Image Semantic Segmentation Based On Attention Mechanism And Lightweight Network

As an important task in computer vision,there is a wide range of prospects for semantic segmentation,such as autonomous driving,medical imaging and augmented reality.The early image segmentation algorithms struggle to meet the demands of practical applications in terms of accuracy and efficiency.Image semantic segmentation has made a breakthrough in recent years with the development of deep learning.Two semantic segmentation algorithms based on the DeepLabv3+ model are proposed in this paper.The specific work is as follows.(1)DeepLabv3+ algorithm based on dual attention mechanism is proposed.To aggregate multi-scale features and refine the segmentation results of objects with different scales,the various dilated convolutions are used in the ASPP module of the DeepLabv3+algorithm.However,there are some drawbacks in the expression on global context information and the differences on the channel features for ASPP,which may result in the loss of a lot of details.A novel algorithm called deep Labv3+ algorithm based on dual attention mechanism is proposed to avoid the problems.The feature discrimination ability is improved after adding the position attention module which model the global contextual relationship for ASPP.Furthermore,the channel attention module enhances the learning of important channel features and improves the semantic segmentation accuracy by obtaining the interdependence between different feature channels.Experiments on the PASCAL VOC 2012 augmented dataset demonstrate that the mean intersection over union(MIo U)and the mean pixel accuracy(MPA)of the proposed method are improved by 2.05% and2.19% compared with the DeepLabv3+ algorithm respectively,and the segmentation accuracy is higher than that of other five comparison algorithms.(2)Lightweight image semantic segmentation based on attention mechanism and densely adjacent prediction is proposed.Xception network is employed as the backbone of the DeepLabv3+ model to extract image features,resulting in complex network structure and increased computing complexity.The feature map generated by the ASPP module gives the same weight to each channel and does not take into account the differences between different channel features,thus crucial channel features cannot stand out effortlessly.In addition,high semantic feature map generated by decoder lacks sufficient detailed information,causing poor segmentation results.A novel algorithm named lightweight image semantic segmentation based on attention mechanism and densely adjacent prediction is proposed to solve the problems.The lightweight Mobile Net V2 is regarded as the backbone network to reduce model parameters.After the multi-scale information is extracted by the channel atrous spatial pyramid pooling,each channel of the feature map is weighted to reinforce the learning of important channel features.Moreover,the segmentation results are refined since densely adjacent prediction is utilized to combine high-level and low-level features.Experiments on the PASCAL VOC 2012 augmented dataset reveal higher values of the MIo U and MPA of the proposed method than in the state-of-the-art algorithms.Compared with DeepLabv3+,the parameters and calculation amount are decreased by 184.82 M and 90.83 GFLOPs respectively.The presented algorithm not only improves the segmentation accuracy,but also reduces the computation cost in comparison with the baseline algorithm.