Research On Image Semantic Segmentation Algorithm Based On Deep Learning

In recent years,computer vision-related algorithms have been widely used and practical,the research interest has continued to rise.Among them,the image semantic segmentation task is a core research topic.Although it is practical,it is very challenging.However,due to the increasing demand for related engineering applications,most of the data encountered is high-dimensional and large-volume data.The current algorithm model has difficulty in capturing local image information,segmentation edges are not refined enough,relying on a large number of samples,and segmentation accuracy and efficiency are difficult to balance.And so on,it is very valuable to promote the research of semantic segmentation from the level of solving these problems.Therefore,the main research work and contributions of this thesis are as follows:(1)A semantic segmentation model based on adaptive dilated convolution and codec structure is proposed.First,the deep residual network Res Net and the dilated convolution network are studied and analyzed.An adaptive dilated convolution is proposed and the original residual network structure is optimized.Based on this,a multi-scale encoder is designed to make the model correct the image.The subtle local structure is more sensitive,while capturing more image context information.In addition,a decoder that integrates multi-scale features is designed to obtain more comprehensive feature mapping while retaining more image detail information,and can combine features at all levels without increasing the amount of network parameters.The experimental results show that the model has a strong representation ability for images with complex targets,and the accuracy of semantic segmentation is also improved to a certain extent.(2)A conditional generative adversarial segmentation model based on mixed tags is proposed.First,the structure of the conditional generative confrontation network is analyzed,and a generative network module based on mixed label guidance is designed.In order to generate the network and introduce additional guidance information,a discriminant network module based on mixed label loss is proposed.A training algorithm for conditional generative confrontation networks is developed,which mixes real-world images and fake images artificially added with random noise during the training process to provide feedback supervision for network training.Experimental results show that the model not only reduces the need for a large number of samples,but also improves the ability to refine the edges of complex objects in the image,and the generalization performance is also improved.(3)From the perspective of segmentation network prediction performance optimization,the principle characteristics and physical storage forms of several types of low-precision data are deeply studied,a memory layout format is proposed,and the operation optimization of the input data of the network through the DP4A instruction set is designed in detail.Method,with the help of the deep learning algorithm library CuDNN provided by Nvidia company to compare the proposed models on the Tensor RT platform.The experimental results show that when the input data is low precision,the semantic segmentation model can improve the real-time performance of the network without reducing the accuracy,which has great reference value for the implementation and practical application of the algorithm.