Research On Image Semantic Segmentation Based On Fully Convolutional Network

As an important part of scene understanding in computer vision,the main task of image semantic segmentation is to predict category label pixel by pixel in an image.It can achieve fine-grained prediction,so that each pixel is marked to a certain category wit hin its closed region.Therefore,image semantic segmentation can deeply understand image scene,and it plays a significant role in application fields such as autonomous vehicle,virtual reality,and land monitoring.Currently,the mainstream methods in semantic segmentation are based on deep learning,which can train the network to fit the annotation data to obtain the best-fitting network model.Existing networks have greatly improved in accuracy,but because of the excessively stacked network layers,the memory consumption is serious,which limits its application in fields with high real-time requirements such as automatic driving and Internet of things.Therefore,this thesis aims to study a real-time semantic segmentation method based on a fully convolutional network.The main structure of the thesis is as follows:First,in view of the large amount of calculation and low efficiency of the existing high-precision models and inspired by the lightweight network model,this thesis proposes a real-time semantic segmentation method based on an asymmetric codec.The encoder part of the network adopts a newly designed bottleneck residual module based on depth-wise separable convolution,dilated convolution,and factorized convolution to extract local information and contextual information,without significantly increasing the amount of calculation.At the same time,a channel shuffle mechanism is used to promote the exchange of information between channels.In the decoder,a novel non-local pyramid module is able to extract global information and multi-scale information,meanwhile,perform segmentation result mapping.The experiment results have proved that compared with the current network models,the model achieves the trade-off between the accuracy and speed.Then,aiming at the problem of undifferentiated fusion of high-level features and low-level features in existing codec models,this thesis proposes a real-time semantic segmentation method based on attention mechanism.The encoder-decoder parts of this method is constructed with down-sampling module,bottleneck residual module and up-sampling module.At the same time,a newly designed global attention guidance module based on the attention mechanism is introduced at the jump connection between codecs to guide the low-level features of the encoder structure and the high-level features of the decoder structure to better integrate,so that the low-level features and the high-level features can be better integrated,and the accuracy is improved.In addition,a brand-new global average pooling module is used between the encoder and the decoder to introduce global information.The experiment results demonstrated that this method can obtain high segmentation accuracy under the condition of limited computing resources,such as mobile devices.