Image Segmentation Method Based On Context Aggregation And Location Awareness

In recent years,with the development of neural network algorithms,deep learning technology has made breakthroughs in various tasks of image recognition,image semantic segmentation and instance segmentation have become the hot spots in computer vision research.Semantic segmentation has important applications in autonomous driving,medical diagnosis and other fields.The goal is to classify a given image pixel by pixel according to different categories.However,semantic segmentation can only divide categories simply as a pixel classification problem,while instance segmentation can divide each instance of the object contained in the image.Because the number of instances in various categories is different,the instance segmentation task is more challenging.This paper probes the problems existing in semantic segmentation and instance segmentation and completes the following works:In the task of semantic segmentation,this thesis analyzes the shortcomings of the current model based on full convolutional neural network,and proposes an improved semantic segmentation model.Due to the limitation of convolution kernel,most methods tend to model the relationship between local regions,but seldom explore the remote pixel correlation.The two difficulties of current research are how to realize the aggregation of multi-scale context and how to obtain dense context information effectively.To solve these two difficulties,this thesis proposes a dual context aggregation module(Dual Context Aggregation Module,DCM).DCM is divided into two attention modules,which establish the dependency relationship between spatial location and channels respectively to obtain dense context information.The spatial attention module adopts a crossing structure which integrates the horizontal and vertical dependencies of the target position.In the channel attention module,the self-attention method is also used to explore the relationship between each pair of channels.In order to obtain target information of different scales,this thesis also designs a two-step decoder structure to add low-level features into the network to improve the recognition rate of small-scale targets.A large number of evaluation experiments on the benchmarks show that the introduction of attention module improves the performance of the model.Compared with the current method,it shows a competitive advantage.For instance segmentation task,the existing work is mainly based on two methods: the methods of detecting and then dividing,and the methods of grouping pixels first and dividing different instances.Both of them implement prediction in two stages.The result of segmentation depends on the quality of bounding box or the accuracy of pixel classification,so it is impossible to train the category and mask simultaneously.However,some model optimization of single stage is poor and cannot achieve high accuracy.In order to solve these problems,this thesis proposes an instance segmentation method based on context aggregation and location awareness,which generates the prediction of category and instance mask of different spatial positions in the image simultaneously.First,the feature maps are divided into several grids,and the prediction network is divided into two parallel branches: semantic branch and mask branch.In these two branches,the category of each grid position is connected with the mask,and then identifies the instance.In addition,two coordinate feature channels are added to the convolutional network to make the network more sensitive to the spatial position.The spatial attention module and the channel attention module are applied to the two prediction branches,so that the dependence between different positions can be established to improve the accuracy of the model.Extensive ablation experiments on MS COCO2017 dataset confirmed the advancement of this instance segmentation model.