Research On Image Semantic Segmentation Based On Fully Convolutional Neural Network

The traditional image segmentation mainly aims to segment foreground objects from background images.However,this method cannot handle complex scenes.Image semantic segmentation,a new technique which combines deep-learning,is to classify an image in pixel-level and separate it into different parts based on semantics.It can analyze and understand complex scenes,e.g.,to evaluate road conditions and help unmanned driving detecting pedestrians.The convolutional neural network plays an influential role in promoting the rapid development of semantic segmentation.Specifically,to train a network by inputting images with corresponding annotations,then let it complete feature extraction and segmentation by itself.This can enhance the accuracy of semantic segmentation.On account of the objects of natural images that usually have different sizes and ratios,and a variety of textures,it is crucial to get details and contextual information as many as possible while making pixel-level predictions.Based on the consideration of a larger convolution kernel could have a positive impact on object location and segmentation,this thesis comes up with a contextual structure based fully convolutional neural network(CS-FCNN).The network consists of a basic net and a contextual structure,respectively.The basic net is used to sufficiently extract the features of an image,and contextual structure is used to extract global information and multi-scale information.Then,it does the semantic segmentation with fusing the information from low layers,global information from high layers,and multi-scale information.The main contributes of CS-FCNN are listed as follows:(1)it improved the realization of convolution kernel,enlarged the range of receptive field,optimized extraction of global information,and solved the problem of too many parameters raised while increasing the size of the kernel;(2)it supports multi-scale feature extraction.In this case,the local and global information can be extracted effectively when segmenting multi-scale objects.The experiment indicates that CS-FCNN performs better than some existing segmentation algorithms on classic public datasets.Though the performance of the improved semantic segmentation algorithm has made progress,there are still problems with segmentation a nd edge optimization.Due to CS-FCNN continuously using atrous convolution with the same rate parameters,the origin feature map losses details.To figure out the issue,some measures have been taken to optimize the way of combination among atrous convolut ions.Therefore,the thesis further optimizes the method mentioned above and puts forward modified CS-FCNN(MCS-FCNN).Additionally,the technique in CS-FCNN to enlarge feature maps is crude.So,MSC-FCNN restores the native resolution of an image by sharing information between intermediate layer feature maps.It increases the performance of MCS-FCNN in two points: the accuracy of segmentation and the edge smoothness of objects.With applying CNN to image semantic segmentation can not only make it fits better in complex scenes but also address multiple objects segmentation issues.This paper introduces a new method to improve traditional CNN.From experimental results,the improved net performs better when taking computation complexity and efficiency into consideration.