Research Of Image Semantic Segmentation Based On Deep Learning

In recent years,with the continuous development of deep learning technology,semantic segmentation of images based on deep learning technology has become a hot research field in artificial intelligence,and its research results are also widely used in autopilot,robot navigation,medical image analysis and so on.Deep semantic segmentation model's network structure is relatively large,the number of layers and parameters is relatively large,and the features of different layers extracted are also different.The shallow features are generally low-level boundary features of the image,and these features are very favorable for the positioning of objects in the image.The high-level features are generally high-level abstract image semantic information.These features are conducive to the classification and recognition of different objects in the image area,but they lose more spatial information.Semantic image segmentation not only requires accurate positioning and segmentation of different objects in the image,but also identifies their semantic categories.Therefore,this paper studies how to use the features of the deep segmentation model and how to change the model network structure to reduce the number of model parameters.Major works and innovations of this paper are listed as follows:(1)Building the current mainstream deep learning framework Caffe in the Linux environment.(2)Aiming at the problem that the network structure of deep segmentation model is huge and the computation is complex,a semantic segmentation model with dilated convolution is proposed and design its network structure.The features of the FC6 layer with the most parameters are extracted by using dilated convolution,building and training the proposed model in the Caffe framework.The experimental results show that the proposed model can maintain high segmentation performance while reduce the number of parameters and computational complexity.(3)Exploring the impact of fusing high-level features on semantic segmentation performance,design and transform the model structure.Due to the lack of training datasets,using transfer learning to train the FCN32s model,fused high-level features'FCN32s model FCN8s model and fused high-level features' FCN8s model in the Stanford background class dataset.Through the comparison and analysis of the experimental results,we can see that the appropriate fusion of some high-level features can effectively improve the performance of semantic segmentation.If the fusion is too much,the improvement of semantic segmentation performance is not significant.(4)Aiming at the fact that most of the deep segmentation models use the features of different layers only simply summed up the features of the corresponding pixels,we propose a semantic segmentation method based on the weighted fusion.This method transforms the VGG16 classification model into a segmentation model,and creates a weight control module to control the fusion of each position feature of different layers before upsample.The model is built and trained in the Caffe framework,experimental results show that the designed model can make full use of features of different layers,and is more conducive to the correct segmentation of the boundary of different objects and the interior of the same object,improves the semantic segmentation performance of the outdoor scene image.