Research On Image Semantic Segmentation Algorithm Based On DCNN

The appearance of artificial intelligence makes the production and life of human beings enter the era of intelligence.Robots have replaced the labor force,replacing humans to complete dangerous heavy work;The advent of autonomous driving could eliminate the need for humans to drive themselves;Meitu,retouching software is already everywhere.However,these intelligent implementations are inseparable from the research of semantic segmentation technology.With the rapid development of deep learning,the research of image semantic segmentation technology has become very mature,but there are still some shortcomings.In recent years,semantic segmentation has been improved in a single way.Most algorithms only change the network structure,leading to no obvious breakthrough in performance.The quantity of training data also has a great influence on the result of semantic segmentation.In this paper,in order to improve the performance of semantic segmentation,and solve the problems such as large difference in target size,similarity in features between classes,large difference in features within classes,a new semantic segmentation system for arbitrary size is constructed,and the structure and training method of semantic segmentation network for large size are improved.First,we compare the segmentation performance of each feature extraction network,and respectively construct a simple semantic segmentation model for the Res Net,Dense Net and Xception as feature extraction networks,and conduct experiments on them.We determine the optimal segmentation effect of Xception as the basis of the new semantic segmentation network.Because the current semantic segmentation network can not achieve satisfactory segmentation effect for small-size input,we construct a new semantic segmentation system which can achieve satisfactory effect for input of any size.The system is composed of large size semantic segmentation model and small size semantic segmentation model.When the width or the height of the input image is greater than 200,the image is input into the large-size semantic segmentation model;otherwise,the image is input into the small-size semantic segmentation model.Because the background of small size image is fuzzy and can only segment one semantic,the small size semantic segmentation network proposed in this paper can only segment pedestrians.The experimental results show that the performance of large size semantic segmentation still needs to be improved.In terms of network structure,we mainly adopt the attention mechanism.The aim is toenhance the network's focus on key features through the attention mechanism.In the process of training,attention mechanism can shorten the distance of effective features and enlarge the proportion of effective features,so that the whole network pays more attention to the key information.Based on the understanding of attention mechanism,three kinds of attention modules are proposed according to the principles of spatial attention,channel attention and mixed attention.Considering that the attention module can be inserted into the network at will,we propose six attention modules for the fusion process and upsampling process.According to the experimental results,a new encoder was constructed by combining the mixed attention module and Xception network,and a new decoder was constructed by combining the channel domain attention module with the residual mode,thus a new semantic segmentation model was proposed.The performance of the new encoder,decoder and the new semantic segmentation model is better than deeplabv3 plus.In terms of training methods,we use the idea of adversarial training to realize the full-supervised training of the new semantic segmentation model.We add the adversarial loss to the generating network,optimize the network model,and improve the segmentation performance of the new semantic segmentation model once again.On the basis of the full-supervised training model,we add the semi-supervised loss to the generating network,and only use a small part of labeled training data to realize the semi-supervised training of the semantic segmentation model,which solves the problem of semantic labeling difficulty.The experimental results show that the performance of semi-supervised segmentation is better than that of deeplab V3 plus.