| Image segmentation technology has always played an important role in computer vision tasks,and is widely used in fields such as automatic driving,medical image processing,and automatic translation.Although the semantic segmentation technology has been greatly improved compared to the traditional method,it's classification accuracy is still low when dealing with small objects or fuzzy objects.In addition,Some pooling and hole convolution will be used to obtain a large receptive field when constructing and training the deep learning models,which can make the pixel classification more accurate.However,the pixel loss during the operation affects the final segmentation accuracy,and the problem that pixels cannot be efficiently restored in upsampling will cause blurred boundary of the segmentation result.Semantic segmentation requires real-time processing,so how to balance accuracy and speed has become an important task of semantic segmentation.The main research contents of this paper are as follows:1)Based on the U-shaped structure,a real-time semantic segmentation model of gated multi-layer fusion is proposed.The model uses a horizontal U-shaped connection with gating.The structure uses the advantage of the attention mechanism to target screening.In the process of semantic information transmission in the horizontal parallel layer,the attention is paid to the target pixel to obtain more Target pixel information,filtering other useless feature content,also known as filtering background information.The advantage of this structure is that it extracts more important target pixels and provides more target features for the up-sampling parallel layer.The final result is a good segmentation result.In addition,multiple layers of fusion are used in the upsampling process to complement the advantages of different layers of semantic information,thereby improving pixel diversity and making pixel restoration more accurate.While adding more structures to improve accuracy,the downsampling structure is simplified,and the advantage of 1×1 convolution is used to perform multiple dimensionality reduction processes to reduce the operating parameters of the model,thereby enabling the model to ensure good real-time results.2)Improved the gated multi-layer fusion semantic segmentation model.The gating structure of the horizontal connection is changed to a dual-channel gating structure,and the two gating structures are used to obtain more comprehensive target characteristics according to the different semantics of different levels,and the transfer performance of the horizontal connection is enhanced.In addition,the upsampling structure is changed to a dual-channel upsampling structure,and different extraction methods of different structures are used to obtain different upsampling feature maps,so that the upsampling feature maps contain more comprehensive semantic information,and then the obtained feature maps are fused.The final upsampling feature map obtained is more accurate,and the output accuracy is improved.The model structure proposed in this paper makes an important contribution in solving computer vision tasks,and at the same time lays the foundation for subsequent target detection and target tracking.Finally,the model is proposed to be tested on the CamVid dataset.When the input size is 512×1024,the mIOU reaches 74.1%,37 frames/s. |
