaging and urban road segmentation for unmanned vehicles.However,there is still a very serious gap between implementability and application requirements in reality.Currently,there is an explosion of visual data and a hot research area in computer vision with semantic segmentation tasks for pixel-level classification is taking shape.Translated with www.Deep L.com/Translator(free version)This paper will design the algorithm for real-time semantic segmentation,and study the balance between the accuracy and speed of the semantic segmentation model and the class imbalance of the dataset.The first part focuses on the real-time semantic segmentation algorithm research on multi-channel and multi-Loss value supervision;the second part focuses on the real-time semantic segmentation algorithm research on convolution factorization and class weight reassignment.The main research contents and contributions of this paper are as follows:1)because of the existing data set category imbalance and the uncertainty of small object position information,a method based on multi-channel multi-Loss value supervised learning is introduced.A real-time semantic segmentation network based on multi-channel multi-loss supervision is designed.On the one hand,the objects are accurately segmented according to multi-channel without increasing the computational burden;The experimental results show that the method proposed in this paper can achieve an average intersection ratio of 67.13% and a processing speed of 55 fps on the Cityscapes semantic segmentation dataset.2)Aiming at the loss of segmentation details and confusion of contextual semantic information that is common in real-time semantic segmentation algorithms,and an efficient multi-scale split fusion network is proposed.The network solves the problem based on the effective fusion processing of contextual semantic features and spatial location features.First,convolution factorization is introduced into the convolutional layer,and the semantic segmentation model can achieve real-time effects with relatively few parameters.Then,the multi-scale attention fusion module is used to aggregate the two low-resolution feature maps output by the two encoding modules into global semantic information,thereby eliminating the loss of details and semantic confusion in the process of semantic segmentation.The method achieves better performance than existing real-time methods on semantic segmentation datasets.At the same time,it also has high efficiency,and the speed can reach 60.6fps.The experimental results show that the method proposed in this paper can achieve an average intersection ratio of 72.00%on the Cityscapes semantic segmentation dataset. |
