Image Semantic Segmentation Algorithms Based On Feature Fusion And Non-local Features

Semantic segmentation of images is an important branch in the field of computer vision.It has a deep influence on satellite image segmentation,medical image diagnosis and driverless technology.More and more deep learning methods about semantic segmentation are proposed after the appearance of full convolutional network.At present,most researches of semantic segmentation are divided into two aspects.One tries to improve the accuracy of semantic segmentation by using relative algorithms.The other makes networks lighter to meet the needs of real-time requirement.This thesis is committed to improving the accuracy of semantic segmentation.The thesis proposes a feature fusion network which emphasizes the different importance of different semantic level features.The network can solve the imbalance of semantic information on encoder-decoder structure.The thesis also proposes another network by using recurrent neutral network to build non-local attention.It effectively solves the problem of space correlation in semantic segmentation.The feature fusion network mainly proposes three modules,including semantic enhancement module,attention module and dense decoding module.The semantic enhancement module constructs a strategy to enrich semantic information based on deformable convolution.These modules are connected to different scale features of backbone.This module can effectively overcome the shortcoming that low-level features lack context information.At the same time,it does not increase computation.In the process of constructing the attention module,the high-level semantic features and low-level semantic features are fully integrated to construct the attention weight of the fusion features.The weight is used to weight the channel to obtain more context information.Therefore,the network can make a choice among different channels automatically.The dense decoding module adds dense links on decoder side so that the overall output of the network gives different importance to different semantic level features.This module not only maximizes the contribution of high-level features,but also preserves the spatial details provided by low-level features.The experiments are tested on PASCAL VOC2012,Cityscapes,and ADE20 K datasets by using multiple NVIDIA Tesla V100 graphics cards.This experiment achieves 81.9%,80.0%,43.76%m Io U on each of the three datasets.The experiment proves the effectiveness of the feature fusion method and corresponding module.The non-local attention network divides the output of backbone into different feature blocks,which respectively correspond to the input of each time sequence of the recurrent neural network.This method combines different timing outputs to form attention features.Then the features are fused with input features so that each position of the output contains entire space information.In this thesis,different feature segmentation methods are listed and their accuracy and speed are tested on the Cityscapes dataset.This thesis proves this novel mechanism can improve the accuracy and has a similar time consumption with convolutional neural networks.